A neural network-based operator home broadband complaint prediction method and system

Abstract

The application discloses a neural network-based operator home broadband complaint prediction method and system, and relates to the technical field of big data mining and analysis. Since home broadband problems are difficult to be quickly solved, the scheme comprises the following steps: obtaining user complaint work orders, extracting device state information and alarm information at the time of user complaints, constructing feature engineering and generating a supervised learning time series dataset after mining and preprocessing; dividing the dataset, designing a neural network architecture based on LSTM, training the neural network by using the divided dataset, and finally outputting an LSTM multivariate time series prediction model; optimizing the training results of the prediction model, introducing a time series attention mechanism and a combined loss function, and improving the performance of the model in adapting to the complaint prediction task; and deploying the optimized prediction model to an actual business system, positioning potential complaint customers and predicting complaint reasons by accessing real-time device state information and alarm information. The application can improve the efficiency of on-site handling by installation and maintenance personnel.

Description

Technical Field

[0001] This invention relates to the field of big data mining and analysis technology, specifically to a method and system for predicting operator home broadband complaints based on neural networks. Background Technology

[0002] With the growth of operators' home broadband business, the number of complaints about home broadband services has been increasing year by year. Handling customer complaints has become an important part of improving customer service satisfaction.

[0003] When customers experience network problems, installation and maintenance personnel often need to rely on their experience to conduct on-site inspections of various network devices in the customer's home, as well as upstream OLTs, BRAS, and other equipment, step by step to pinpoint the cause of the problem and then resolve it. During this process, frontline installation and maintenance personnel often spend a significant amount of time identifying the cause of the problem and accumulate valuable experience through extensive work. Summary of the Invention

[0004] This invention addresses the needs and shortcomings of current technological development by providing a method and system for predicting complaints about home broadband services from operators based on neural networks. By predicting the reasons for complaints from home broadband customers within a certain period, it improves the efficiency of broadband fault location and reduces the workload of front-line installation and maintenance personnel.

[0005] In a first aspect, the present invention provides a method for predicting operator home broadband complaints based on neural networks, and the technical solution adopted to solve the above-mentioned technical problem is as follows: A method for predicting operator-subsidized home broadband complaints based on neural networks includes the following steps: S1. Obtain user complaint work orders, extract device status information and alarm information at the time of user complaints from the work orders, mine and preprocess them, construct feature engineering and generate supervised learning time series datasets to lay the data foundation for subsequent prediction model training. S2. Divide the supervised learning time series dataset into time sequences, design an LSTM-based neural network architecture in combination with the business requirements of complaint prediction, complete the model training through forward propagation, loss calculation, back propagation and parameter update, and finally output an LSTM multivariate time series prediction model adapted to the complaint prediction task. S3. Optimize the training results of the LSTM multivariate temporal prediction model by adjusting the network structure and hyperparameters, analyzing the model performance bottlenecks and proposing data-level improvement schemes. At the same time, introduce the temporal attention mechanism and the combined loss function to improve the performance of the model in the complaint prediction task. S4. Deploy the optimized LSTM multivariate time series prediction model to the actual business system. By accessing real-time device status information and alarm information, locate potential complaining customers and predict the reasons for complaints.

[0006] Optionally, step S1 specifically includes: S1.1 Obtain user complaint work orders, mine and extract key information from the work orders, specifically including four data closely related to the complaint: OLT device status information, ONU device status information, PON port transmit and receive information, and alarm information, and construct feature engineering based on the extracted information. S1.2. Based on the constructed feature engineering, multivariate time series data are generated. Standardized formulas are used to process the multivariate time series data to ensure that each feature is in the same dimension, thus eliminating the impact of dimensional differences on subsequent analysis. S1.3. Based on the standardized data, identify outliers in the data and remove work order data with outliers; at the same time, identify and delete duplicate records to avoid deviations in data analysis caused by outlier and duplicate data. S1.4. Further normalize the equipment status data output in step S1.3, compress the feature values ​​to a specified range, and eliminate the interference caused by the differences in units and magnitudes between different data. S1.5. Based on the normalized data, with a time step of T and a prediction step of H, construct a supervised learning dataset, where: Input data X=[x_t-T+1,x_t-T+2,...,x_t]∈R^(T×M), where T is the time step, M is the feature dimension, x_t-T+1 is the start time of the input sequence, and x_t is the end time of the input sequence; The output data y=[y_t+1,y_t+2,...,y_t+H]∈R^H forms the input and output data format of the time series prediction model; where H is the prediction step size, y_t+1 is the first future time to be predicted, and y_t+H is the last future time to be predicted.

