Indoor distribution fault identification method and system based on LSTM-Autoencoder
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-20
- Publication Date
- 2026-03-13
Smart Images
Figure CN121665280A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of communications, and specifically relates to an indoor distribution fault identification method and system based on LSTM-Autoencoder. Background Technology
[0002] Indoor distributed antenna systems (DAS) are a crucial component of mobile communication network coverage, and their stable operation directly impacts users' communication experience. However, existing fault detection methods for DAS generally suffer from drawbacks such as limited data dimensions, limited fault identification capabilities, insufficient positioning accuracy, and poor real-time performance. This is particularly true for complex nonlinear fault modes and latent faults, where traditional methods struggle to effectively detect and locate them. Currently, mainstream DAS fault detection methods primarily rely on single terminal measurement report (MR) data and simple threshold-based rules. These methods perform poorly when handling complex nonlinear fault modes (such as latent, intermittent, and compound faults), failing to effectively capture the complex correlations between signals. Furthermore, traditional methods typically only achieve coarse-grained fault location at the site or floor level, unable to provide precise physical location information, leading to low efficiency for on-site troubleshooting by maintenance personnel. In terms of data processing, existing technologies mostly employ batch offline processing modes, resulting in significant response delays between data acquisition and fault identification, failing to meet the high demands of real-time maintenance in the 5G-A and 6G eras. Simultaneously, static threshold-based decision mechanisms are sensitive to environmental fluctuations, exhibiting high false alarm rates and increasing unnecessary maintenance costs. While existing research has attempted to incorporate machine learning methods for fault detection, most still rely on supervised learning paradigms, requiring a large number of labeled fault samples for model training. However, in real-world operation and maintenance scenarios, fault samples are scarce and labeling costs are high, limiting the widespread application of these methods. In recent years, unsupervised deep learning models such as LSTM-Autoencoder have shown potential in time series anomaly detection, but their application in indoor distributed system fault detection is still in its early stages. Current technologies have not yet solved key issues such as multi-source data fusion, high-precision spatial mapping, real-time stream processing, and dynamic adaptive decision-making, making it impossible to build an end-to-end intelligent operation and maintenance closed loop. Summary of the Invention
[0003] To address the aforementioned issues, this application provides an indoor distribution system fault identification method and system based on LSTM-Autoencoder, which solves the problems of existing indoor distribution system fault detection methods, such as single data dimension, limited fault identification capability, insufficient positioning accuracy, poor real-time performance, and difficulty in effectively detecting and locating complex nonlinear fault modes and latent faults.
[0004] An indoor distribution fault identification method based on LSTM-Autoencoder includes: The raw data from the indoor distribution station is acquired and preprocessed. The raw data includes test logs, network configuration data, and performance data. The raw data to be processed is divided into datasets, and time series samples are constructed. A signal reconstruction model is established based on time series samples, and reconstruction error is generated; Fault determination is performed based on reconstruction error and preset dynamic threshold, and alarms are triggered based on the determination results; The fault alarm information is matched with a predefined grid data table, and the fault location information is output.
[0005] According to a specific embodiment of the present invention, the data preprocessing of the raw data from the indoor distribution station further includes: The raw data from the indoor distribution station is obtained, cleaned, and parsed to obtain the tower site code, operator, floor, and pixel coordinate information. The indoor map raster system divides the physical space into raster units and assigns a raster ID and latitude and longitude coordinates to the pixel coordinates of each raster unit.
[0006] According to a specific embodiment of the present invention, dividing the raw data into datasets and constructing time series samples further includes: The raw data to be processed is grouped using tower site code, operator, floor and grid ID as combined data, and the combined data of each grid cell is divided into training set, validation set and test set in chronological order; Based on the combined data of each grid cell in the dataset, a sequence of signal measurement values for a continuous time window is constructed, and time series samples of the training set, validation set, and test set are obtained respectively.
[0007] According to a specific embodiment of the present invention, establishing a signal reconstruction model based on time series samples and generating reconstruction errors further includes: The LSTM-Autoencoder deep learning framework is trained based on time series samples to obtain a signal reconstruction model; Reconstruction error is generated based on the original time series input to the signal reconstruction model and the reconstructed time series output.
[0008] According to a specific embodiment of the present invention, fault determination based on reconstruction error and a preset dynamic threshold, and triggering an alarm based on the determination result, further includes: The reconstruction error Eseq is compared with a preset dynamic threshold T. If Eseq > T, it is determined to be a fault. An alarm is triggered when a fault occurs and the fault condition persists for more than N consecutive time windows. The formula for calculating the preset dynamic threshold T is as follows: T = μ + k * σ In the formula, μ and σ are the mean and standard deviation of the errors of all normal sequences in the validation set, respectively, and k is an adjustable parameter.
[0009] According to a specific embodiment of the present invention, it further includes: Multiple fault decisions occurring in the same or adjacent grid cells within a preset time period are merged into a single fault alarm event, and the fault start time, duration, and maximum reconstruction error are recorded.
[0010] According to a specific embodiment of the present invention, matching fault alarm information with a predefined grid data table and outputting fault location information further includes: The system queries a predefined grid data table based on the grid ID in the fault alarm information and outputs the corresponding fault location information, which includes geographic coordinates and floor.
[0011] According to a specific embodiment of the present invention, it further includes: Based on the fault location information, the status of adjacent grids of the fault grid is analyzed. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault-affected area is generated, which includes the set of the fault grid and adjacent grids.
[0012] An indoor distribution fault identification system based on LSTM-Autoencoder includes: The data acquisition and processing module is used to acquire raw data from indoor distribution stations for data preprocessing. The raw data includes test logs, network configuration data, and performance data. The time-series feature construction module is used to partition the raw data and construct time-series samples. The signal reconstruction model building module is used to build a signal reconstruction model based on time series samples and generate reconstruction errors; The fault identification module is used to determine faults based on reconstruction errors and preset dynamic thresholds, and to trigger alarms based on the determination results. The fault location module is used to match fault alarm information with a predefined grid data table and output fault location information.
[0013] According to a specific embodiment of the present invention, the data acquisition and processing module further includes: The data processing module is used to acquire the raw data from the indoor distribution station, perform data cleaning and parsing, and obtain tower site code, operator, floor and pixel coordinate information; The raster mapping module is used to divide the physical space into raster units based on the indoor map raster system, and assign raster ID and latitude and longitude coordinates to the pixel coordinates of each raster unit.
