Digital smart factory equipment operation management system based on deep learning

Abstract

The invention provides a digital smart factory equipment operation management system based on deep learning, and the system comprises a data collection module, a data preprocessing module, a feature engineering module, a model training and fusion module, a dynamic optimization module, a fault judgment and early warning module and a data storage module which are in communication connection in sequence. According to the method, by providing a multi-algorithm fusion deep learning model architecture and integrating the advantages of time sequence prediction and a nonlinear classification model, the problem that complex data association cannot be mined by a single algorithm is solved, a precise data processing scheme adaptive to industrial data characteristics is designed, and the data quality and the characteristic discrimination degree are improved; and by constructing a model dynamic optimization mechanism, the problem of poor model generalization ability is solved, early-stage accurate early warning of faults is realized, the early warning advance time is prolonged, the non-planned shutdown probability is reduced, the real-time performance and robustness of the system are improved, and the method is adaptive to high-noise industrial scenes.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) technology, and in particular to a digital smart factory equipment operation and management system based on deep learning. Background Technology

[0002] Driven by Industry 4.0, digital smart factories have become the core of manufacturing transformation. As the core carrier of production, the operational stability of equipment directly determines production efficiency, product quality, and operating costs. With the improvement of factory automation levels, equipment types are becoming increasingly complex (CNC machine tools, industrial robots, etc.). The industrial data generated during operation exhibits characteristics of multi-dimensionality, nonlinearity, high noise, strong temporality, and uneven distribution. Furthermore, the data formats are inconsistent and the transmission protocols are heterogeneous, posing significant challenges to data processing and fault prediction.

[0003] Existing factory equipment operation management systems often rely on single machine learning algorithms (such as LSTM and XGBoost), which can only capture certain data features and cannot comprehensively uncover complex correlations. Once trained, these models remain unchanged, lacking dynamic optimization mechanisms. Their generalization ability drops sharply when faced with changes in operating conditions, data shifts, or new fault types, resulting in significant overfitting and bias issues. Preprocessing mechanisms are simple (mean filling, simple filtering), making it difficult to handle the high noise and numerous anomalies in industrial data, leading to low data quality. Feature extraction is limited to basic statistics, failing to integrate deep temporal features and static equipment attributes, resulting in single-dimensional features and insufficient support. Alarms rely on sudden feature mutations at the time of fault occurrence, failing to capture early, weak features. Alarms are only triggered when a fault is imminent or after it has occurred, with warning times typically less than one hour. This leaves factories with insufficient time for preventative maintenance, and sudden faults lead to prolonged downtime. Summary of the Invention

[0004] The technical problem this invention aims to solve is overcoming the shortcomings of existing technologies. This invention proposes a deep learning-based digital smart factory equipment operation and management system. To address the issues raised by existing factory equipment operation and management systems, which often rely on single machine learning algorithms, only capturing certain data features and failing to comprehensively uncover complex correlations; models remain fixed after training, lacking dynamic optimization mechanisms, resulting in a sharp drop in generalization ability when faced with changes in operating conditions, data shifts, or new fault types, leading to significant overfitting and bias problems; simple preprocessing mechanisms struggle to handle the high noise and numerous anomalies of industrial data, resulting in low data quality; feature extraction is limited to basic statistics, failing to integrate deep temporal features and equipment static attributes, resulting in a single feature dimension and insufficient support; alarms rely on feature mutations at the time of fault occurrence, failing to capture early, weak features, and only providing warnings when a fault is imminent or occurs, with warning times typically less than one hour, leaving factories with insufficient preventative maintenance time and prolonged downtime due to sudden faults. The technical solution adopted by this invention is: A deep learning-based digital smart factory equipment operation and management system, wherein the method is executed by the equipment operation and management system and includes the following steps: Step 1: By pre-setting multiple types of sensors and data interfaces, collect the time-series operation data and static attribute data of the target device in real time, and perform format standardization processing on the collected data to obtain a standardized dataset; Step 2: Perform data cleaning, data standardization, and feature engineering on the standardized dataset in sequence to obtain the fault prediction feature set; data cleaning removes outliers, missing values, and duplicate data; data standardization maps the cleaned data to a preset interval; and feature engineering extracts time-series features, static features, and cross features. Step 3: Input the fault prediction feature set into a pre-set multi-algorithm fusion deep learning model to calculate the real-time fault probability value of the target device; the model adjusts the parameters in real time based on a dynamic optimization mechanism. Step 4: Compare the real-time fault probability value with the preset fault threshold. If it is greater than or equal to the threshold, generate a warning message containing the fault type, warning level, and expected fault occurrence time. If it is less than the threshold, output the normal operation result and store the relevant data in the database for dynamic model optimization.