[0007] Optionally, step S2 specifically includes: S2.1. According to the preset ratio, the constructed supervised learning time series dataset is divided into training set, validation set and test set in chronological order; S2.2 Based on the business nature of complaint prediction, design the parameters of the LSTM-based neural network architecture, including the number of network layers, the number of neurons in each layer, and the type of activation function, to ensure that the architecture adapts to the long-term and short-term dependencies of time series data and meets the task requirements of complaint prediction. S2.3 During the neural network training process, the following operations are performed sequentially: S2.3.1 Input the training set data into the neural network for forward propagation. The predicted output results are obtained step by step through the calculation of each layer of neurons, thus completing the mapping from input data to predicted values. S2.3.2. The mean squared error loss function is used to calculate the difference between the predicted output of the training set and the true label, and to quantify the model prediction error. S2.3.3 Based on the gradient information of the training set loss function, the gradient of each network weight in the neural network is solved by the BackPropagation Through Time (BPTT) algorithm, providing a quantitative basis for weight updates; S2.3.4. The Adam optimization algorithm is used for parameter iterative updates. First, the statistical information of the gradient of the training set is obtained through first-order moment estimation and second-order moment estimation. Then, the network weights of the neural network are adjusted according to the parameter update formula to continuously reduce the prediction error of the neural network on the training set. At the same time, the loss change of the validation set is used to monitor the overfitting of the neural network. The training is terminated by an early stopping strategy. After the training is completed, the performance of the neural network is verified by the test set. Finally, an LSTM multivariate time series prediction model adapted to the complaint prediction task is output.

[0008] Further optionally, step S3 specifically includes: S3.1 Based on the output results of the trained LSTM multivariate temporal prediction model on the validation and test sets, the initial single-layer LSTM network architecture is optimized into a 3-layer LSTM network structure. At the same time, combined with the characteristics of the complaint prediction task, key hyperparameters such as the number of temporal attention heads and the number of hidden layer neurons are adjusted to improve the model's ability to capture complex temporal dependencies. S3.2. Regarding the issue that the adjusted LSTM multivariate time series prediction model does not show significant improvement in accuracy on the validation and test sets and that the inference speed is reduced by 4 times, the improvement solution is to expand the dataset size over time. Regarding the issue that the LSTM multivariate time series prediction model has low precision and F1-score on the test set, the improvement solution is to add more feature variables and extend the data time span in subsequent model updates. S3.3. Based on the upgraded 3-layer LSTM network in step S3.1, a temporal attention mechanism is introduced. Attention scores are calculated and normalized to obtain attention weights, and then weighted summation is performed to obtain the attention context vector, thereby strengthening the weight of key temporal features. At the same time, the loss function is optimized by using a combination of mean squared error, mean absolute error and Huber loss to balance the influence of different error indicators and improve the model's adaptability to data distribution skew. S3.4 After optimizing the loss function, the LSTM multivariate time series prediction model is retrained using the training set. The training process is monitored and early stopping is implemented using the validation set. Finally, the performance improvement of the optimized LSTM multivariate time series prediction model is verified using the test set.

[0009] Optionally, step S4 involves interpreting the prediction results output by the LSTM multivariate time series prediction model for business purposes, generating a prediction report and pushing it to relevant business departments; at the same time, it compares the actual complaint data with the prediction results, records the deviation, and provides feedback data for the continuous iterative optimization of the model.

[0010] Secondly, the present invention provides a neural network-based operator home broadband complaint prediction system, and the technical solution adopted to solve the above-mentioned technical problems is as follows: A neural network-based system for predicting home broadband complaints from telecom operators, comprising: The work order acquisition and processing module is used to acquire user complaint work orders, extract the device status information and alarm information at the time of the user complaint from the work orders, mine and preprocess them, construct feature engineering and generate supervised learning time series datasets, and lay the data foundation for subsequent prediction model training. The data partitioning module is used to divide the constructed supervised learning time series dataset into training set, validation set and test set according to a preset ratio and in chronological order. The model training module is used to design an LSTM-based neural network architecture based on the business needs of complaint prediction. Then, using the training set, validation set, and test set, the model is trained through forward propagation, loss calculation, back propagation, and parameter update, and finally outputs an LSTM multivariate time series prediction model adapted to the complaint prediction task. The model optimization module is used to optimize the training results of the LSTM multivariate temporal prediction model. It improves the model's performance in the complaint prediction task by adjusting the network structure and hyperparameters, analyzing model performance bottlenecks, and proposing data-level improvement schemes. At the same time, it introduces a temporal attention mechanism and a combined loss function. The model deployment and usage module is used to deploy the optimized LSTM multivariate time series prediction model to the actual business system. By accessing real-time device status information and alarm information, it can locate potential complaining customers and predict the reasons for complaints.

[0011] Optionally, the work order acquisition and processing module involved specifically includes: The work order acquisition and extraction unit is used to acquire user complaint work orders, mine and extract key information from the work orders, specifically including four data closely related to the complaint: OLT device status information, ONU device status information, PON port transmit and receive information, and alarm information. Feature engineering is constructed based on the extracted information. The data standardization unit is used to generate multivariate time series data based on the constructed feature engineering. The multivariate time series data is processed using a standardization formula to ensure that each feature is in the same dimension, thereby eliminating the impact of dimensional differences on subsequent analysis. The data cleaning unit is used to identify outliers in the standardized data and remove work order data with outliers; at the same time, it identifies and deletes duplicate records to avoid deviations in data analysis caused by outlier and duplicate data. The normalization processing unit is used to further normalize the equipment status-related data output by the data cleaning unit, compress the feature values ​​to a specified range, and eliminate the interference caused by differences in units and magnitudes between different data. The dataset construction unit is used to construct a supervised learning dataset based on normalized data, with a time step of T and a prediction step of H, wherein: Input data X=[x_t-T+1,x_t-T+2,...,x_t]∈R^(T×M), where T is the time step, M is the feature dimension, x_t-T+1 is the start time of the input sequence, and x_t is the end time of the input sequence; The output data y=[y_t+1,y_t+2,...,y_t+H]∈R^H forms the input and output data format of the time series prediction model; where H is the prediction step size, y_t+1 is the first future time to be predicted, and y_t+H is the last future time to be predicted.