[0014] According to a specific embodiment of the present invention, the time-series feature construction module further includes: The dataset partitioning module is used to group the raw data being processed using tower site codes, operators, floors, and grid IDs as combined data, and to divide the combined data of each grid cell into training set, validation set, and test set in chronological order. The time-series feature extraction module is used to construct a sequence of signal measurement values for a continuous time window based on the combined data of each grid cell in the dataset, thereby obtaining time-series samples for the training set, validation set, and test set, respectively.
[0015] According to a specific embodiment of the present invention, the signal reconstruction model establishment module further includes: The model building module is used to train the LSTM-Autoencoder deep learning framework based on time series samples to obtain a signal reconstruction model; The error calculation module is used to generate reconstruction error based on the original time series input to the signal reconstruction model and the reconstructed time series output.
[0016] According to a specific embodiment of the present invention, the fault identification module includes: The fault determination module is used to compare the reconstruction error Eseq with the preset dynamic threshold T. When Eseq > T, it is determined to be a fault. The alarm triggering module is used to trigger an alarm when a fault occurs and the fault state lasts for more than N consecutive time windows.
[0017] According to a specific embodiment of the present invention, the fault identification module further includes: The alarm aggregation module is used to merge multiple fault decisions that occur in the same grid cell or adjacent grid cells within a preset time into a single fault alarm event, and record the fault start time, duration, and maximum reconstruction error.
[0018] According to a specific embodiment of the present invention, it further includes: The fault impact area analysis module is used to analyze the status of adjacent grids of the fault grid based on the fault location information. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault impact area is generated, which includes the set of the fault grid and adjacent grids.
[0019] Compared with the prior art, this application has the following advantages: 1. Overcoming the limitations of linear fault detection: Existing technologies are mostly based on threshold rules or linear models, which can only detect obvious anomalies in a single indicator. The LSTM-Autoencoder deep learning model used in this invention has powerful nonlinear pattern learning capabilities. Through multi-source data fusion and time-series pattern analysis, it can intelligently learn normal signal patterns and capture complex and hidden correlations between signals, thereby effectively identifying complex fault types, including latent faults, intermittent faults, and compound faults, greatly improving the comprehensiveness and depth of fault detection.
[0020] 2. Achieving a leap from "detecting anomalies" to "locating faults": Traditional methods often only issue general alarms such as "an anomaly at a certain site." This invention, through indoor map grid management and multi-dimensional data correlation analysis, improves fault location accuracy from floor-level to meter-level grid accuracy. Furthermore, by analyzing the error propagation of adjacent grids, the scope of fault impact can be determined, providing maintenance personnel with precise repair guidance, greatly shortening on-site troubleshooting time and improving on-site troubleshooting efficiency.
[0021] 3. Significantly improves real-time detection: Existing technologies mostly involve batch offline processing, resulting in fault detection delays of up to several hours. This invention constructs an end-to-end real-time stream processing pipeline, achieving second-level response from data access and model inference to alarm output, reducing fault detection time from hours to seconds, and significantly reducing the duration of fault impact.
[0022] 4. Significantly reduce false alarm rate and operation and maintenance costs: Through dynamic adaptive threshold and alarm suppression aggregation mechanism, this invention can effectively distinguish between normal environmental fluctuations and real equipment failures, significantly reduce false alarms, avoid ineffective work of operation and maintenance personnel, and thus significantly reduce manual operation and maintenance costs.
[0023] 5. Solving the problem of dependence on labeled data: Traditional supervised learning methods require a large number of labeled fault samples, which are difficult to obtain in actual operation and maintenance. This invention adopts an unsupervised learning paradigm, which only requires a large amount of easily obtainable normal historical data to complete model training, perfectly solving the industry pain points of few fault samples and difficulty in labeling, and is more suitable for practical engineering applications.
[0024] 6. Possesses self-evolution and self-adaptation capabilities: Existing static models cannot adapt to network changes and require frequent manual parameter adjustments. This invention incorporates dynamic threshold updates and online / periodic model retraining mechanisms. Through automatic model updates and feedback learning mechanisms, it can continuously adapt to changes in the network environment, overcoming the drawbacks of traditional static rules requiring frequent manual adjustments. This enables the system to continuously learn new network patterns, automatically adapt to dynamic changes in the network environment, maintain long-term detection accuracy, and achieve intelligent operation and maintenance. After implementation, it is expected to reduce manual operation and maintenance costs, improve fault detection rates, increase network availability to 99.99%, and ultimately establish a new paradigm of intelligent operation and maintenance from "passive response" to "proactive prevention," providing technical support for network operation and maintenance in the 5G-A and 6G eras.
[0025] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0027] Figure 1 This is a flowchart of an indoor distribution fault identification method based on LSTM-Autoencoder according to an embodiment of the present invention.
[0028] Figure 2 This is a flowchart of a method for preprocessing raw data from an indoor distribution station according to an embodiment of the present invention.
[0029] Figure 3 This is a flowchart of a method for dividing raw data into datasets and constructing time series samples according to an embodiment of the present invention.
[0030] Figure 4 This is a flowchart of a method for establishing a signal reconstruction model based on time series samples according to an embodiment of the present invention.
[0031] Figure 5 This is a flowchart of a method for determining faults based on reconstruction errors and preset dynamic thresholds, and triggering alarms based on the determination results, according to an embodiment of the present invention.
[0032] Figure 6This is a flowchart of a method for matching fault alarm information with a predefined grid data table and outputting fault location information according to an embodiment of the present invention.
[0033] Figure 7 This is a schematic diagram of the overall intelligent operation and maintenance process according to an embodiment of the present invention.
[0034] Figure 8 This is an LSTM-Autoencoder model architecture diagram provided according to an embodiment of the present invention.
[0035] Figure 9 This is a structural diagram of an indoor distribution fault identification system based on LSTM-Autoencoder according to an embodiment of the present invention.
[0036] Figure 10 This is a structural diagram of a data acquisition and processing module provided according to an embodiment of the present invention.
[0037] Figure 11 This is a structural diagram of a timing feature construction module provided according to an embodiment of the present invention.
[0038] Figure 12 This is a structural diagram of a signal reconstruction model establishment module provided according to an embodiment of the present invention.
[0039] Figure 13 This is a structural diagram of a fault identification module provided according to an embodiment of the present invention.
