Universal intelligent state recognition system and method based on LightGBM

Abstract

The invention provides a universal intelligent state recognition system and method based on LightGBM, and the system comprises a data preprocessing and sample dividing module, a model training and evaluation module, a state recognition and visual analysis module, and a universal adaptation and lightweight deployment engine. According to the method, data-driven recessive feature self-mining can be realized, multi-parameter coupling features can be automatically extracted from massive time series data, and the limitation of artificial feature design is avoided; quick response of on-site lightweight deployment is guaranteed, a model structure and a deployment mode are optimized, and lightweight model training and reasoning can be realized on a common notebook computer; the expert experience implicit migration is realized, and the diagnosis experience of a designer is solidified into reusable model knowledge through historical fault annotation data, so that the dependence of field guarantee on artificial expert decision can be reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent equipment condition monitoring and fault diagnosis, and specifically relates to a general intelligent state recognition system and method based on Light Gradient Boosting Machine (LightGBM), which is particularly suitable for efficiently performing state recognition and fault diagnosis on the massive bus data processing scenarios of complex equipment such as vehicles, and has the capabilities of lightweight deployment, fast calculation, and visual decision support. Background Art

[0002] In the field of equipment condition monitoring, the current mainstream technical routes have the following limitations:

[0003] 1. The inherent defects of traditional Built-In Test (BIT) systems: Most of the built-in test modules commonly deployed in existing equipment adopt single-parameter threshold alarms or simple multi-parameter logic combination rules (such as "temperature > X °C and rotational speed < Y rpm" to trigger early warnings). Although they can identify typical faults with clear characteristics, when facing hidden faults with multi-parameter coupling, due to the lack of cross-dimensional correlation analysis capabilities, complex working condition faults still rely on designers to go to the site to retrieve bus historical data or perform manual diagnosis through offline analysis tools, resulting in a long average fault handling time.

[0004] 2. The implementation bottleneck of traditional machine learning methods - high dependence on feature engineering: For scenarios with clear fault characteristics, automatic recognition can be achieved by extracting time-frequency domain features and combining algorithms such as SVM and random forests. However, when the fault shows hidden associations of multi-source signals, the efficiency of manual feature extraction is low and it is difficult to cover key information dimensions.

[0005] 3. Deployment obstacles for ensuring on-site intelligent diagnosis: High deployment cost of deep learning models: Although deep networks such as CNN and LSTM can automatically mine deep features of data, model inference requires GPU acceleration, while only ordinary laptops can be equipped at the equipment support site, which cannot meet the deployment requirements of deep learning models.

[0006] Therefore, there is an urgent need in this field for a general intelligent state recognition solution that can both automatically mine complex features and be efficiently deployed and run on lightweight hardware. Summary of the Invention

[0007] (1) Technical Problems to be Solved

[0008] To address the shortcomings of traditional BIT systems in identifying hidden faults coupled with multiple parameters, the high dependence of traditional machine learning on manual feature engineering, and the hardware limitations in ensuring the deployment of deep learning models in the field, this invention proposes a general intelligent condition recognition system and method based on LightGBM. This system aims to solve the problems of weak hidden fault identification capabilities of traditional BIT systems, high dependence of traditional machine learning on feature engineering, and high deployment costs of deep learning models. It enables intelligent condition recognition that requires no manual feature engineering, can run on conventional computing equipment, and is adaptable to various equipment types.

[0009] (II) Technical Solution

[0010] To address the aforementioned technical problems, this invention proposes a general intelligent state recognition system based on LightGBM. This system includes a data preprocessing and sample partitioning module, a model training and evaluation module, a state recognition and visualization analysis module, and a generalized adaptation and lightweight deployment engine.

[0011] The data preprocessing and sample partitioning module is used to partition and flatten the equipment bus time series data through a multi-scale sliding window, generate standardized feature vectors, and automatically extract or map state labels from the data source to form a labeled training sample set.