[0005] Preferably, in step one, the time-series operating data includes temperature, vibration, pressure, current, voltage, rotational speed, and operating time data; the static attribute data includes equipment model, manufacturer, years of operation, historical maintenance records, rated parameters, and installation location data.

[0006] Preferably, in step two, the data cleaning includes: Outlier handling: An improved IQR algorithm combined with the DBSCAN algorithm is used. The IQR algorithm screens potential outliers, and the DBSCAN algorithm performs secondary verification to delete real outliers. Missing value handling: For missing values ​​<5%, use K-nearest neighbor filling; for missing values ​​between 5% and 30%, use LSTM neural network filling; for missing values ​​≥30%, delete the field. Duplicate data processing: Based on dual verification of timestamp and data value, duplicate data with the same timestamp and a data value difference of <0.01 are deleted.

[0007] Preferably, in step two, the data standardization adopts an improved Z-score algorithm, introducing a time-series weighting factor, the formula of which is: ,in, The result is the standardized result at time t. This is the original data. , Let be the mean and standard deviation of the sliding window at time t. .

[0008] Preferably, in step two, the feature engineering process includes: Temporal feature extraction: Extracting time-domain and frequency-domain features; the power spectral density formula is as follows: ; Static feature extraction: Categorical data uses one-hot encoding or target encoding, while numerical data uses Min-Max standardization; Cross-feature construction: Construct interactive features of operating years and vibration peak values ​​to capture the correlation of different types of features.

[0009] Preferably, in step three, the multi-algorithm fusion deep learning model includes: The base model layer includes LSTM, GRU, and BiLSTM time-series prediction models, as well as XGBoost, LightGBM, and CatBoost nonlinear classification models, each outputting its respective fault probability prediction. ; Feature fusion layer: Weighted fusion based on attention mechanism, attention weights , fusion features ; Meta-classifier layer: Employs a logistic regression model to output the final fault probability value. .

[0010] Preferably, in step three, the dynamic optimization mechanism includes: Performance monitoring: Real-time calculation of accuracy, recall, precision, F1 score, and MAE. The accuracy formula is: The MAE formula is ; Optimization trigger: If any performance indicator exceeds the threshold for ≥5 consecutive sampling periods, a new fault type is detected, or the KL divergence is ≥0.2, an update is triggered; Model update: Incremental learning is adopted. The parameters of the base model are updated using the SGD algorithm, the learning rate is adjusted by cosine annealing, and the parameters of the fusion layer and the meta-classifier layer are recalculated.

[0011] Preferably, in step four, the preset fault threshold is dynamically adjusted using the following formula: ; Warning levels are based on Classification: [0, 0.2) is Level 1, [0.2, 0.4) is Level 2, and ≥0.4 is Level 3.

[0012] Preferably, in step four, the expected failure time is predicted using polynomial fitting, with the following formula: Solve time .