[0040] Figure label: 01-Data Acquisition and Processing Module; 02-Time Series Feature Construction Module; 03-Signal Reconstruction Model Establishment Module; 04-Fault Identification Module; 05-Fault Location Module; 06-Fault Affected Area Analysis Module; 011-Data Processing Module; 012-Raster Mapping Module; 021 - Dataset partitioning module; 022 - Temporal feature extraction module; 031 - Model building module; 032 - Error calculation module; 041-Fault determination module; 042-Alarm triggering module; 043-Alarm aggregation module. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0042] Example 1 Additional aspects and advantages of embodiments of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of embodiments of the invention. Figures 1-8 This invention provides an indoor distribution fault identification method based on LSTM-Autoencoder, comprising: S1: Obtain raw data from indoor distribution stations and perform data preprocessing. The raw data includes test logs, network configuration data, and performance data.
[0043] S2: Divide the raw data into datasets and construct time series samples.
[0044] S3: Establish a signal reconstruction model based on time series samples and generate reconstruction error.
[0045] S4: Determine the fault based on the reconstruction error and the preset dynamic threshold, and trigger an alarm based on the determination result.
[0046] S5: Match fault alarm information with predefined grid data tables and output fault location information.
[0047] Specifically, step S1, acquiring the raw data from the indoor distribution station and performing data preprocessing, further includes: S11: Obtain the raw data from the indoor distribution station, perform data cleaning and parsing to obtain the tower site code, operator, floor, and pixel coordinate information.
[0048] In a specific embodiment of the present invention, raw data from indoor distribution stations is first collected through methods such as AR testing, including test logs, network configuration, and performance data. Then, the raw data is cleaned using the 3-sigma (Raida criterion). After data cleaning, the test logs are deeply analyzed. The analysis process is as follows: First, the tower site name, operator, and test type metadata are extracted from the filename. Then, the extrainfo field is parsed to obtain information such as signal strength, floor level, and pixel coordinates. Specifically, the system deploys a high-performance file monitoring service to monitor a specified file directory in real time. When a new CSV test log file is uploaded, the monitoring service immediately captures the event and drives the parsing engine to extract key business metadata such as tower site name, operator, and test type from the filename. The parsing engine reads the CSV file content line by line and performs deep analysis on the extrainfo field containing context to extract information including signal strength, floor level, and pixel coordinates.
[0049] In a specific embodiment of the present invention, after data parsing is completed, feature extraction and classification are performed on the parsed data. The extracted data features include numerical features, classification features, and time-series features. The numerical features include continuous numerical features such as NR serving cell RSRP, LTE serving cell RSRP, RSRP values of neighboring cells, and Wi-Fi signal strength. The classification features include discrete features such as PCI values and frequency points, processed through an embedding layer or one-hot encoding. The time-series features include time-series samples, each containing signal measurement values at continuous time steps. After feature extraction, each feature is standardized to eliminate the influence of units.
[0050] S12: Based on the indoor map raster system, the physical space is divided into raster units, and each raster unit is assigned a raster ID and latitude and longitude coordinates for its pixel coordinates.
[0051] In one specific embodiment of the present invention, the acquired pixel coordinates are matched with a predefined indoor map grid system. This indoor map grid system divides the physical space into grid units with meter-level precision and assigns a unique grid ID to each grid unit. Then, a coordinate mapping algorithm is used to convert the pixel coordinates into latitude and longitude coordinates.
[0052] In one specific embodiment of the present invention, after the grid mapping is completed, the mapping data of each grid cell is associated with the database. The database is queried using the tower site name as the keyword to supplement complete asset information for each test record, such as tower site code, equipment model, etc., thereby realizing the full-link association from signal measurement data to physical assets.
[0053] In a specific embodiment of the present invention, by constructing an end-to-end real-time stream processing pipeline to perform data cleaning, parsing and rasterization mapping processing on the raw data, second-level data access and millisecond-level model inference are achieved, reducing the fault detection time from hours to seconds, and significantly reducing the duration of fault impact.
[0054] Specifically, step S2, which involves partitioning the raw data and constructing time-series samples, further includes: S21: The raw data to be processed is grouped using tower site code, operator, floor and grid ID as combined data, and the combined data of each grid cell is divided into training set, validation set and test set in chronological order.
[0055] In one specific embodiment of this invention, an unsupervised learning paradigm is used for model training. Using fault data to construct the training set would cause the model to learn incorrect patterns, thus affecting the accuracy of fault detection. Therefore, this invention selects confirmed normal historical data to construct the training set to ensure that the LSTM-Autoencoder model learns the "baseline" pattern of indoor distribution station health. Validation and test sets are constructed using data labeled as normal / faulty for model tuning and final performance evaluation. Specifically, this invention divides the training, test, and validation sets according to chronological order, using data with earlier timestamps to construct the training set and data with later timestamps to construct the validation and test sets. Using historical data for training to detect future faults avoids the problem of "data leakage."
[0056] In a specific embodiment of the present invention, {tower site code, operator, floor and grid ID} are combined into a combined data, and the data of each grid unit is independently divided according to the combined data to ensure that the training data and test data in the same grid are continuous in time, rather than randomly interleaved, thereby ensuring the authenticity of the evaluation.
[0057] S22: Construct a sequence of signal measurement values for a continuous time window based on the combined data of each grid cell in the dataset, and obtain time series samples for the training set, validation set and test set respectively.
[0058] In a specific embodiment of the present invention, after completing the division of the training set, validation set, and test set, the present invention uses the sliding window technique to establish time series samples for each dataset. Each sample contains signal measurement values with continuous time steps. For the combined data of each grid cell, the data is sorted by time, and a sliding window (e.g., window length L=10, representing the measurement values of 10 consecutive time points) and step size (usually 1) are set. The window is slid only on the normal continuous time series to generate the time series samples of the training set. The window is slid on the mixed data composed of normal data and fault data to generate the time series samples of the test set / validation set. If a sequence contains data of at least one fault time point, the sequence is marked as a "fault sequence"; otherwise, it is marked as a "normal sequence".
[0059] For example, in the data preparation and filtering phase, the initial dataset is first filtered from the preprocessed data pool according to the following criteria: 1. Time range: Select a continuous time window (e.g., January 1, 2023 to June 30, 2023).
[0060] 2. Data Quality: Exclude all records with missing data, field errors, or obvious anomalies (e.g., RSRP consistently 0).
[0061] 3. Normal data filtering: Based on simple business rules (e.g., RSRP>-110 dBm and SINR>0), high-probability normal samples are filtered out to form the normal dataset Dnormal, and the remaining samples form the pending dataset Dpending.