[0012] Optionally, the model training modules involved specifically include: The network architecture design unit is used to design LSTM-based neural network architecture parameters based on the business nature of complaint prediction, including the number of network layers, the number of neurons in each layer, and the type of activation function, to ensure that the architecture adapts to the long-term and short-term dependencies of time series data and meets the task requirements of complaint prediction. The network training unit performs the following operations sequentially during neural network training: ① Inputting training set data into the neural network for forward propagation, and gradually obtaining the predicted output through the calculation of each layer of neurons, completing the mapping from input data to predicted values; ② Using the mean squared error loss function, calculating the difference between the predicted output of the training set and the true label, quantifying the model's prediction error; ③ Based on the gradient information of the training set loss function, solving the gradient of each network weight in the neural network through the Back Propagation Through Time (BPTT) algorithm, providing a quantitative basis for weight updates; ④ Using the Adam optimization algorithm for iterative parameter updates, first obtaining the statistical information of the training set gradient through first-order moment estimation and second-order moment estimation, and then adjusting the network weights of the neural network according to the parameter update formula, continuously reducing the prediction error of the neural network on the training set. At the same time, the validation set is used to monitor the model overfitting, and the training is terminated through an early stopping strategy. After training, the performance of the neural network is verified using the test set, and finally, an LSTM multivariate time series prediction model adapted to the complaint prediction task is output.

[0013] Optionally, the model optimization module may perform the following operations to optimize the training results of the LSTM multivariate time series prediction model: Based on the output results of the trained LSTM multivariate temporal prediction model on the validation and test sets, the initial single-layer LSTM network architecture is optimized into a 3-layer LSTM network structure. At the same time, combined with the characteristics of the complaint prediction task, key hyperparameters such as the number of temporal attention heads and the number of hidden layer neurons are adjusted to improve the model's ability to capture complex temporal dependencies. To address the issues of insignificant accuracy improvement and a 4-fold decrease in inference speed on the validation and test sets for the adjusted LSTM multivariate time series prediction model, the improvement solution is to expand the dataset size over time. To address the issues of low precision and F1-score on the test set for the LSTM multivariate time series prediction model, the improvement solution is to add more feature variables and extend the data time span in subsequent model updates. Based on the upgraded 3-layer LSTM network, a temporal attention mechanism is introduced. Attention weights are obtained by calculating attention scores and normalizing them, and then weighted summation is used to obtain the attention context vector, which strengthens the weight of key temporal features. At the same time, the loss function is optimized by using a combination of mean squared error, mean absolute error and Huber loss to balance the influence of different error indicators and improve the model's adaptability to data distribution skew. After optimizing the loss function, the LSTM multivariate time series prediction model is retrained using the training set. The training process is monitored and early stopping is implemented using the validation set. Finally, the performance improvement of the optimized model is verified using the test set.

[0014] Optionally, the systems involved may also include: The report generation and feedback update module is used to interpret the prediction results output by the LSTM multivariate time series prediction model for business purposes, generate prediction reports and push them to relevant business departments; at the same time, it compares the actual complaint data with the prediction results, records the deviation, and provides feedback data for the continuous iterative optimization of the model.

[0015] The present invention provides a method and system for predicting operator home broadband complaints based on neural networks, which has the following advantages compared with the prior art: This invention utilizes historical complaint work order data from home broadband users and employs an LSTM multivariate time series prediction model algorithm for model training. This model can accurately locate potential complaining customers and predict the reasons for their complaints, providing guidance for frontline installation and maintenance personnel in handling user faults. This significantly improves the efficiency of on-site handling by installation and maintenance personnel and enhances the satisfaction of home broadband customers. Attached Figure Description

[0016] Appendix Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention; Appendix Figure 2 This is a module connection block diagram of Embodiment 2 of the present invention. Detailed Implementation

[0017] To make the technical solution, the technical problem solved, and the technical effect of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments.

[0018] Example 1: Refer to Appendix Figure 1 This embodiment proposes a method for predicting operator-subsidized home broadband complaints based on neural networks, which includes the following steps: S1. Obtain user complaint work orders, extract device status information and alarm information at the time of user complaints from the work orders, perform mining and preprocessing, construct feature engineering and generate supervised learning time series datasets to lay the data foundation for subsequent prediction model training.