[0013] A deep learning-based digital smart factory equipment operation and management system includes a data acquisition module, a data preprocessing module, a feature engineering module, a model training and fusion module, a dynamic optimization module, a fault diagnosis and early warning module, and a data storage module, which are sequentially connected in communication to execute the method described in any one of claims 1-9. The functions of each module are as follows: Data acquisition module: Collects and standardizes data through multiple types of sensors and industrial communication interfaces; Data preprocessing module: Performs data cleaning and standardization, and outputs clean data; Feature engineering module: Extracts multi-dimensional features and outputs a fault prediction feature set; Model training and fusion module: Constructs a multi-algorithm fusion model and outputs real-time fault probability values; Dynamic optimization module: Monitors model performance, triggers and executes incremental updates; Fault diagnosis and early warning module: compares with thresholds and generates early warning information or normal results; Data storage module: Stores various types of data and ensures efficient querying.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a multi-algorithm fusion deep learning model architecture, which integrates the advantages of time series prediction and nonlinear classification models, solves the problem that a single algorithm cannot uncover complex data correlations, designs a precise data processing solution adapted to the characteristics of industrial data, and improves data quality and feature discrimination.

[0015] This invention also addresses the problem of poor model generalization ability by constructing a dynamic model optimization mechanism, enabling accurate early warning of faults, extending the warning lead time, reducing the probability of unplanned downtime, improving system real-time performance and robustness, and adapting to high-noise industrial scenarios. Attached Figure Description

[0016] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is a block diagram of the equipment operation and management system of the present invention; Figure 2 This is a flowchart illustrating the steps of the equipment operation and management method of the present invention. Detailed Implementation

[0017] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0018] Specific embodiments of the present invention are described below with reference to the accompanying drawings.

[0019] Please see Figures 1-2 To address the shortcomings of existing systems, this embodiment proposes a deep learning-based method for the operation and management of equipment in a digital smart factory. This method is executed by the equipment operation and management system and includes the following steps: S1. Data Acquisition: By pre-setting multiple types of sensors and data interfaces, the system collects real-time operating data and static attribute data of the target device, performs format standardization processing on the collected data, and obtains a standardized dataset.

[0020] The time-series operational data includes temperature, vibration, pressure, current, voltage, speed, and operating time data; the static attribute data includes equipment model, manufacturer, years of operation, historical maintenance records, rated parameters, and installation location data; the format standardization process converts heterogeneous data into JSON format, unifying field names, data units, and timestamps in the format "YYYY-MM-DDHH:MM:SS.ms".

[0021] S2. Data Processing: Perform data cleaning, data standardization, and feature engineering on the standardized dataset in sequence to obtain the fault prediction feature set; data cleaning removes outliers, missing values, and duplicate data; data standardization maps the cleaned data to a preset interval; and feature engineering extracts time-series features, static features, and cross features.

[0022] Data cleaning includes: Outlier handling: An improved IQR algorithm combined with the DBSCAN algorithm is used. The IQR algorithm filters potential outliers (IQR=Q3-Q1, threshold range [Q1-1.5×IQR,Q3+1.5×IQR]), and the DBSCAN algorithm (setting the neighborhood radius ε and the minimum number of points MinPts) performs secondary verification to delete real outliers.

[0023] Missing value handling: For missing values ​​<5%, use K-nearest neighbor filling (weighted mean, weight is inversely proportional to distance); for missing values ​​between 5% and 30%, use LSTM neural network filling; for missing values ​​≥30%, delete the field.

[0024] Duplicate data processing: Based on dual verification of timestamp and data value, duplicate data with the same timestamp and a data value difference of <0.01 are deleted.

[0025] Data standardization employs an improved Z-score algorithm, introducing a time-series weighting factor, as shown in the formula: ,in, The result is the standardized result at time t. This is the original data. , Let be the mean and standard deviation of the sliding window at time t. (α takes values ​​of 0.1-0.5, (T is the time at the center of the window, and T is the window length).

[0026] Feature engineering processes include: Time-series feature extraction: Extracting time-domain features (mean, variance, peak value, etc.) and frequency-domain features (amplitude spectrum peak value, power spectral density, etc.). The power spectral density formula is... , ( The frequency domain data after FFT transformation. (Sampling frequency).

[0027] Static feature extraction: Categorical data uses one-hot encoding or target encoding, while numerical data uses Min-Max normalization. ).

[0028] Cross-feature construction: Construct interactive features such as "years of operation - peak vibration" to capture the correlation between different types of features.