[0062] During the data labeling phase, data records from the normal dataset Dnormal that occurred during the fault period and within the fault-affected area are first removed. Then, data in the pending dataset Dpending is labeled with a "fault" tag, thus confirming the labeling of the maintenance records.
[0063] During the dataset partitioning phase, for each unique combination Gi of {tower address code, operator, floor, grid ID}, all timestamp-sorted data Data_Gi within the corresponding grid cell are retrieved. Data from the first 70% time window of Data_Gi is selected, and only data labeled "normal" is retained to construct the training set (e.g., for 6 months of data, normal data from the first 4 months is selected). Data from the first 15% time window is selected to construct the validation set, which includes labeled normal and faulty data (e.g., data from the 5th month). Data from the last 15% time window is selected to construct the test set, which also includes labeled normal and faulty data (e.g., data from the 6th month).
[0064] Through the above process, training, validation, and test sets were obtained. The training set consists of a large number of clean and continuous normal time-series samples, used for unsupervised model training. The validation and test sets consist of mixed time-series samples composed of normal and faulty time-series samples. The validation set is used for model selection and hyperparameter tuning to ensure the model's generalization ability. The test set is used to simulate real-world scenarios and ultimately evaluate the model's ability to predict faults. This method of dividing the dataset can minimize data leakage and ensure the authenticity and reliability of the model evaluation results.
[0065] Specifically, step S3, which establishes a signal reconstruction model based on time series samples and generates reconstruction errors, further includes: S31: The LSTM-Autoencoder deep learning framework is trained based on time series samples to obtain a signal reconstruction model.
[0066] In a specific embodiment of this invention, the LSTM-Autoencoder deep learning framework is a deep learning model combining LSTM (Long Short-Term Memory) and an autoencoder. It is primarily used to process sequential data and extract features. LSTM is used to process data with strong time dependencies, while the autoencoder enables dimensionality reduction and reconstruction. This invention trains a signal reconstruction model by inputting normal time-series samples into the LSTM-Autoencoder deep learning framework. This model learns the multi-dimensional signal feature temporal patterns of each grid under normal conditions through an encoder-decoder structure. When abnormal or fault signals occur, the model cannot effectively reconstruct the signal because its pattern deviates from the normal state, resulting in a high reconstruction error. This error is used as an anomaly score for fault detection in indoor distribution stations.
[0067] In a specific embodiment of the present invention, the LSTM-Autoencoder model architecture is as follows: Figure 8As shown, the structure consists of an encoder, a bottleneck layer, a decoder, and an output layer. The encoder, a multi-layered Long Short-Term Memory (LSTM) network, is responsible for understanding and compressing the input sequence information. The LSTM layer captures long-term dependencies in the time series, the encoder compresses the input sequence into low-dimensional feature representations (latent variables), and the decoder reconstructs the latent variables into the output sequence. The LSTM layer comprises a first encoding LSTM layer and a second encoding LSTM layer. The first encoding LSTM layer receives the sequence output from the embedding layer or the original numerical feature sequence as a primary feature extractor, learning the short-term dependencies and basic features at each time step in the input sequence, outputting a hidden state sequence, and passing the hidden state h1 and cell state c1 of the last time step to the next layer. Its hidden state dimension is typically set to H1 (e.g., 128 dimensions). The second encoding LSTM layer receives the hidden state sequence output from the first encoding LSTM layer and acts as a high-level feature abstractor, learning longer-term and more complex temporal dependencies. It further abstracts and compresses the features output by the first encoding LSTM layer, using the hidden state h2 of its last time step as the context vector C of the entire input sequence. The dimension of C is H2 (e.g., 64-dimensional). The bottleneck layer is one or more fully connected layers connecting the encoder and decoder, and is crucial for the model to achieve effective feature compression and generalization. The fully connected layer forces the context vector C (dimension H2) output by the encoder to compress into a smaller space (dimension H2 / 2), forming an "information bottleneck." This process forces the model to discard redundant and noisy information, retaining only the most essential and representative features, effectively preventing overfitting and improving the model's generalization ability and robustness. The decoder is a mirror-symmetric structure of the encoder, and its role is to reconstruct the original sequence from the context vector C (or the vector processed by the bottleneck layer). The decoder includes a first decoding LSTM layer and a second decoding LSTM layer, where the initial hidden state and cell state of the first decoding LSTM layer are set to the final state of the encoder (or the state transformed by the bottleneck layer). At each time step, its input is the output of the previous time step or a repetition of the context vector. A common and efficient strategy is to repeat the context vector C L times, forming a fixed input sequence [C, C, ..., C]. This initiates the decoding process, generating initial features for the reconstructed sequence. The second decoding LSTM layer receives the hidden state sequence output from the first decoding LSTM layer. Further processing is performed to gradually recover the detailed structure and features of the time series. The output layer is a fully connected layer, which can be followed by a linear activation function. The second decoding LSTM layer outputs the hidden state hdec2t at each time step t. This maps the high-dimensional hidden states of the decoder LSTM back to the original feature space.It transforms hdec2t (dimension H1) into a vector with the exact same dimension D as the original input xt, i.e., the reconstructed value x't. The final output is the reconstructed complete sequence X', with the same dimension as the input X.
[0068] In a specific embodiment of the present invention, the specific process of training the LSTM-Autoencoder deep learning framework based on time series samples is as follows: In the LSTM-Autoencoder deep learning model, an indoor substation signal time series X = [x1, x2, ..., xL] of length L and feature dimension D is input. The encoder progressively reads the input sequence X and compresses it into a fixed-length context vector C. This vector is a condensed representation of the input sequence, containing the most informative pattern features. Subsequently, the decoder, using the context vector C as its initial state, progressively outputs a reconstructed sequence X' = [x'1, x'2, ..., x'L], aiming to make the reconstructed sequence X' infinitely close to the original input time series X. During training, the model learns reconstruction using a large amount of normal data. When a normal sequence is input, the model can reconstruct accurately with low reconstruction error. However, when an abnormal sequence containing faults is input, because its pattern deviates from the learned normal pattern, the model cannot reconstruct effectively, leading to a significant increase in reconstruction error.
[0069] S32: Generate reconstruction error based on the original time series input to the signal reconstruction model and the reconstructed time series output.