[0019] This step specifically includes the following operations: S1.1 Obtain user complaint work orders, mine and extract key information from the work orders, specifically including four data closely related to the complaint: OLT device status information, ONU device status information, PON port transmit and receive information, and alarm information, and construct feature engineering based on the extracted information. S1.2. Based on the constructed feature engineering, multivariate time series data are generated, and a standardized formula is used ( ,in These are the original eigenvalues. The characteristic mean, The multivariate time series data are processed to ensure that each feature is on the same dimension (i.e., the standard deviation of the features), thereby eliminating the influence of dimensional differences on subsequent analysis. S1.3. Based on the standardized data, identify outliers in the data and remove work order data with outliers; at the same time, identify and delete duplicate records to avoid deviations in data analysis caused by outlier and duplicate data. S1.4. Further normalize the equipment status data output in step S1.3, compress the feature values ​​to a specified range, and eliminate the interference caused by the differences in units and magnitudes between different data. S1.5. Based on the normalized data, with a time step of T and a prediction step of H, construct a supervised learning dataset, where: Input data ,in, For time step, For feature dimension, This represents the starting time of the input sequence. This represents the termination time of the input sequence; Output data This forms the input and output data format adapted to the time series prediction model; among which, To predict the step size, For the first future moment of the prediction, This refers to the last future moment predicted.

[0020] S2. Divide the supervised learning time series dataset into time sequences, design an LSTM-based neural network architecture based on the business requirements of complaint prediction, complete the model training through forward propagation, loss calculation, backpropagation and parameter update, and finally output an LSTM multivariate time series prediction model adapted to the complaint prediction task.

[0021] This step specifically includes the following operations: S2.1. Divide the constructed supervised learning time series dataset into training set, validation set and test set according to the preset ratio (such as 7:2:1 or 8:1:1) in chronological order. S2.2 Based on the business nature of complaint prediction, design the parameters of the LSTM-based neural network architecture, including the number of network layers, the number of neurons in each layer, and the type of activation function, to ensure that the architecture adapts to the long-term and short-term dependencies of time series data and meets the task requirements of complaint prediction. S2.3 During the neural network training process, the following operations are performed sequentially: S2.3.1 Input the training set data into the neural network for forward propagation. The predicted output results are obtained step by step through the calculation of each layer of neurons, thus completing the mapping from input data to predicted values. S2.3.2. The mean squared error loss function is used to calculate the difference between the predicted output of the training set and the true label, quantifying the model prediction error; the expression of the mean squared error loss function is as follows: , In the formula, This represents the value of the loss function; N represents the set of model parameters; N represents the number of samples in the dataset. Indicates the first The true value of each sample; Indicates the first The predicted value for each sample; S2.3.3. Based on the gradient information of the training set loss function, the gradient of each network weight in the neural network is solved using the Back Propagation Through Time (BPTT) algorithm, providing a quantitative basis for weight updates. The formula for the Back Propagation Through Time (BPTT) algorithm is as follows: , In the formula, This represents the gradient of the loss function with respect to the model parameters; Represents the loss function; Represents the set of model parameters; Indicates the total number of time steps; This represents the hidden state at time t; This represents the gradient of the loss function with respect to the hidden state at time t; This represents the gradient of the hidden state with respect to the model parameters at time t; S2.3.4. The Adam optimization algorithm is used for parameter iterative updates. First, the statistical information of the gradient of the training set is obtained through first-order moment estimation and second-order moment estimation. Then, the network weights of the neural network are adjusted according to the parameter update formula to continuously reduce the prediction error of the neural network on the training set. At the same time, the loss change of the validation set is used to monitor the overfitting of the neural network. The training is terminated by an early stopping strategy. After the training is completed, the performance of the neural network is verified by the test set. Finally, an LSTM multivariate time series prediction model adapted to the complaint prediction task is output.

[0022] First-order moment estimation formula: , In the formula, express The first moment estimate at time t; Indicates the number of iterations / update rounds; express The first moment estimate at time t; This represents the attenuation coefficient of the first moment; This represents the weight coefficient of the current gradient; express The gradient value at time step; Second-order moment estimation formula: , In the formula, express The estimated value of the second moment at time t; Indicates the number of iterations / update rounds; express The estimated value of the second moment at time t; This represents the attenuation coefficient of the second moment; This represents the weight coefficient of the current gradient; express The gradient value at time step; express Element-wise square of the gradient value at each time step; Parameter update formula: , In the formula, express Model parameters at time points; express Model parameters at time points; Indicates the learning rate; express The deviation correction value of the first moment at time; express The deviation correction value of the second moment at time; express The element-wise square root of the second moment deviation correction value at time 1; This represents the smoothing term, which is a minimal constant.

[0023] S3. Optimize the training results of the LSTM multivariate temporal prediction model by adjusting the network structure and hyperparameters, analyzing the model performance bottlenecks and proposing data-level improvement schemes. At the same time, introduce the temporal attention mechanism and the combined loss function to improve the model's performance in the complaint prediction task.