[0012] The model training and evaluation module is used to load labeled training sample sets, perform data cleaning, label encoding and feature standardization, train the LightGBM multi-classification model, evaluate model performance and output visual evaluation results;

[0013] The status recognition and visualization analysis module is used to load newly acquired equipment bus data, use a pre-trained LightGBM model to predict status, generate a text report containing fault status and its time period, and perform visualization analysis.

[0014] The generalized adaptation and lightweight deployment engine is used to decouple the system from specific equipment parameters, supports dynamic adaptation to different fault types through configuration files, and ensures the efficient operation of the system on conventional computing devices.

[0015] Furthermore, the data preprocessing and sample partitioning module uses at least two sliding windows of different time lengths to partition the input time series data; flattens the multi-parameter data in each window into one-dimensional feature vectors in time step order; automatically associates the corresponding state labels to each feature vector by parsing the file name or metadata of the input data; and stores the generated labeled sample set in two formats simultaneously: one for manual review and the other for direct model call.

[0016] Furthermore, the model training and evaluation module preprocesses the loaded sample data, including removing time columns, filtering non-numerical columns, and performing data type conversion; it uses a normalizer to standardize the feature data and persists the normalizer parameters; it uses LGBMClassifier under the LightGBM framework to train a multi-classification model and saves the trained model; it automatically generates model evaluation results, including one or more of the following: classification report, confusion matrix visualization, and feature importance ranking graph; and it performs a horizontal performance comparison of models generated by sliding windows at different scales.

[0017] Furthermore, the state recognition and visualization analysis module uses a sliding window with the same parameters as the training phase to divide the new data; loads a pre-trained normalizer to standardize the newly divided feature vectors; loads a pre-trained LightGBM model for inference to obtain predicted labels and map them to readable state names; generates a detailed text report recording the identified states and the specific time periods in which they occurred; and uses a dimensionality reduction algorithm to visualize the high-dimensional feature data and generate a state distribution map.

[0018] Furthermore, the generalized adaptation and lightweight deployment engine includes a parameter-independent processing unit and a dynamic label mapping unit. The parameter-independent processing unit treats the input bus parameters as time-series variables without physical meaning, and forms a standardized feature matrix that does not depend on the specific number and meaning of parameters through unified window partitioning and flattening operations. The dynamic label mapping unit manages the correspondence between fault types and digital labels by maintaining an external configuration file, and only needs to update the configuration file when adapting to new scenarios.

[0019] Furthermore, this invention also proposes a general intelligent state identification method based on the aforementioned general intelligent state identification system. The entire processing flow of this general intelligent state identification method is independent of the meaning and quantity of the physical parameters of the input data, and only adapts to different equipment or fault types by modifying the tag mapping configuration file. The method includes the following steps:

[0020] A standardized training sample set is generated by dividing, flattening and labeling historical equipment bus data through a multi-scale sliding window.

[0021] The training sample set is cleaned and standardized, the LightGBM multi-class classification model is trained, and the model is evaluated and saved.

[0022] For newly acquired equipment bus data, a sliding window with the same parameters as in the training phase is used for partitioning and standardization;

[0023] Use a pre-trained LightGBM model to perform state recognition and prediction on new data;

[0024] Generate status identification reports and visual analysis results.

[0025] Furthermore, the steps for generating a standardized training sample set include: extracting numeric labels from the filenames of the original data using regular expressions and associating them with feature vectors segmented by a sliding window.

[0026] Furthermore, the steps for training the LightGBM multi-class classification model include: encoding the class labels using LabelEncoder and standardizing the feature data using StandardScaler, while persisting the encoder and standardizer parameters for subsequent inference.

[0027] Furthermore, the steps for generating visualization analysis results include: using the t-SNE algorithm to reduce the dimensionality of the high-dimensional prediction results features to two-dimensional or three-dimensional space and plotting them to visualize the clustering of different state categories.