[0070] In a specific embodiment of the present invention, based on the original input sequence X and the reconstructed sequence X' output by the model, the mean squared error (MSE) is used as the error function to calculate the reconstruction error Eseq at each time step and each feature dimension in each sequence, including sequence-level error, time-step-level error, and feature-level error. Specifically, by calculating the average MSE of all elements in the entire sequence, a scalar value representing the overall degree of anomaly in the sequence is obtained, i.e., the sequence-level error. By averaging the MSE over all feature dimensions at each time step, the time-step-level error is obtained, which is used to identify the specific time point where the anomaly occurred. By averaging the MSE over all time steps for each feature dimension, the feature-level error is obtained, which is used to assist in the analysis of which signal indicators showed anomalies.
[0071] Specifically, step S4 involves determining the fault based on the reconstruction error and a preset dynamic threshold, and triggering an alarm based on the determination result, further including: S41: Compare the reconstruction error Eseq with the preset dynamic threshold T. If Eseq > T, it is determined to be a fault. The formula for calculating the preset dynamic threshold T is as follows: T = μ + k * σ(1) In the formula, μ and σ are the mean and standard deviation of the errors of all normal sequences in the validation set, respectively, and k is an adjustable parameter.
[0072] In one specific embodiment of the present invention, the system sets a dynamic, adaptive anomaly detection threshold (i.e., a preset dynamic threshold T) for each grid cell. This dynamic threshold enables accurate identification and avoids the limitations of using a fixed threshold. The threshold is calculated as follows: Before model deployment, the reconstruction error Eseq of all normal sequences in the validation set is used to calculate its mean μ and standard deviation σ. The threshold is set as T = μ + k * σ, where k is an adjustable parameter (usually 2.5~4) used to control the sensitivity of the system. The larger the value of k, the lower the false alarm rate, but the risk of missed alarms increases. After the system is running, the error values of sequences confirmed as normal can be added to the historical error set periodically (e.g., weekly), and the mean μ and standard deviation σ are updated on a rolling basis, thereby achieving dynamic adaptation of the threshold to adapt to the slow changes in the network environment.
[0073] In a specific embodiment of the present invention, a logical judgment is performed based on the reconstruction error Eseq of the current sequence and the preset dynamic threshold T of the corresponding raster unit. If Eseq > T, it is determined to be a fault; otherwise, it is determined to be normal.
[0074] S42: An alarm is triggered when a fault occurs and the fault state continues for more than N consecutive time windows.
[0075] S43: Merge multiple fault decisions that occur in the same grid cell or adjacent grid cells within a preset time into a single fault alarm event, and record the fault start time, duration, and maximum reconstruction error.
[0076] In a specific embodiment of the present invention, the fault decision in step S41 generates a large number of primary abnormal events. Directly outputting each fault event would lead to an alarm storm. To prevent repeated and frequent alarm storms for persistent single faults and improve the effectiveness of alarms, the present invention introduces alarm suppression and alarm aggregation functions to optimize the primary decision. First, the duration of the fault is filtered, and an alarm is triggered only when the abnormal state lasts for more than N consecutive time windows (e.g., more than 5 minutes). This filters out short-term faults caused by momentary jitter. Then, multiple abnormal decisions occurring in the same or adjacent grids within a short period of time are aggregated into a single fault alarm, and its start time, duration, and maximum abnormal score (i.e., reconstruction error) are recorded.
[0077] Specifically, step S5 matches the fault alarm information with the predefined grid data table and outputs fault location information, further including: S51: Query the predefined grid data table based on the grid ID in the fault alarm information, and output the corresponding fault location information, which includes geographical coordinates and floor.
[0078] In a specific embodiment of the present invention, upon receiving a fault alarm message, the abstract anomaly score (i.e., reconstruction error) is precisely located to its specific physical location by combining spatial topology data. Specifically, by inputting the grid ID corresponding to the fault alarm into the system, the system queries a predefined grid data table using the grid ID as the keyword to obtain all location information of the grid. The predefined grid data table stores metadata such as grid ID, tower site code, floor, pixel coordinates, geographic latitude and longitude coordinates of the grid center point, geographic fence polygon coordinates of the grid, and adjacent grid IDs.
[0079] S52: Analyze the status of adjacent grids of the fault grid based on the fault location information. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault-affected area is generated, which includes the set of the fault grid and adjacent grids.
[0080] In a specific embodiment of this invention, fault location not only precisely pinpoints the fault to a single grid cell, but also strives to analyze the spatial propagation and impact range of the fault. This invention defines the impact range of a fault by analyzing the states of adjacent grid cells of the faulty grid, ultimately generating a comprehensive "fault impact area" to provide maintenance personnel with a global fault view and guidance on repair priorities. Specifically, upon receiving a fault alarm message (which includes the faulty grid ID and the timestamp of the fault occurrence), the system determines all adjacent grid cells by querying the faulty grid ID, then searches for reconstruction error data (anomaly scores) within a time window close to the faulty grid's fault period. Subsequently, the reconstruction error of each adjacent grid cell is compared with its own independent preset dynamic threshold. If the reconstruction error Eseq of an adjacent grid cell is greater than its own preset dynamic threshold T, then the adjacent grid cell is determined to be in an abnormal state; otherwise, the adjacent grid cell is in a normal state. Finally, the system dynamically generates a fault impact area, which is logically a collection of geospatial objects, i.e., a collection containing the faulty grid cell and its adjacent grid cells in an abnormal state. If the reconstruction error of one or more adjacent grid cells exceeds a preset threshold, the fault is determined to have a spatial propagation effect.
[0081] Example 2 This invention also provides an indoor distribution fault identification system based on LSTM-Autoencoder, such as... Figures 9-13 As shown, it includes: The data acquisition and processing module 01 is used to acquire raw data from the indoor distribution station for data preprocessing. The raw data includes test logs, network configuration data, and performance data.
[0082] The time series feature construction module 02 is used to divide the raw data into datasets and construct time series samples.
[0083] The signal reconstruction model establishment module 03 is used to establish a signal reconstruction model based on time series samples and generate reconstruction errors.
[0084] The fault identification module 04 is used to determine faults based on reconstruction errors and preset dynamic thresholds, and to trigger alarms based on the determination results.
[0085] The fault location module 05 is used to match fault alarm information with a predefined grid data table and output fault location information.
[0086] The fault impact area analysis module 06 is used to analyze the status of adjacent grids of the fault grid based on the fault location information. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault impact area is generated, which includes the set of the fault grid and adjacent grids.