[0029] S3. Data Judgment: Input the fault prediction feature set into a preset multi-algorithm fusion deep learning model to calculate the real-time fault probability value of the target device; the model adjusts parameters in real time based on a dynamic optimization mechanism. Multi-algorithm fusion deep learning models include: The base model layer includes LSTM, GRU, and BiLSTM time-series prediction models, as well as XGBoost, LightGBM, and CatBoost nonlinear classification models, each outputting its respective fault probability prediction. .

[0030] Feature fusion layer: Weighted fusion based on attention mechanism, attention weights , ( (Base model F1 score percentage), fusion features .

[0031] Meta-classifier layer: Employs a logistic regression model to output the final failure probability value. (w is the weight, b is the bias).

[0032] Dynamic optimization mechanisms include: Performance monitoring: Real-time calculation of accuracy, recall, precision, F1 score, and MAE. The accuracy formula is as follows: The MAE formula is .

[0033] Optimization trigger: If any performance indicator exceeds the threshold for ≥5 consecutive sampling periods (accuracy ≥90%, etc.), a new fault type is detected, or KL divergence ≥0.2, an update is triggered.

[0034] Model update: Incremental learning is used, with the base model parameters updated using the SGD algorithm and the learning rate adjusted by cosine annealing. The parameters of the fusion layer and the meta-classifier layer are recalculated.

[0035] S4. Output Results: Compare the real-time fault probability value with the preset fault threshold. If it is greater than or equal to the threshold, generate a warning message containing the fault type, warning level, and expected fault occurrence time. If it is less than the threshold, output the normal operation result and store the relevant data in the database for dynamic model optimization.

[0036] The preset fault threshold is dynamically adjusted using the following formula: (β ranges from 0.6 to 0.8, K is the number of historical fault samples, and L is the number of historical normal samples); the warning level is based on... Classification: [0, 0.2) is Level 1, [0.2, 0.4) is Level 2, and ≥0.4 is Level 3.

[0037] The expected failure time is predicted using a polynomial fitting method, as shown in the formula: Solve time .

[0038] A deep learning-based digital smart factory equipment operation and management system is proposed. This system adopts a closed-loop architecture of data acquisition, preprocessing, feature engineering, model fusion prediction, dynamic optimization, and fault early warning. It works collaboratively through seven modules. The specific technical solution is as follows: System Overall Architecture The design adopts a layered approach, with each layer having the following functions: The perception layer consists of various types of industrial sensors and equipment controllers, which collect time-series data and static attribute data. Transport layer: Employs industrial Ethernet, MQTT / OPCUA protocols, and 5G edge computing to achieve low-latency, high-reliability transmission; Data layer: Composed of a distributed database (HadoopHBase + MySQL) and a Redis cache, responsible for data storage and fast querying; Algorithm layer: The core layer, which includes data preprocessing, feature engineering, multi-algorithm fusion models, and dynamic optimization algorithms; Application layer: Provides tools such as equipment status monitoring and fault early warning notifications for operation and maintenance personnel to use.

[0039] The seven modules include: data acquisition module, data preprocessing module, feature engineering module, model training and fusion module, dynamic optimization module, fault diagnosis and early warning module, and data storage module.

[0040] 1. Data Acquisition Module Data collection targets: key production equipment such as CNC machine tools and industrial robots; Data collected: Time-series operational data (temperature, vibration, etc., sampling frequency 1-100Hz) and static attribute data (equipment model, years of operation, etc.); Data acquisition methods: direct sensor acquisition, controller interface reading, and gateway aggregation. Standardize the format: Convert to JSON format and unify the field names, units and timestamp formats.

[0041] 2. Data Preprocessing Module Outlier handling: Improved IQR+DBSCAN dual detection to avoid misjudgment by a single algorithm; Missing value handling: Adaptively select the filling method or delete invalid fields based on the missing value ratio; Deduplication: Double-checked verification removes redundant data; Data standardization: Time-weighted Z-score standardization, focusing on recent data volatility.