[0024] This step specifically includes the following operations: S3.1 Based on the output results of the trained LSTM multivariate temporal prediction model on the validation and test sets, the initial single-layer LSTM network architecture is optimized into a 3-layer LSTM network structure. At the same time, combined with the characteristics of the complaint prediction task, key hyperparameters such as the number of temporal attention heads and the number of hidden layer neurons are adjusted to improve the model's ability to capture complex temporal dependencies. S3.2. Regarding the issue that the adjusted LSTM multivariate time series prediction model does not show significant improvement in accuracy on the validation and test sets and that the inference speed is reduced by 4 times, the improvement solution is to expand the dataset size over time. Regarding the issue that the LSTM multivariate time series prediction model has low precision and F1-score on the test set, the improvement solution is to add more feature variables and extend the data time span in subsequent model updates. S3.3. Based on the upgraded 3-layer LSTM network in step S3.1, a temporal attention mechanism is introduced. Attention scores are calculated and normalized to obtain attention weights, and then weighted summation is performed to obtain the attention context vector, thereby strengthening the weight of key temporal features. At the same time, the loss function is optimized by using a combination of mean squared error, mean absolute error and Huber loss to balance the influence of different error indicators and improve the model's adaptability to data distribution skew. Temporal attention formula: , , , In the formula, Indicates the first Attention score at each time step; The weight vector representing the attention score; This represents the transpose operation of a vector; The weight matrix representing the attention mechanism; Represents a concatenated matrix of temporal hidden state sequences; The bias vector representing the attention mechanism; Indicates the first Attention weights for each time step; The attention context vector is a weighted sum of the hidden states at all time steps, which centrally reflects the key information in the time sequence. This represents the hidden state after weighting; Formula for combined loss function: , In the formula, , , These are the weighting coefficients of the loss function. ; This represents the mean squared error loss. Indicates the mean absolute error loss. This indicates Huber's loss.

[0025] S3.4 After optimizing the loss function, the LSTM multivariate time series prediction model is retrained using the training set. The training process is monitored and early stopping is implemented using the validation set. Finally, the performance improvement of the optimized LSTM multivariate time series prediction model is verified using the test set.

[0026] S4. Deploy the optimized LSTM multivariate time series prediction model to the actual business system. By accessing real-time device status information and alarm information, locate potential complaining customers and predict the reasons for complaints. At the same time, interpret the prediction results output by the LSTM multivariate time series prediction model for business purposes, generate prediction reports and push them to relevant business departments. Compare the actual complaint data with the prediction results, record the deviation, and provide feedback data for the continuous iteration and optimization of the model.

[0027] Example 2: Refer to Appendix Figure 2 This embodiment proposes a network-based system for predicting operator-managed home broadband complaints, which includes: The work order acquisition and processing module is used to acquire user complaint work orders, extract the device status information and alarm information at the time of the user complaint from the work orders, mine and preprocess them, construct feature engineering and generate supervised learning time series datasets, and lay the data foundation for subsequent prediction model training. The data partitioning module is used to divide the constructed supervised learning time series dataset into training set, validation set and test set according to a preset ratio and in chronological order. The model training module is used to design an LSTM-based neural network architecture based on the business needs of complaint prediction. Then, using the training set, validation set, and test set, the model is trained through forward propagation, loss calculation, back propagation, and parameter update, and finally outputs an LSTM multivariate time series prediction model adapted to the complaint prediction task. The model optimization module is used to optimize the training results of the LSTM multivariate temporal prediction model. It improves the model's performance in the complaint prediction task by adjusting the network structure and hyperparameters, analyzing model performance bottlenecks, and proposing data-level improvement schemes. At the same time, it introduces a temporal attention mechanism and a combined loss function. The model deployment and usage module is used to deploy the optimized LSTM multivariate time series prediction model to the actual business system. By accessing real-time device status information and alarm information, it can locate potential complaining customers and predict the reasons for complaints.

[0028] In this embodiment, the work order acquisition and processing module specifically includes: The work order acquisition and extraction unit is used to acquire user complaint work orders, mine and extract key information from the work orders, specifically including four data closely related to the complaint: OLT device status information, ONU device status information, PON port transmit and receive information, and alarm information. Feature engineering is constructed based on the extracted information. The data standardization unit is used to generate multivariate time series data based on the constructed feature engineering. The multivariate time series data is processed using a standardization formula to ensure that each feature is in the same dimension, thereby eliminating the impact of dimensional differences on subsequent analysis. The data cleaning unit is used to identify outliers in the standardized data and remove work order data with outliers; at the same time, it identifies and deletes duplicate records to avoid deviations in data analysis caused by outlier and duplicate data. The normalization processing unit is used to further normalize the equipment status-related data output by the data cleaning unit, compress the feature values ​​to a specified range, and eliminate the interference caused by differences in units and magnitudes between different data. The dataset construction unit is used to construct a supervised learning dataset based on normalized data, with a time step of T and a prediction step of H, wherein: Input data X=[x_t-T+1,x_t-T+2,...,x_t]∈R^(T×M), where T is the time step, M is the feature dimension, x_t-T+1 is the start time of the input sequence, and x_t is the end time of the input sequence; The output data y=[y_t+1,y_t+2,...,y_t+H]∈R^H forms the input and output data format of the time series prediction model; where H is the prediction step size, y_t+1 is the first future time to be predicted, and y_t+H is the last future time to be predicted.