[0087] Specifically, the data acquisition and processing module 01 also includes: Data processing module 011 is used to acquire raw data from indoor distribution stations, perform data cleaning and parsing, and obtain tower site code, operator, floor and pixel coordinate information.
[0088] In a specific embodiment of the present invention, the data processing module 011 performs data cleaning on the acquired raw data. After data cleaning, the test logs are deeply analyzed. The analysis process is as follows: First, the tower site name, operator, and test type metadata are extracted from the file name. Then, the extrainfo field is parsed to obtain information such as signal strength, floor level, and pixel coordinates. Specifically, the system deploys a high-performance file monitoring service to monitor a specified file directory in real time. When a new CSV test log file is uploaded, the monitoring service immediately captures the event. The parsing engine first extracts key business metadata such as the tower site name, operator, and test type from the file name. The parsing engine reads the CSV file content line by line and performs deep analysis on the extrainfo field containing the context to extract information such as signal strength, floor level, and pixel coordinates.
[0089] The raster mapping module 012 is used to divide the physical space into raster units based on the indoor map raster system, and to assign a raster ID and latitude and longitude coordinates to the pixel coordinates of each raster unit.
[0090] In a specific embodiment of the present invention, the acquired pixel coordinates are matched with a predefined indoor map grid system by the grid mapping module 012. This indoor map grid system divides the physical space into grid units with meter-level precision and assigns a unique grid ID to each grid unit. Then, a coordinate mapping algorithm converts the pixel coordinates into latitude and longitude coordinates. After completing the grid mapping, the mapping data of each grid unit is associated with a database. The database is queried using the tower site name as the keyword to supplement complete asset information for each test record, such as tower site code and equipment model, thereby achieving a full-link association from signal measurement data to physical assets.
[0091] Specifically, the time-series feature construction module 02 also includes: The dataset partitioning module 021 is used to group the raw data being processed using tower site codes, operators, floors, and grid IDs as combined data, and to divide the combined data of each grid cell into training set, validation set, and test set in chronological order.
[0092] In a specific embodiment of the present invention, a dataset partitioning module 021 divides the processed raw data into training, testing, and validation sets according to chronological order. The training set is constructed using data with earlier timestamps, while the validation and testing sets are constructed using data with later timestamps. Historical data is used for training to detect future faults, avoiding the problem of "data leakage." {tower address code, operator, floor, and grid ID} are combined into a single dataset, and the data in each grid cell is independently partitioned according to this combined dataset. This ensures that the training and testing data within the same grid are temporally continuous, rather than randomly interleaved, thereby ensuring the authenticity of the evaluation.
[0093] The time-series feature extraction module 022 is used to construct a sequence of signal measurement values for a continuous time window based on the combined data of each grid cell in the dataset, thereby obtaining time-series samples for the training set, validation set, and test set, respectively.
[0094] In a specific embodiment of the present invention, the time series feature extraction module 022 is used to establish time series samples for each dataset. Each sample contains signal measurement values with continuous time steps. For the combined data of each grid cell, the data is sorted by time, and a sliding window (e.g., window length L=10, representing the measurement values of 10 consecutive time points) and step size (usually 1) are set. The window is slid only on normal continuous time series to generate time series samples for the training set. The window is slid on mixed data composed of normal data and fault data to generate time series samples for the test set / validation set. If a sequence contains data of at least one fault time point, the rectification sequence is marked as a "fault sequence"; otherwise, it is marked as a "normal sequence".
[0095] Specifically, the signal reconstruction model establishment module 03 also includes: Model building module 031 is used to train the LSTM-Autoencoder deep learning framework based on time series samples to obtain a signal reconstruction model.
[0096] In a specific embodiment of the present invention, the model building module 031 is used to train the LSTM-Autoencoder deep learning framework to obtain a signal reconstruction model. The model learns the multi-dimensional signal feature temporal pattern of each grid under normal conditions through the encoder-decoder structure. When abnormal or fault signals occur, the model cannot effectively reconstruct them because their patterns deviate from the normal state, resulting in a high reconstruction error. This error is used as an anomaly score for fault detection in indoor distribution stations.
[0097] In a specific embodiment of the present invention, the specific process of training the LSTM-Autoencoder deep learning framework based on the model building module 031 is as follows: In model building module 031, a time series of indoor substation signals, X = [x1, x2, ..., xL], with length L and feature dimension D, is input. The encoder of the LSTM-Autoencoder deep learning framework progressively reads the input sequence X and compresses it into a fixed-length context vector C. Subsequently, the decoder, using the context vector C as the initial state, progressively outputs a reconstructed sequence X' = [x'1, x'2, ..., x'L]. During training, the model learns reconstruction using a large amount of normal data. When a normal sequence is input, the model can reconstruct accurately with a low reconstruction error. However, when an abnormal sequence containing faults is input, because its pattern deviates from the learned normal pattern, the model cannot reconstruct effectively, resulting in a significant increase in reconstruction error.
[0098] Error calculation module 032 is used to generate reconstruction error based on the original time series input to the signal reconstruction model and the reconstructed time series output.
[0099] In a specific embodiment of the present invention, based on the original input sequence X and the reconstructed sequence X' output by the model, the error calculation module 032 calculates the reconstruction error Eseq at each time step and each feature dimension in each sequence, including sequence-level error, time-step-level error, and feature-level error. Specifically, by calculating the average MSE of all elements in the entire sequence, a scalar value representing the overall degree of anomaly in the sequence is obtained, i.e., the sequence-level error. By averaging the MSE over all feature dimensions at each time step, the time-step-level error is obtained, which is used to identify the specific time point where the anomaly occurred. By averaging the MSE over all time steps for each feature dimension, the feature-level error is obtained, which is used to assist in the analysis of which signal indicators showed anomalies.
[0100] Specifically, the fault identification module 04 includes: The fault determination module 041 is used to compare the reconstruction error Eseq with the preset dynamic threshold T. When Eseq > T, it is determined to be a fault.