[0028] (III) Beneficial Effects

[0029] This invention proposes a general intelligent state recognition system and method based on LightGBM. The general intelligent state recognition system includes a data preprocessing and sample partitioning module, a model training and evaluation module, a state recognition and visualization analysis module, and a generalized adaptation and lightweight deployment engine. This invention can achieve data-driven implicit feature self-mining, automatically extracting multi-parameter coupled features from massive time-series data, avoiding the limitations of manual feature design; ensuring rapid response of lightweight deployment in the field, with optimized model structure and deployment methods, enabling lightweight model training and inference on ordinary laptops; realizing implicit transfer of expert experience, solidifying the diagnostic experience of designers into reusable model knowledge through historical fault annotation data, reducing the dependence on human expert decision-making in field support.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. Highly versatile and adaptable to multiple scenarios

[0032] By using parameter independence processing and dynamic label mapping mechanism, the model and physical object are completely decoupled, and can be seamlessly applied to various equipment types such as different parts and systems of armored vehicles; it supports the input of any number of bus parameters, eliminating the need to redevelop models for different vehicle models or fault scenarios, and significantly reducing system adaptation costs.

[0033] 2. Lightweight deployment, highly adaptable to on-site conditions

[0034] The core algorithm and its dependent libraries are small in size and can run smoothly on regular laptops; they support on-site support personnel in quickly identifying the status of the situation.

[0035] 3. High degree of intelligence, reducing reliance on manual labor.

[0036] It automatically completes the entire process from data loading and sample segmentation to fault identification without manual intervention; it provides visualized test reports and dimensionality reduction analysis charts to help on-site support personnel quickly locate faults and formulate maintenance strategies. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the five types of state data in the embodiment;

[0038] Figure 2 This is a schematic diagram of the training data file generated after sample partitioning and label injection.

[0039] Figure 3 This is a diagram showing the model file and normalizer file saved after training is complete.

[0040] Figure 4 A schematic diagram of the confusion matrix for model evaluation under different sample partitioning scales;

[0041] Figure 5 A schematic diagram illustrating the feature importance of the model evaluation under different sample partitioning scales;

[0042] Figure 6 A diagram illustrating the performance comparison of the model under different sample partitioning scales;

[0043] Figure 7 A schematic diagram of the verification output file generated after state recognition of new data;

[0044] Figure 8 A visualization diagram of t-SNE dimensionality reduction for new data prediction results. Detailed Implementation

[0045] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0046] This embodiment proposes a general intelligent state recognition system based on LightGBM, including a data preprocessing and sample partitioning module, a model training and evaluation module, a state recognition and visualization analysis module, and a generalized adaptation and lightweight deployment engine.

[0047] The data preprocessing and sample partitioning module is used to standardize the bus time-series data of the equipment. This module partitions the input data using multi-scale sliding windows (e.g., 15 seconds, 10 seconds, 5 seconds), flattening the multi-parameter time-series data within each window into standardized feature vectors. The data preprocessing and sample partitioning module also includes a label mapping unit, which automatically extracts fault type labels from data file names or metadata and associates them with the corresponding feature vectors to form a standardized training sample set. The output is a data file that can be directly used by the model (e.g., a Pickle file) and a file that can be manually viewed (e.g., an Excel file).

[0048] The model training and evaluation module is used to build, train, and evaluate LightGBM multi-class classification models. This module loads preprocessed sample data and performs data cleaning (such as removing timestamps, filtering non-numerical data, and forcibly converting data types), label encoding, and feature standardization. It builds and trains the LightGBM classifier (LGBMClassifier), saving the model file and the corresponding standardizer (StandardScaler) metadata upon completion. This module also automatically evaluates the model, outputting a classification report (accuracy, precision, recall, F1 score), a confusion matrix visualization, a feature importance ranking chart, and comparing the performance of models trained with different window sizes, providing a basis for selecting the optimal window size.

[0049] The Status Recognition and Visualization Analysis module is used for real-time or offline status recognition of newly acquired equipment data. This module loads a pre-trained LightGBM model and a normalizer, employing the sliding window method consistent with the training phase to partition and normalize the new data. Then, it performs model inference to predict the status label for each time window. The prediction results are mapped to readable fault status names, generating a detailed text report containing the fault status and its corresponding time period. Simultaneously, this module utilizes dimensionality reduction algorithms such as t-SNE to project high-dimensional features into a two-dimensional space and plot them, visually demonstrating the clustering of different statuses to assist human decision-making. All raw data, prediction results, and visualizations are saved and output.