[0029] In this embodiment, the model training module specifically includes: The network architecture design unit is used to design LSTM-based neural network architecture parameters based on the business nature of complaint prediction, including the number of network layers, the number of neurons in each layer, and the type of activation function, to ensure that the architecture adapts to the long-term and short-term dependencies of time series data and meets the task requirements of complaint prediction. The network training unit performs the following operations sequentially during neural network training: ① Inputting training set data into the neural network for forward propagation, and gradually obtaining the predicted output through the calculation of each layer of neurons, completing the mapping from input data to predicted values; ② Using the mean squared error loss function, calculating the difference between the predicted output of the training set and the true label, quantifying the model's prediction error; ③ Based on the gradient information of the training set loss function, solving the gradient of each network weight in the neural network through the Back Propagation Through Time (BPTT) algorithm, providing a quantitative basis for weight updates; ④ Using the Adam optimization algorithm for iterative parameter updates, first obtaining the statistical information of the training set gradient through first-order moment estimation and second-order moment estimation, and then adjusting the network weights of the neural network according to the parameter update formula, continuously reducing the prediction error of the neural network on the training set. At the same time, the validation set is used to monitor the model overfitting, and the training is terminated through an early stopping strategy. After training, the performance of the neural network is verified using the test set, and finally, an LSTM multivariate time series prediction model adapted to the complaint prediction task is output.

[0030] In this embodiment, the model optimization module performs the following operations to optimize the training results of the LSTM multivariate time series prediction model: Based on the output results of the trained LSTM multivariate temporal prediction model on the validation and test sets, the initial single-layer LSTM network architecture is optimized into a 3-layer LSTM network structure. At the same time, combined with the characteristics of the complaint prediction task, key hyperparameters such as the number of temporal attention heads and the number of hidden layer neurons are adjusted to improve the model's ability to capture complex temporal dependencies. To address the issues of insignificant accuracy improvement and a 4-fold decrease in inference speed on the validation and test sets for the adjusted LSTM multivariate time series prediction model, the improvement solution is to expand the dataset size over time. To address the issues of low precision and F1-score on the test set for the LSTM multivariate time series prediction model, the improvement solution is to add more feature variables and extend the data time span in subsequent model updates. Based on the upgraded 3-layer LSTM network, a temporal attention mechanism is introduced. Attention weights are obtained by calculating attention scores and normalizing them, and then weighted summation is used to obtain the attention context vector, which strengthens the weight of key temporal features. At the same time, the loss function is optimized by using a combination of mean squared error, mean absolute error and Huber loss to balance the influence of different error indicators and improve the model's adaptability to data distribution skew. After optimizing the loss function, the LSTM multivariate time series prediction model is retrained using the training set. The training process is monitored and early stopping is implemented using the validation set. Finally, the performance improvement of the optimized model is verified using the test set.

[0031] It should be added that the system in this embodiment may further include: The report generation and feedback update module is used to interpret the prediction results output by the LSTM multivariate time series prediction model for business purposes, generate prediction reports and push them to relevant business departments; at the same time, it compares the actual complaint data with the prediction results, records the deviation, and provides feedback data for the continuous iterative optimization of the model.

[0032] In summary, the neural network-based method and system for predicting operator home broadband complaints of this invention, by utilizing historical complaint work order data of home broadband users and training the model using the LSTM multivariate time series prediction model algorithm, can accurately locate potential complaining customers and predict the reasons for their complaints. This provides guidance for frontline installation and maintenance personnel to handle user faults, significantly improves the efficiency of on-site handling by installation and maintenance personnel, and enhances the satisfaction of home broadband customers.

[0033] The above specific examples illustrate the principles and implementation methods of the present invention in detail. These embodiments are merely for the purpose of helping to understand the core technical content of the present invention. Based on the above specific embodiments of the present invention, any improvements and modifications made to the present invention by those skilled in the art without departing from the principles of the present invention should fall within the patent protection scope of the present invention.

Claims

1. A method for predicting operator-submitted home broadband complaints based on neural networks, characterized in that, Includes the following steps: S1. Obtain user complaint work orders, extract device status information and alarm information at the time of user complaints from the work orders, mine and preprocess them, construct feature engineering and generate supervised learning time series datasets to lay the data foundation for subsequent model training. S2. Divide the supervised learning time series dataset into time sequences, design an LSTM-based neural network architecture in combination with the business requirements of complaint prediction, complete the model training through forward propagation, loss calculation, back propagation and parameter update, and finally output an LSTM multivariate time series prediction model adapted to the complaint prediction task. S3. Optimize the training results of the LSTM multivariate temporal prediction model by adjusting the network structure and hyperparameters, analyzing the model performance bottlenecks and proposing data-level improvement schemes. At the same time, introduce the temporal attention mechanism and the combined loss function to improve the performance of the model in the complaint prediction task. S4. Deploy the optimized LSTM multivariate time series prediction model to the actual business system. By accessing real-time device status information and alarm information, locate potential complaining customers and predict the reasons for complaints.

2. The method for predicting operator-based home broadband complaints according to claim 1, characterized in that, Step S1 specifically includes: S1.1 Obtain user complaint work orders, mine and extract key information from the work orders, specifically including four data closely related to the complaint: OLT device status information, ONU device status information, PON port transmit and receive information, and alarm information, and construct feature engineering based on the extracted information. S1.

2. Based on the constructed feature engineering, multivariate time series data are generated. Standardized formulas are used to process the multivariate time series data to ensure that each feature is in the same dimension, thus eliminating the impact of dimensional differences on subsequent analysis. S1.

3. Based on the standardized data, identify outliers in the data and remove work order data with outliers; at the same time, identify and delete duplicate records to avoid deviations in data analysis caused by outlier and duplicate data. S1.