[0101] In a specific embodiment of the present invention, the system sets a dynamic and adaptive anomaly detection threshold (i.e., a preset dynamic threshold T) for each grid cell. This dynamic threshold enables accurate identification and avoids the limitations of using a fixed threshold. The threshold is calculated as follows: Before model deployment, the reconstruction error Eseq of all normal sequences in the validation set is used to calculate its mean μ and standard deviation σ. The threshold is set as T = μ + k * σ, where k is an adjustable parameter (usually 2.5~4) used to control the sensitivity of the system. The larger the value of k, the lower the false alarm rate, but the risk of missed alarms increases. After the system is running, the error values of sequences confirmed as normal can be added to the historical error set periodically (e.g., weekly), and the mean μ and standard deviation σ are updated on a rolling basis, thereby achieving dynamic adaptation of the threshold to adapt to the slow changes in the network environment. The present invention uses a fault detection module 041, which performs logical judgment based on the reconstruction error Eseq of the current sequence and the preset dynamic threshold T of the corresponding grid cell. When Eseq > T, it is determined to be a fault; otherwise, it is determined to be normal.
[0102] The alarm triggering module 042 is used to trigger an alarm when a fault occurs and the fault state lasts for more than N consecutive time windows.
[0103] The alarm aggregation module 043 is used to merge multiple fault decisions that occur in the same grid cell or adjacent grid cells within a preset time into a single fault alarm event, and to record the fault start time, duration and maximum reconstruction error.
[0104] In a specific embodiment of the present invention, to prevent repeated and frequent alarm storms for persistent single faults and improve the effectiveness of alarms, the present invention introduces an alarm triggering module 042 and an alarm aggregation module 043 to optimize the initial decision-making. First, the alarm triggering module 042 filters the duration of the fault, triggering an alarm only when the abnormal state lasts for more than N consecutive time windows (e.g., more than 5 minutes), thus filtering out short-term faults caused by momentary jitter. Then, the alarm aggregation module 043 aggregates multiple abnormal decisions occurring in the same or adjacent grids within a short period into a single fault alarm, recording its start time, duration, and maximum abnormal score (i.e., reconstruction error).
[0105] Specifically, the fault location module 05 is used to query a predefined grid data table based on the grid ID in the fault alarm information and output the corresponding fault location information, which includes geographical coordinates and floor.
[0106] In a specific embodiment of the present invention, upon receiving a fault alarm message, the fault location module 05, combined with spatial topology data, precisely locates the abstract anomaly score (i.e., reconstruction error) to a specific location in physical space. Specifically, by inputting the grid ID corresponding to the fault alarm into the fault location module 05, the system queries a predefined grid data table using the grid ID as the keyword to obtain all location information of the grid. The predefined grid data table stores metadata such as grid ID, tower site code, floor, pixel coordinates, geographic latitude and longitude coordinates of the grid center point, geographic fence polygon coordinates of the grid, and adjacent grid IDs.
[0107] Specifically, the fault impact area analysis module 06 is used to analyze the status of adjacent grids of the fault grid based on the fault location information. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault impact area is generated, which includes the set of the fault grid and adjacent grids.
[0108] In a specific embodiment of this invention, fault location not only aims to precisely pinpoint the fault to a single grid, but also strives to analyze the spatial propagation and impact range of the fault. This invention uses a fault impact area analysis module 06 to analyze the status of adjacent grids of the faulty grid to define the fault impact range, ultimately generating a comprehensive "fault impact area" to provide maintenance personnel with a global fault view and guidance on repair priorities. Specifically, when the fault impact area analysis module 06 receives a fault alarm message (which includes the faulty grid ID and the timestamp of the fault occurrence), it determines all its adjacent grids by querying the faulty grid ID, then searches for reconstruction error data (anomaly score) within a time window close to the faulty grid's fault period. Subsequently, it compares the reconstruction error of each adjacent grid with its own independent preset dynamic threshold. If the reconstruction error Eseq of an adjacent grid is greater than its own preset dynamic threshold T, then the adjacent grid is determined to be in an abnormal state; otherwise, the adjacent grid is in a normal state. Finally, the system dynamically generates a fault impact area, which is logically a collection of geospatial objects, i.e., a collection containing the faulty grid and its adjacent grids in an abnormal state. If the reconstruction error of one or more adjacent grid cells exceeds a preset threshold, the fault is determined to have a spatial propagation effect.
[0109] In summary, the indoor distribution fault identification method and system based on LSTM-Autoencoder provided by this invention has the following advantages: 1. Overcoming the limitations of linear fault detection: Existing technologies are mostly based on threshold rules or linear models, which can only detect obvious anomalies in a single indicator. The LSTM-Autoencoder deep learning model used in this invention has powerful nonlinear pattern learning capabilities. Through multi-source data fusion and time-series pattern analysis, it can intelligently learn normal signal patterns and capture complex and hidden correlations between signals, thereby effectively identifying complex fault types, including latent faults, intermittent faults, and compound faults, greatly improving the comprehensiveness and depth of fault detection.
[0110] 2. Achieving a leap from "detecting anomalies" to "locating faults": Traditional methods often only issue general alarms such as "an anomaly at a certain site." This invention, through indoor map grid management and multi-dimensional data correlation analysis, improves fault location accuracy from floor-level to meter-level grid accuracy. Furthermore, by analyzing the error propagation of adjacent grids, the scope of fault impact can be determined, providing maintenance personnel with precise repair guidance, greatly shortening on-site troubleshooting time and improving on-site troubleshooting efficiency.
[0111] 3. Significantly improves real-time detection: Existing technologies mostly involve batch offline processing, resulting in fault detection delays of up to several hours. This invention constructs an end-to-end real-time stream processing pipeline, achieving second-level response from data access and model inference to alarm output, reducing fault detection time from hours to seconds, and significantly reducing the duration of fault impact.
[0112] 4. Significantly reduce false alarm rate and operation and maintenance costs: Through dynamic adaptive threshold and alarm suppression aggregation mechanism, this invention can effectively distinguish between normal environmental fluctuations and real equipment failures, significantly reduce false alarms, avoid ineffective work of operation and maintenance personnel, and thus significantly reduce manual operation and maintenance costs.
[0113] 5. Solving the problem of dependence on labeled data: Traditional supervised learning methods require a large number of labeled fault samples, which are difficult to obtain in actual operation and maintenance. This invention adopts an unsupervised learning paradigm, which only requires a large amount of easily obtainable normal historical data to complete model training, perfectly solving the industry pain points of few fault samples and difficulty in labeling, and is more suitable for practical engineering applications.