[0050] Generalized adaptation and lightweight deployment engine: This is the core innovation of this invention, including a parameter-independent processing unit and a dynamic tag mapping unit.

[0051] The parameter independence processing unit is used to treat the input bus parameters as pure time-series variables, discarding the prior assumptions about their physical meaning. Through a unified sliding window and flattening operation, it converts any number of sensor data with any physical meaning into a feature matrix in a fixed format, thereby decoupling the model from the specific equipment type.

[0052] The dynamic tag mapping unit can maintain the correspondence between fault types and digital tags through an external configuration file. When adapting to new equipment or new fault types, only this configuration file needs to be updated, without modifying the core algorithm code.

[0053] The lightweight runtime environment is as follows: the entire system is built on a lightweight Python library with a small total dependency library size (<500MB), requires no GPU acceleration, and can efficiently complete training and inference tasks on a regular laptop with 8GB of memory and an Intel i5 processor.

[0054] The intelligent state recognition method using the above-mentioned general intelligent state recognition system mainly includes the following steps:

[0055] Through the data preprocessing and sample partitioning module, the historical bus data is processed and labeled using a multi-scale sliding window to generate a standardized training sample set.

[0056] The model training and evaluation module cleans, encodes, and standardizes the samples to train the LightGBM multi-class classification model, and then evaluates and compares its performance.

[0057] Through the state recognition and visualization analysis module, new data is loaded and preprocessed in the same way, and a pre-trained model is used to predict the state and generate visualization reports and charts.

[0058] Through a generalized adaptation mechanism, the above process is ensured to be independent of the specific number and physical meaning of sensor parameters, and can be adapted to new application scenarios simply by modifying the tag configuration file.

[0059] This invention proposes a sample partitioning and label mapping method based on a multi-scale sliding window for intelligent preprocessing of equipment bus data. First, file existence verification and forced data format alignment ensure the validity and consistency of input data (e.g., Broken Shaft 1.xlsx, Overvoltage 2.xlsx, etc.). Simultaneously, regular expressions are used to extract numerical labels from filenames (e.g., "Normal 4.xlsx" → Label 4), establishing a direct mapping between filenames and fault types. For fault characteristics at different time scales, three window sizes (15 seconds, 10 seconds, and 5 seconds) are used for sample partitioning. Fixed-step sliding and sample integrity verification generate standardized feature vectors and append labels, forming a standard data format of [feature vector, label]. Finally, a structured column name system (e.g., t1_FeatureA, t2_FeatureA, ..., tN_FeatureB) is automatically generated based on the original feature column names and window size. The results are stored in Excel and Pickle files, supporting manual verification and direct model access, respectively, providing high-quality input data for subsequent intelligent recognition model training.

[0060] This invention proposes a multi-scale intelligent identification model training and evaluation method based on LightGBM for fault classification and performance optimization of equipment bus data. The method includes the following steps: First, preprocessed multi-scale sample data (15-second, 10-second, and 5-second window sizes) is loaded. A rigorous data preprocessing procedure ensures the quality of the input data, including detecting and removing time columns, filtering non-numerical columns, and performing forced conversion of the numerical types of the remaining columns to avoid model training anomalies caused by inconsistent data types. Then, a label encoder is used to digitally map fault categories, ensuring the model can handle categorical labels. Standardization is then used to eliminate differences in feature dimensions. Simultaneously, the standardizer and its metadata (including feature names, sample sizes, etc.) are persistently stored for direct retrieval during subsequent inference, ensuring data processing consistency.