[0042] 3. Feature Engineering Module Temporal features: Extracting features in the time domain (peak value, kurtosis, etc.) and frequency domain (power spectral density, etc.); Static features: Encoding categorical data and Min-Max standardization for numerical data; Cross-features: Construct interactive features between time-series and static features to improve discriminative power.

[0043] 4. Model Training and Fusion Module The base model layer consists of 6 heterogeneous models (3 time-series prediction models + 3 non-linear classification models), each trained to output predicted values. Feature fusion layer: Weighted fusion using attention mechanism, higher-performing models receive higher weights; Meta-classifier layer: The logistic regression model is optimized twice to output the final failure probability value.

[0044] 5. Dynamic Optimization Module Performance monitoring: Real-time calculation of multi-dimensional performance metrics; Optimization triggering: Multiple conditions trigger updates (performance exceeding limits, new fault types, etc.); Model update: Incremental learning updates parameters, and the learning rate is dynamically adjusted.

[0045] 6. Fault Diagnosis and Early Warning Module Dynamic threshold: adaptively calculated based on historical data; Warning levels: divided into three levels, providing clear action guidelines; Failure time prediction: Polynomial fitting predicts failure occurrence time; Fault type identification: Feature association analysis determines the fault type.

[0046] 7. Data storage module Hybrid storage architecture: Hadoop HBase stores massive amounts of time-series data, MySQL stores static data, and Redis caches high-frequency data; Data backup: Regular backups + off-site disaster recovery ensure data security.

[0047] System hardware and software environment Data acquisition devices include: PT100 temperature sensor (accuracy ±0.1℃), ICP-093 vibration sensor (range 0-50g), and KG5100 IoT gateway; Edge computing node: Intel Core i7-12700H CPU, NVIDIA RTX 3060 GPU, 32GB DDR5 memory, 1TB SSD; Server-side: Intel Xeon Gold 5318Y CPU, NVIDIA A100 GPU, 128GB DDR4 memory, 10TB HDD + 2TB SSD, 3-node distributed database cluster; Clients: 15.6-inch industrial tablet PC, Android / iOS smartphones.

[0048] Software environment: Operating systems: Ubuntu 20.04 LTS (edge ​​node), CentOS 7.9 (server), Windows 10 IoT / Android 12 / iOS 16 (client). Development languages: Python 3.9, Java 11, Vue 3.0; Frameworks and tools: Pandas 1.5.3, TensorFlow 2.10.0, PyTorch 1.13.1, HadoopHBase 2.4.12, etc.; Communication protocols: MQTT 3.1.1, OPCUA 1.04, Modbus TCP.

[0049] Data acquisition module implementation Sensor deployment: Deploy sensors according to the parts of the equipment that are sensitive to equipment failure, such as deploying vibration and temperature sensors at the end of the spindle of a CNC machine tool; Data transmission: Sensor signals are converted into digital signals by a conditioner, aggregated by the gateway and converted into JSON format, and then transmitted to the edge nodes via the MQTT protocol.

[0050] Data preprocessing module implementation Outlier handling: IQR algorithm: Calculate quartiles Q1 and Q3, IQR = Q3 - Q1, and filter out potential outliers; DBSCAN algorithm: ε=0.5×σ, MinPts=5, verify potential outliers and delete real outliers; Missing value handling: Missing percentage calculation: ; K-nearest neighbor fill (K=5): Weighted average fill, where the weight is inversely proportional to the distance; LSTM padding: Input 10 data points before and after the missing value. The model structure is input layer (20-dimensional) → hidden layer (64 neurons) → output layer (1-dimensional).

[0051] For duplicate data processing, the absolute error of the same timestamp data value is calculated. If the error is less than 0.01, it is judged as duplicate data, and the first record is retained.

[0052] Data standardization: Sliding window T=60s, step size=1s, calculate the mean within the window. Standard deviation ; Time series weighting factor ; Standardized calculation: .

[0053] Feature Engineering Module Implementation Temporal feature extraction Time domain characteristics: mean Peak 10 features; Frequency domain features: The frequency domain data is obtained by FFT transformation, and six features such as amplitude spectrum and power spectral density are calculated.