[0114] 6. Possesses self-evolution and self-adaptation capabilities: Existing static models cannot adapt to network changes and require frequent manual parameter adjustments. This invention incorporates dynamic threshold updates and online / periodic model retraining mechanisms. Through automatic model updates and feedback learning mechanisms, it can continuously adapt to changes in the network environment, overcoming the drawbacks of traditional static rules requiring frequent manual adjustments. This enables the system to continuously learn new network patterns, automatically adapt to dynamic changes in the network environment, maintain long-term detection accuracy, and achieve intelligent operation and maintenance. After implementation, it is expected to reduce manual operation and maintenance costs, improve fault detection rates, increase network availability to 99.99%, and ultimately establish a new paradigm of intelligent operation and maintenance from "passive response" to "proactive prevention," providing technical support for network operation and maintenance in the 5G-A and 6G eras.
[0115] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for indoor distribution system fault identification based on LSTM-Autoencoder, characterized in that, include: The raw data from the indoor distribution station is acquired and preprocessed. The raw data includes test logs, network configuration data, and performance data. The raw data to be processed is divided into datasets, and time series samples are constructed. A signal reconstruction model is established based on time series samples, and reconstruction error is generated; Fault determination is performed based on reconstruction error and preset dynamic threshold, and alarms are triggered based on the determination results; The fault alarm information is matched with a predefined grid data table, and the fault location information is output.
2. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 1, characterized in that, The data preprocessing of the raw data from the indoor distribution station further includes: The raw data from the indoor distribution station is obtained, cleaned, and parsed to obtain the tower site code, operator, floor, and pixel coordinate information. The indoor map raster system divides the physical space into raster units and assigns a raster ID and latitude and longitude coordinates to the pixel coordinates of each raster unit.
3. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 2, characterized in that, The step of partitioning the raw data and constructing time series samples further includes: The raw data to be processed is grouped using tower site code, operator, floor and grid ID as combined data, and the combined data of each grid cell is divided into training set, validation set and test set in chronological order; Based on the combined data of each grid cell in the dataset, a sequence of signal measurement values for a continuous time window is constructed, and time series samples of the training set, validation set, and test set are obtained respectively.
4. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 3, characterized in that, The establishment of a signal reconstruction model based on time series samples and the generation of reconstruction error further include: The LSTM-Autoencoder deep learning framework is trained based on time series samples to obtain a signal reconstruction model; Reconstruction error is generated based on the original time series input to the signal reconstruction model and the reconstructed time series output.
5. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 1, characterized in that, The step of determining the fault based on the reconstruction error and a preset dynamic threshold, and triggering an alarm based on the determination result, further includes: The reconstruction error Eseq is compared with a preset dynamic threshold T. If Eseq > T, it is determined to be a fault. An alarm is triggered when a fault occurs and the fault condition persists for more than N consecutive time windows. The formula for calculating the preset dynamic threshold T is as follows: T = μ + k * σ In the formula, μ and σ are the mean and standard deviation of the errors of all normal sequences in the validation set, respectively, and k is an adjustable parameter.
6. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 5, characterized in that, Also includes: Multiple fault decisions occurring in the same or adjacent grid cells within a preset time period are merged into a single fault alarm event, and the fault start time, duration, and maximum reconstruction error are recorded.
7. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 1, characterized in that, The step of matching fault alarm information with a predefined grid data table and outputting fault location information further includes: The system queries a predefined grid data table based on the grid ID in the fault alarm information and outputs the corresponding fault location information, which includes geographic coordinates and floor level.
8. The indoor distribution fault identification method based on LSTM-Autoencoder according to claim 7, characterized in that, Also includes: Based on the fault location information, the state of adjacent grids of the fault grid is analyzed. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault influence area is generated, wherein the fault influence area includes the set of the fault grid and adjacent grids.
9. An indoor distribution fault identification system based on LSTM-Autoencoder, characterized in that, include: The data acquisition and processing module is used to acquire raw data from the indoor distribution station and perform data preprocessing. The raw data includes test logs, network configuration data, and performance data. The time-series feature construction module is used to partition the raw data and construct time-series samples. The signal reconstruction model building module is used to build a signal reconstruction model based on time series samples and generate reconstruction errors; The fault identification module is used to determine faults based on reconstruction errors and preset dynamic thresholds, and to trigger alarms based on the determination results. The fault location module is used to match fault alarm information with a predefined grid data table and output fault location information.
10. The indoor distribution fault identification system based on LSTM-Autoencoder according to claim 9, characterized in that, The data acquisition and processing module also includes: The data processing module is used to acquire the raw data from the indoor distribution station, perform data cleaning and parsing, and obtain tower site code, operator, floor and pixel coordinate information; The raster mapping module is used to divide the physical space into raster units based on the indoor map raster system, and assign raster ID and latitude and longitude coordinates to the pixel coordinates of each raster unit.
11. The indoor distribution fault identification system based on LSTM-Autoencoder according to claim 9, characterized in that, The time-series feature construction module also includes: The dataset partitioning module is used to group the raw data being processed using tower site codes, operators, floors, and grid IDs as combined data, and to divide the combined data of each grid cell into training set, validation set, and test set in chronological order. The time-series feature extraction module is used to construct a sequence of signal measurement values for a continuous time window based on the combined data of each grid cell in the dataset, thereby obtaining time-series samples for the training set, validation set, and test set, respectively.
12. The indoor distribution fault identification system based on LSTM-Autoencoder according to claim 9, characterized in that, The signal reconstruction model establishment module also includes: The model building module is used to train the LSTM-Autoencoder deep learning framework based on time series samples to obtain a signal reconstruction model; The error calculation module is used to generate reconstruction error based on the original time series input to the signal reconstruction model and the reconstructed time series output.
13. The indoor distribution fault identification system based on LSTM-Autoencoder according to claim 9, characterized in that, The fault identification module includes: The fault determination module is used to compare the reconstruction error Eseq with the preset dynamic threshold T. When Eseq > T, it is determined to be a fault. The alarm triggering module is used to trigger an alarm when a fault occurs and the fault state lasts for more than N consecutive time windows.
14. The indoor distribution fault identification system based on LSTM-Autoencoder according to claim 13, characterized in that, The fault identification module also includes: The alarm aggregation module is used to merge multiple fault decisions that occur in the same grid cell or adjacent grid cells within a preset time into a single fault alarm event, and record the fault start time, duration, and maximum reconstruction error.
15. The indoor distribution fault identification system based on LSTM-Autoencoder according to claim 9, characterized in that, Also includes: The fault impact area analysis module is used to analyze the status of adjacent grids of the fault grid based on the fault location information. When the reconstruction error of adjacent grids is greater than a preset threshold, a fault impact area is generated, wherein the fault impact area includes the set of the fault grid and adjacent grids.