[0061] During the model training phase, a multi-classification model (LGBMClassifier) ​​based on LightGBM was constructed. Key hyperparameters were set (e.g., number of leaf nodes num_leaves = 31, learning rate = 0.05, number of iterations n_estimators = 100, etc.), and feature names were passed in to support feature importance analysis. After model training, it and related metadata (e.g., expected feature dimensions, sample size, etc.) were saved as separate files to ensure model reusability and ease of deployment. During the model evaluation phase, predictions were performed using test set data, generating detailed classification reports (including metrics such as accuracy, precision, recall, and F1 score). A confusion matrix (ConfusionMatrixDisplay) was used to visually demonstrate the model's classification performance across different fault categories, supporting Chinese labels for easy manual interpretation. Furthermore, LightGBM's built-in feature importance analysis function was used to plot a gain-based feature importance graph, identifying the key features that contribute most to the classification results and providing guidance for subsequent feature engineering optimization.

[0062] Finally, training results for different sample sizes (15 seconds, 10 seconds, and 5 seconds) are compared horizontally, generating model performance comparison reports and visualization charts (such as bar charts of accuracy and macro-average F1 score). This comprehensively evaluates the impact of window size on model performance and provides data support for selecting the optimal window size in practical applications. All intermediate results (such as confusion matrices, feature importance maps, and performance comparison charts) are saved as high-resolution images for easy subsequent analysis and decision-making. Through the above process, this invention achieves full automation from data loading, preprocessing, model training to performance evaluation, providing an efficient and reliable solution for intelligent fault identification of equipment bus data.

[0063] This invention proposes a method for fault identification and visualization analysis of equipment bus data based on a pre-trained model, used to achieve rapid status assessment and decision support for new vehicle models. First, the bus data of the new vehicle model to be tested (e.g., vehicle #3.xlsx) is loaded, ensuring that the data includes time columns and feature columns consistent with the training data. Standardized feature vectors are generated using sample partitioning logic completely consistent with the training phase (including fixed-step sliding window, sample integrity verification, etc.). Then, a pre-trained LightGBM model and its corresponding normalizer are loaded to standardize the feature data, ensuring that the input data distribution is consistent with the training data. After the model infers from the new data, it generates prediction labels and maps them to readable fault state names (e.g., "controller overheating," "broken axle," etc.), while recording the time range corresponding to each sample, forming a timestamped prediction result.

[0064] To support rapid decision-making by on-site support personnel, the system automatically generates detailed test reports, including the fault status and its corresponding time period (e.g., "2023-10-01 10:00:00~2023-10-01 10:15:00, controller overheating"), and saves them as text files. Furthermore, the t-SNE dimensionality reduction algorithm is used to map high-dimensional features to a two-dimensional space, creating a distribution map of the prediction results. This visually displays the clustering of different fault states, aiding in the identification of abnormal patterns. All prediction results (including original features, time range, prediction labels, and status) are stored in Excel files, supporting subsequent manual review and analysis. Through this process, the present invention achieves a one-stop fault identification function from data loading, sample partitioning, model inference to result visualization, providing an efficient and reliable intelligent decision support tool for equipment support on-site.

[0065] This invention proposes a generalized intelligent state recognition method decoupled from physical objects. Through parameter independence processing and an adaptive sample partitioning mechanism, a universal recognition framework is constructed for any equipment type and fault scenario. This method abandons the prior assumptions of traditional methods regarding specific parameter combinations, treating bus parameters as a set of time-series variables without physical meaning. Regardless of the number of input parameters (e.g., one model has 10 sensor parameters, another has 20) or their physical meaning (temperature, pressure, speed, etc.), the system performs feature flattening and vectorization processing through a standardized sliding window (15 / 10 / 5 seconds) to form a unified [time step × number of parameters] feature matrix, ensuring the model is completely independent of the physical characteristics of specific equipment. Simultaneously, the correspondence between fault types and labels is dynamically maintained through an external configuration file. When applied to a new vehicle model, only the label mapping table needs to be updated to match the newly defined fault states, without modifying the core algorithm. This solution achieves universality across the entire process from data preprocessing and feature engineering to model inference through a closed-loop technology of unified feature space, dynamic labeling system, and adaptive processing engine. This allows the same system to be seamlessly applied to different fault identification scenarios of different parts and systems of armored vehicles, significantly reducing the deployment threshold and maintenance cost of intelligent diagnostic systems.