[0054] Static feature extraction Category data: Equipment model target code, manufacturer's unique thermal code; Numerical data: Min-Max standardization.

[0055] Cross-feature construction Six cross features were constructed, such as (F1 = years of operation × peak vibration, F2 = rated power / current variance, etc.), and the final feature set contained 30 features.

[0056] Model training and fusion module implementation Basic model parameter settings Feature fusion layer implementation Calculate the F1 score of the basic model validation set, such as LSTM (0.90), BiLSTM (0.92), etc.; Performance rating Attention weight ; Fusion features .

[0057] Meta-classifier layer implementation Logistic Regression Model: Input Fusion Features Minimize the cross-entropy loss function and optimize the parameters w and b using gradient descent.

[0058] Implementation of dynamic optimization module Performance monitoring With a monitoring frequency of 1 minute per instance, calculate the accuracy, recall, and other metrics for the most recent 1000 samples.

[0059] Optimize triggering An update is triggered if the performance metric exceeds the threshold five times consecutively, a new fault type is detected, or the KL divergence is ≥0.2.

[0060] Model update Incremental data: Of the most recent 10,000 samples, 15% were faulty. Base model: SGD optimizer, cosine annealing learning rate ; Fusion layer and meta-classifier layer: recalculate weights and parameters, and the updated performance must meet the requirements of accuracy ≥95% and recall ≥92%.

[0061] Implementation of Fault Diagnosis and Early Warning Module Dynamic threshold setting Select data from the most recent 3 months, K=1000, L=10000. , ; Dynamic threshold .

[0062] Warning level classification Level 1 Warning: The system prompts that an inspection must be conducted within 24 hours. Level 2 warning: A reminder to perform preventative maintenance within 8 hours. Level 3 Warning: The system prompts you to stop and inspect it immediately.

[0063] Disaster occurrence time prediction A 3rd-order polynomial is used to fit the failure probability values ​​over the past 10 minutes, and the solution is obtained. time This refers to the estimated downtime.

[0064] Fault type identification The current feature set is matched with the feature templates of 12 fault types. If the matching degree is ≥80%, the fault type is determined; otherwise, the model is updated.

[0065] Data storage module implementation HadoopHBase: Table name "device_running_data", row key "device_id+timestamp", partitioned by device type and time; MySQL: Stores model parameters, prediction results, etc., with table structures including model_parameters, prediction_results, etc. Redis: Caches real-time device status, recent prediction results, etc., and sets expiration policies. System deployment and operation process System Deployment Edge nodes are deployed in the workshop and connected to IoT gateways; The server cluster is deployed in the data center and is responsible for model training and data storage. The client installs the application and accesses the server via LAN or 5G.

[0066] Operation process Data acquisition and transmission → Data preprocessing → Feature engineering → Model prediction → Dynamic optimization → Fault diagnosis and early warning → Data storage.

[0067] Compared to existing factory equipment operation and management systems, this invention addresses the core pain points of existing technologies, such as "limitations of single algorithms, insufficient adaptation to industrial data processing, weak early fault identification, poor model anti-interference, and low operational efficiency." Through multi-dimensional technological innovation, it forms synergistic advantages, achieving not only a breakthrough improvement at the technological level but also creating significant economic and management value in practical applications.

[0068] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A deep learning-based method for the operation and management of equipment in a digital smart factory, characterized in that: The method is executed by the equipment operation management system and includes the following steps: Step 1: By pre-setting multiple types of sensors and data interfaces, collect the time-series operation data and static attribute data of the target device in real time, and perform format standardization processing on the collected data to obtain a standardized dataset; Step 2: Perform data cleaning, data standardization, and feature engineering on the standardized dataset in sequence to obtain the fault prediction feature set; data cleaning removes outliers, missing values, and duplicate data; data standardization maps the cleaned data to a preset interval; and feature engineering extracts time-series features, static features, and cross features. Step 3: Input the fault prediction feature set into a pre-set multi-algorithm fusion deep learning model to calculate the real-time fault probability value of the target device; the model adjusts the parameters in real time based on a dynamic optimization mechanism. Step 4: Compare the real-time fault probability value with the preset fault threshold. If it is greater than or equal to the threshold, generate a warning message containing the fault type, warning level, and expected fault occurrence time. If it is less than the threshold, output the normal operation result and store the relevant data in the database for dynamic model optimization.

2. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step one, the time-series operating data includes temperature, vibration, pressure, current, voltage, speed, and operating time data; the static attribute data includes equipment model, manufacturer, years of operation, historical maintenance records, rated parameters, and installation location data.

3. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step two, the data cleaning includes: Outlier handling: An improved IQR algorithm combined with the DBSCAN algorithm is used. The IQR algorithm screens potential outliers, and the DBSCAN algorithm performs secondary verification to delete real outliers. Missing value handling: For missing values ​​<5%, use K-nearest neighbor filling; for missing values ​​between 5% and 30%, use LSTM neural network filling; for missing values ​​≥30%, delete the field. Duplicate data processing: Based on dual verification of timestamp and data value, duplicate data with the same timestamp and a data value difference of <0.01 are deleted.

4. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step two, the data standardization adopts an improved Z-score algorithm, introducing a time-series weighting factor, as shown in the formula: ,in, The result is the standardized result at time t. The original data, , Let be the mean and standard deviation of the sliding window at time t. .

5. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step two, the feature engineering process includes: Temporal feature extraction: Extracting time-domain and frequency-domain features; the power spectral density formula is as follows: ; Static feature extraction: Categorical data uses one-hot encoding or target encoding, while numerical data uses Min-Max standardization; Cross-feature construction: Construct interactive features of operating years and vibration peak values ​​to capture the correlation of different types of features.

6. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step three, the multi-algorithm fusion deep learning model includes: The base model layer includes LSTM, GRU, and BiLSTM time-series prediction models, as well as XGBoost, LightGBM, and CatBoost nonlinear classification models, each outputting its respective fault probability prediction. ; Feature fusion layer: Weighted fusion based on attention mechanism, attention weights , fusion features ; Meta-classifier layer: Employs a logistic regression model to output the final fault probability value. 。 7. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step three, the dynamic optimization mechanism includes: Performance monitoring: Real-time calculation of accuracy, recall, precision, F1 score, and MAE. The accuracy formula is: The MAE formula is ; Optimization trigger: If any performance indicator exceeds the threshold for ≥5 consecutive sampling periods, a new fault type is detected, or the KL divergence is ≥0.2, an update is triggered; Model update: Incremental learning is adopted. The parameters of the base model are updated using the SGD algorithm, the learning rate is adjusted by cosine annealing, and the parameters of the fusion layer and the meta-classifier layer are recalculated.

8. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step four, the preset fault threshold is dynamically adjusted using the following formula: ； Warning levels are based on Classification: [0, 0.2) is Level 1, [0.2, 0.4) is Level 2, and ≥0.4 is Level 3.

9. The method for operation and management of digital smart factory equipment based on deep learning according to claim 1, characterized in that, In step four, the expected failure time is predicted using polynomial fitting, with the following formula: Solve time .

10. A deep learning-based digital smart factory equipment operation and management system, characterized in that, The system includes a data acquisition module, a data preprocessing module, a feature engineering module, a model training and fusion module, a dynamic optimization module, a fault diagnosis and early warning module, and a data storage module, which are connected in sequence to execute the method described in any one of claims 1-9. The functions of each module are as follows: Data acquisition module: Collects and standardizes data through multiple types of sensors and industrial communication interfaces; Data preprocessing module: Performs data cleaning and standardization, and outputs clean data; Feature engineering module: Extracts multi-dimensional features and outputs a fault prediction feature set; Model training and fusion module: Constructs a multi-algorithm fusion model and outputs real-time fault probability values; Dynamic optimization module: Monitors model performance, triggers and executes incremental updates; Fault diagnosis and early warning module: compares with thresholds and generates early warning information or normal results; Data storage module: Stores various types of data and ensures efficient querying.

Images