4. Further normalize the equipment status data output in step S1.3, compress the feature values ​​to a specified range, and eliminate the interference caused by the differences in units and magnitudes between different data. S1.

5. Based on the normalized data, with a time step of T and a prediction step of H, construct a supervised learning dataset, where: Input data X=[x_t-T+1,x_t-T+2,...,x_t]∈R^(T×M), where T is the time step, M is the feature dimension, x_t-T+1 is the start time of the input sequence, and x_t is the end time of the input sequence; The output data y=[y_t+1,y_t+2,...,y_t+H]∈R^H forms the input and output data format of the time series prediction model; where H is the prediction step size, y_t+1 is the first future time to be predicted, and y_t+H is the last future time to be predicted.

3. The method for predicting operator-based home broadband complaints according to claim 1, characterized in that, Step S2 specifically includes: S2.

1. According to the preset ratio, the constructed supervised learning time series dataset is divided into training set, validation set and test set in chronological order; S2.2 Based on the business nature of complaint prediction, design the parameters of the LSTM-based neural network architecture, including the number of network layers, the number of neurons in each layer, and the type of activation function, to ensure that the architecture adapts to the long-term and short-term dependencies of time series data and meets the task requirements of complaint prediction. S2.3 During the neural network training process, the following operations are performed sequentially: S2.3.1 Input the training set data into the neural network for forward propagation. The predicted output results are obtained step by step through the calculation of each layer of neurons, thus completing the mapping from input data to predicted values. S2.3.

2. The mean squared error loss function is used to calculate the difference between the predicted output of the training set and the true label, and to quantify the model prediction error. S2.3.3 Based on the gradient information of the training set loss function, the gradient of each network weight in the neural network is solved by the time backpropagation algorithm, providing a quantitative basis for weight update; S2.3.

4. The Adam optimization algorithm is used for parameter iterative updates. First, the statistical information of the gradient of the training set is obtained through first-order moment estimation and second-order moment estimation. Then, the network weights of the neural network are adjusted according to the parameter update formula to continuously reduce the prediction error of the neural network on the training set. At the same time, the loss change of the validation set is used to monitor the overfitting of the neural network. The training is terminated by an early stopping strategy. After the training is completed, the performance of the neural network is verified by the test set. Finally, an LSTM multivariate time series prediction model adapted to the complaint prediction task is output.

4. The method for predicting operator-based home broadband complaints according to claim 3, characterized in that, Step S3 specifically includes: S3.1 Based on the output results of the trained LSTM multivariate temporal prediction model on the validation and test sets, the initial single-layer LSTM network architecture is optimized into a 3-layer LSTM network structure. At the same time, combined with the characteristics of the complaint prediction task, the number of temporal attention heads and the number of hidden layer neurons are adjusted to improve the model's ability to capture complex temporal dependencies. S3.

2. Regarding the issue that the adjusted LSTM multivariate time series prediction model does not show significant improvement in accuracy on the validation and test sets and that the inference speed is reduced by 4 times, the improvement solution is to expand the dataset size over time. Regarding the issue that the LSTM multivariate time series prediction model has low precision and F1-score on the test set, the improvement solution is to add more feature variables and extend the data time span in subsequent model updates. S3.

3. Based on the upgraded 3-layer LSTM network in step S3.1, a temporal attention mechanism is introduced. Attention scores are calculated and normalized to obtain attention weights, and then weighted summation is performed to obtain the attention context vector, thereby strengthening the weight of key temporal features. At the same time, the loss function is optimized by using a combination of mean squared error, mean absolute error and Huber loss to balance the influence of different error indicators and improve the model's adaptability to data distribution skew. S3.4 After optimizing the loss function, the LSTM multivariate time series prediction model is retrained using the training set. The training process is monitored and early stopping is implemented using the validation set. Finally, the performance improvement of the optimized LSTM multivariate time series prediction model is verified using the test set.

5. The method for predicting operator-based home broadband complaints according to claim 1, characterized in that, In step S4, the prediction results output by the LSTM multivariate time series prediction model are interpreted for business purposes, a prediction report is generated and pushed to the relevant business departments; at the same time, the actual complaint data is compared with the prediction results, the deviation is recorded, and feedback data is provided for the continuous iterative optimization of the model.

6. A network-based system for predicting operator complaints about home broadband, characterized in that, It includes: The work order acquisition and processing module is used to acquire user complaint work orders, extract the device status information and alarm information at the time of the user complaint from the work orders, mine and preprocess them, construct feature engineering and generate supervised learning time series datasets, and lay the data foundation for subsequent prediction model training. The data partitioning module is used to divide the constructed supervised learning time series dataset into training set, validation set and test set according to a preset ratio and in chronological order. The model training module is used to design an LSTM-based neural network architecture based on the business needs of complaint prediction. Then, using the training set, validation set, and test set, the model is trained through forward propagation, loss calculation, back propagation, and parameter update, and finally outputs an LSTM multivariate time series prediction model adapted to the complaint prediction task. The model optimization module is used to optimize the training results of the LSTM multivariate temporal prediction model. It improves the model's performance in the complaint prediction task by adjusting the network structure and hyperparameters, analyzing model performance bottlenecks, and proposing data-level improvement schemes. At the same time, it introduces a temporal attention mechanism and a combined loss function. The model deployment and usage module is used to deploy the optimized LSTM multivariate time series prediction model to the actual business system. By accessing real-time device status information and alarm information, it can locate potential complaining customers and predict the reasons for complaints.