[0066] This invention proposes a lightweight deployment solution for field support. By optimizing model architecture and resource management, it enables the efficient operation of a general-purpose intelligent identification model on a regular laptop. The core system components (including data preprocessing, model inference, and result visualization modules) are all built on a lightweight Python environment, with dependent libraries smaller than 500MB, allowing for smooth operation on laptops with 8GB of RAM and an Intel i5 processor. For newly acquired bus data, field support personnel only need to import the data into the system (Excel format), and the system automatically completes the entire process from data loading and sample segmentation to fault identification. This lightweight deployment solution enables seamless migration of intelligent diagnostic capabilities from the laboratory to the field, significantly improving the timeliness and convenience of equipment status monitoring.

[0067] Example: Fault Identification of the Main Generator Controller of an Armored Vehicle

[0068] Data Preparation: Data was collected from the main generator controllers of six armored vehicles of three different models over a period of 10 months. It included 15 sensor channels with second-level sampling, encompassing continuous signals such as engine temperature, input voltage, and generator current, as well as discrete status signals such as chassis low-voltage insulation alarms and controller over-temperature. Based on historical fault records and expert experience, the data was categorized into five states: "Controller Over-temperature" (label 0), "Broken Axle" (label 1), "Overvoltage" (label 2), "Undervoltage" (label 3), and "Normal" (label 4). Redundant channels were removed, retaining 10 key sensor signals, which were then categorized and stored according to health status, forming a structured dataset (e.g., ...). Figure 1 (As shown).

[0069] Sample Segmentation and Label Injection: The data preprocessing and sample segmentation module is run. This module reads the aforementioned data file and extracts the label "4" from the filename "normal4.xlsx" using regular expressions. Subsequently, it iterates through the data using a fixed-step sliding window with three window sizes: 15 seconds, 10 seconds, and 5 seconds. All sensor data within each window is flattened into a feature vector, and the label obtained from the filename is appended, generating a standard feature vector, a label feature vector, and label samples. The final output is an Excel file and a Pickle file directly called by the model (e.g., ...). Figure 2 (As shown).

[0070] Model Training: Run the model training and evaluation module. This module loads training data in Pickle format, automatically removes the time column, filters to ensure all columns are numeric and forces data type conversion. Labels are encoded into numbers using LabelEncoder, features are standardized using StandardScaler, and the standardizer is saved. The model is trained using LGBMClassifier (with num_leaves=31, learning_rate=0.05, n_estimators=100), and the trained model is saved (e.g., ...). Figure 3 (As shown). After training, the module automatically evaluates on the test set and synchronously outputs a confusion matrix (as shown). Figure 4 (as shown), feature importance map (e.g.) Figure 5 (as shown) and a performance comparison chart of different window models (as shown) Figure 6 (As shown).

[0071] On-site verification: Acquire and import newly collected bus data (870,770 records) from the 6th vehicle. Operational status identification and visualization analysis module. This module loads new data (e.g., "Vehicle No. 3.xlsx"), performs partitioning and standardization using the same 10-second window as during training (calling the saved normalizer). Then, it loads the pre-trained 10-second window model for inference, generates a test report, plots a t-SNE dimensionality reduction graph, and saves the prediction results, such as... Figure 7 As shown, taking 10 samples as an example, the dimensionality reduction graph is displayed as follows. Figure 8 As shown.

[0072] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A general-purpose intelligent state recognition system based on LightGBM, characterized in that, The general intelligent state recognition system includes a data preprocessing and sample segmentation module, a model training and evaluation module, a state recognition and visualization analysis module, and a generalized adaptation and lightweight deployment engine; among which... The data preprocessing and sample segmentation module is used to segment and flatten the equipment bus time series data through a multi-scale sliding window, generate standardized feature vectors, and automatically extract or map state labels from the data source to form a labeled training sample set. The model training and evaluation module is used to load the labeled training sample set, perform data cleaning, label encoding and feature standardization, train the LightGBM multi-classification model, evaluate the model performance and output the visual evaluation results. The status recognition and visualization analysis module is used to load newly acquired equipment bus data, use a pre-trained LightGBM model to predict status, generate a text report containing fault status and its time period, and perform visualization analysis. The generalized adaptation and lightweight deployment engine is used to decouple the system from specific equipment parameters, supports dynamic adaptation to different fault types through configuration files, and ensures efficient operation of the system on conventional computing devices.