7. The operator-based home broadband complaint prediction system according to claim 6, characterized in that, The work order acquisition and processing module specifically includes: The work order acquisition and extraction unit is used to acquire user complaint work orders, mine and extract key information from the work orders, specifically including four data closely related to the complaint: OLT device status information, ONU device status information, PON port transmit and receive information, and alarm information. Feature engineering is constructed based on the extracted information. The data standardization unit is used to generate multivariate time series data based on the constructed feature engineering. The multivariate time series data is processed using a standardization formula to ensure that each feature is in the same dimension, thereby eliminating the impact of dimensional differences on subsequent analysis. The data cleaning unit is used to identify outliers in the standardized data and remove work order data with outliers; at the same time, it identifies and deletes duplicate records to avoid deviations in data analysis caused by outlier and duplicate data. The normalization processing unit is used to further normalize the equipment status-related data output by the data cleaning unit, compress the feature values ​​to a specified range, and eliminate the interference caused by differences in units and magnitudes between different data. The dataset construction unit is used to construct a supervised learning dataset based on normalized data, with a time step of T and a prediction step of H, wherein: Input data X=[x_t-T+1,x_t-T+2,...,x_t]∈R^(T×M), where T is the time step, M is the feature dimension, x_t-T+1 is the start time of the input sequence, and x_t is the end time of the input sequence; The output data y=[y_t+1,y_t+2,...,y_t+H]∈R^H forms the input and output data format of the time series prediction model; where H is the prediction step size, y_t+1 is the first future time to be predicted, and y_t+H is the last future time to be predicted.

8. A network-based operator home broadband complaint prediction system according to claim 6, characterized in that, The model training module specifically includes: The network architecture design unit is used to design LSTM-based neural network architecture parameters based on the business nature of complaint prediction, including the number of network layers, the number of neurons in each layer, and the type of activation function, to ensure that the architecture adapts to the long-term and short-term dependencies of time series data and meets the task requirements of complaint prediction. The network training unit performs the following operations sequentially during neural network training: ① Inputting training set data into the neural network for forward propagation, and gradually obtaining the predicted output through the calculation of each layer of neurons, completing the mapping from input data to predicted values; ② Using the mean squared error loss function, calculating the difference between the predicted output of the training set and the true label, quantifying the model's prediction error; ③ Based on the gradient information of the training set loss function, solving the gradient of each network weight in the neural network through the time backpropagation algorithm, providing a quantitative basis for weight updates; ④ Using the Adam optimization algorithm for parameter iterative updates, first obtaining the statistical information of the training set gradient through first-order moment estimation and second-order moment estimation, and then adjusting the network weights of the neural network according to the parameter update formula, continuously reducing the prediction error of the neural network on the training set. At the same time, the validation set is used to monitor the model overfitting, and the training is terminated through an early stopping strategy. After training is completed, the performance of the neural network is verified using the test set, and finally, an LSTM multivariate time series prediction model adapted to the complaint prediction task is output.

9. A network-based operator home broadband complaint prediction system according to claim 8, characterized in that, The model optimization module specifically performs the following operations to optimize the training results of the LSTM multivariate time series prediction model: Based on the output results of the trained LSTM multivariate temporal prediction model on the validation and test sets, the initial single-layer LSTM network architecture is optimized into a 3-layer LSTM network structure. At the same time, combined with the characteristics of the complaint prediction task, the number of temporal attention heads and the number of hidden layer neurons are adjusted to improve the model's ability to capture complex temporal dependencies. To address the issues of insignificant accuracy improvement and a 4-fold decrease in inference speed on the validation and test sets for the adjusted LSTM multivariate time series prediction model, the improvement solution is to expand the dataset size over time. To address the issues of low precision and F1-score on the test set for the LSTM multivariate time series prediction model, the improvement solution is to add more feature variables and extend the data time span in subsequent model updates. Based on the upgraded 3-layer LSTM network, a temporal attention mechanism is introduced. Attention weights are obtained by calculating attention scores and normalizing them, and then weighted summation is used to obtain the attention context vector, which strengthens the weight of key temporal features. At the same time, the loss function is optimized by using a combination of mean squared error, mean absolute error and Huber loss to balance the influence of different error indicators and improve the model's adaptability to data distribution skew. After optimizing the loss function, the LSTM multivariate time series prediction model is retrained using the training set. The training process is monitored and early stopping is implemented using the validation set. Finally, the performance improvement of the optimized model is verified using the test set.

10. A neural network-based operator home broadband complaint prediction system according to claim 6, characterized in that, The system also includes: The report generation and feedback update module is used to interpret the prediction results output by the LSTM multivariate time series prediction model for business purposes, generate prediction reports and push them to relevant business departments; at the same time, it compares the actual complaint data with the prediction results, records the deviation, and provides feedback data for the continuous iterative optimization of the model.

Images