2. The general intelligent state recognition system based on LightGBM as described in claim 1, characterized in that, The data preprocessing and sample partitioning module uses at least two sliding windows of different time lengths to partition the input time series data; it flattens the multi-parameter data in each window into a one-dimensional feature vector in time step order; it automatically associates the corresponding state label to each feature vector by parsing the file name or metadata of the input data; and it stores the generated labeled sample set in two formats: one for manual review and the other for direct model call.

3. The general intelligent state recognition system based on LightGBM as described in claim 1, characterized in that, The model training and evaluation module preprocesses the loaded sample data, including removing time columns, filtering non-numerical columns, and performing data type conversion; it uses a normalizer to standardize the feature data and persists the normalizer parameters; it uses LGBMClassifier under the LightGBM framework to train a multi-classification model and saves the trained model; it automatically generates model evaluation results, including one or more of the following: classification report, confusion matrix visualization, and feature importance ranking graph. A cross-sectional performance comparison was conducted on models generated by sliding windows of different scales.

4. The general intelligent state recognition system based on LightGBM as described in claim 1, characterized in that, The state recognition and visualization analysis module uses a sliding window with the same parameters as the training phase to divide the new data; loads a pre-trained normalizer to standardize the newly divided feature vectors; loads a pre-trained LightGBM model for inference to obtain predicted labels and map them to readable state names; generates a detailed text report to record the identified states and the specific time periods in which they occurred; and uses a dimensionality reduction algorithm to visualize the high-dimensional feature data and generate a state distribution map.

5. The general intelligent state recognition system based on LightGBM as described in claim 1, characterized in that, The generalized adaptation and lightweight deployment engine includes a parameter-independent processing unit and a dynamic label mapping unit. The parameter-independent processing unit treats the input bus parameters as time-series variables without physical meaning, and forms a standardized feature matrix that does not depend on the specific number and meaning of parameters through unified window partitioning and flattening operations. The dynamic label mapping unit manages the correspondence between fault types and numerical labels by maintaining an external configuration file, and only needs to update the configuration file when adapting to new scenarios.

6. A universal intelligent state recognition method based on the universal intelligent state recognition system according to any one of claims 1 to 5, characterized in that, The entire processing flow of the general intelligent state recognition method is independent of the meaning and quantity of the physical parameters of the input data. It only adapts to different equipment or fault types by modifying the tag mapping configuration file, and includes the following steps: A standardized training sample set is generated by dividing, flattening and labeling historical equipment bus data through a multi-scale sliding window. The training sample set is cleaned and standardized, the LightGBM multi-class classification model is trained, and the model is evaluated and saved. For newly acquired equipment bus data, a sliding window with the same parameters as in the training phase is used for partitioning and standardization; Use a pre-trained LightGBM model to perform state recognition and prediction on new data; Generate status identification reports and visual analysis results.

7. The general intelligent state recognition method as described in claim 6, characterized in that, The steps to generate a standardized training sample set include: extracting numeric labels from the filenames of the original data using regular expressions and associating them with feature vectors segmented by a sliding window.

8. The general intelligent state recognition method as described in claim 6, characterized in that, The steps for training the LightGBM multi-class classification model include: encoding the class labels using LabelEncoder, standardizing the feature data using StandardScaler, and persistently saving the encoder and standardizer parameters for subsequent inference.

9. The general intelligent state recognition method as described in claim 6, characterized in that, The steps to generate visualization analysis results include: using the t-SNE algorithm to reduce the dimensionality of the high-dimensional prediction results to two-dimensional or three-dimensional space and plotting them to visualize the clustering of different state categories.

Images