Coal mill fault detection system and method based on intelligent deep learning technology

Abstract

The invention discloses a coal mill fault detection system based on an intelligent deep learning technology and a method thereof. The system and the method are suitable for state monitoring and fault diagnosis of coal mill equipment in the iron and steel industry. According to the system, an improved Diffusion model is adopted to generate diversified fault data, and a fault classification model is constructed based on a ResNet structure, so that continuous real-time monitoring and intelligent diagnosis of coal mill equipment are realized. The system uses vibration signals and PLC data collected by a sensor to generate diversified new data for training a downstream classification model. The system can identify various fault types such as bearing damage, coupler damage and foundation looseness, and provides an alarm prompt through a real-time early warning mechanism, so that the operation reliability and the maintenance efficiency of equipment are remarkably improved. The intelligent fault detection system has the advantages of being stable in data generation, high in fault recognition accuracy and high in application scene adaptability, the problem of scarcity of fault data is solved, and an efficient solution is provided for a predictive maintenance strategy in the field of steel manufacturing.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial equipment fault detection, and specifically relates to a coal mill fault detection system and method based on intelligent deep learning technology. Background Art

[0002] With the rapid development of Industry 4.0 and Internet of Things (IoT) technologies, remote monitoring and fault diagnosis of equipment have become important requirements of modern industry. In the steel industry, coal mill is one of the key equipment in the production process, and its performance directly affects the operating efficiency and product quality of the production line. Coal mills usually need to operate for a long time under high load and complex working conditions, so they are prone to failures such as unbalanced shaft system, loose foundation, bearing damage and other problems. If these faults are not detected and handled in time, they may cause equipment downtime, production interruption, and even cause significant economic losses.

[0003] Traditional coal mill fault detection methods mostly rely on manual inspections and experience-based fault diagnosis methods. These methods are not only inefficient, but also often difficult to capture fault signals in a timely and accurate manner. In recent years, sensor-based condition monitoring technology has been widely used. By installing sensors on coal mills to collect vibration signals, temperature signals and other data, the operating status of the equipment can be monitored in real time. However, how to use these massive data and perform efficient and accurate fault diagnosis still faces many technical challenges.

[0004] In the existing technology, intelligent fault diagnosis methods based on deep learning have gradually gained attention. Deep learning models can learn the normal and abnormal characteristics of equipment operation through large-scale training data, thereby realizing automatic and intelligent fault detection. However, when deep learning technology is applied to coal mill fault detection in industrial sites, the following problems still exist:

[0005] 1. Data scarcity: Coal mill failure events are relatively rare, resulting in insufficient actual failure data to support effective training of deep learning models. This data scarcity often limits the accuracy and robustness of fault detection models.

[0006] 2. Fusion of multi-source heterogeneous data: In the process of coal mill fault detection, vibration signals and data collected by PLC (programmable logic controller) are usually involved. Due to the different sampling frequencies and data characteristics of different sensors, the data are heterogeneous and asynchronous, so the data needs to be effectively processed to be better utilized.

[0007] 3. System real-time and energy efficiency requirements: In industrial field environments, the fault detection system needs to reduce system power consumption as much as possible while ensuring real-time and accuracy to extend equipment life and reduce maintenance costs.

[0008] In view of the above problems, the current market still lacks an efficient, low-power intelligent fault detection system that can adapt to the complex operating environment of coal mills. Therefore, studying a highly robust and high-precision coal mill fault detection system based on deep learning can achieve continuous real-time monitoring and intelligent diagnosis of equipment, which has important application value for improving production efficiency and reducing operating costs. Summary of the invention

[0009] To solve the above problems, the present invention discloses a coal mill fault detection system and method based on intelligent deep learning technology. The system builds an efficient and low-power coal mill fault detection system by generating synthetic fault data, optimizing data preprocessing processes and fusing multi-source data, which can realize real-time monitoring and accurate fault diagnosis of equipment. By adopting an improved diffusion model and ResNet classifier, the present invention realizes the fault type identification and early warning of key components of the coal mill, significantly improving the reliability and operation efficiency of the equipment.

[0010] To achieve the above object, the technical solution of the present invention is as follows:

[0011] The coal mill fault detection system based on intelligent deep learning technology includes a data acquisition module, a data preprocessing and feature extraction module, a core algorithm module and a front-end interface module; the data acquisition module obtains vibration and PLC data from the coal mill sensor, and after being processed by the data preprocessing and feature extraction module, it is input into the core algorithm module for fault detection and classification, and finally the detection results and warning information are displayed in the front-end interface module.

[0012] In the system, the traditional diffusion model is improved to make it more efficient and interpretable in generating high-quality time series data. Specifically, the reverse denoising process of the model is modified to allow the generator network to perform interpretable decomposition while learning time series data, and can decompose time series data into three parts: trend, periodic changes, and non-periodic changes. This improvement not only enhances the model's ability to understand the data generation process, but also improves the interpretability of the generated data. Through this process, the system can effectively generate time series data with multi-dimensional features, providing more diverse data inputs for fault detection models, thereby enhancing their robustness and generalization capabilities.

[0013] The process of adding noise in the forward direction converts the original data x0 into the Pure Gaussian noise

[0014]

[0015] (1) The basis of the forward noise addition process is a random process that gradually transforms the original data into Gaussian noise. At each step, the data is added with noise based on a Gaussian distribution. The system adds noise to three components - trend, cyclical changes, and non-cyclical changes - according to the characteristics of the data. This decomposition allows the generation process to handle different dynamic characteristics in the time series in a more detailed manner.

[0016] (2) For trends, the addition of noise will simulate the possible offset or change of data during long-term operation; the addition of noise to the periodic change part is used to simulate the periodic fluctuations in equipment operation, such as the vibration frequency of the machine under different load conditions; and the addition of noise to the non-periodic change part is mainly used to simulate those unpredictable random events, which may come from environmental changes or instantaneous abnormalities of the equipment.

[0017] (3) Parameter adjustment is also an important step in adding forward noise. In practical applications, the noise intensity and frequency can be adjusted to simulate data under different working conditions. For example, increasing the noise intensity can simulate the state of high-load operation, while reducing the noise intensity can simulate the state of the equipment in relatively stable operation. This flexible adjustment capability enables the generated data to cover a wider range of application scenarios and enhances the adaptability of the model in different environments.

[0018] Reverse data reconstruction, based on the following formula, gradually learns to reversely remove the noise generated by the forward process and constructs a new time series approximating the original data x0

[0019]

[0020] (1) In the forward process, noise is gradually added to the data, and finally it is transformed into complete Gaussian noise. The goal of reverse data reconstruction is to gradually remove the added noise from the noise, and finally reconstruct an approximation of the original data. The basis of this process is an inverse process model, which gradually recovers the data distribution close to the original from the noise by learning denoising.

[0021] (2) During the reverse reconstruction process, the system uses a generative network to learn how to extract useful features from noisy samples and gradually remove the noise through multiple iterations. This process is not just a simple noise removal, but also involves how to recombine the decomposed trends, periodic changes, and non-periodic changes to form a complete time series.

[0022] (3) In order to ensure the consistency of the distribution of the reconstructed data with the original data, the system uses the Fourier loss function during the training process. The introduction of the Fourier loss function can more accurately capture the differences in the frequency domain of the reconstructed data, which is very important for processing data with obvious frequency characteristics such as vibration signals. Through backpropagation and parameter optimization, the model continuously improves itself, so that it can efficiently reconstruct data similar to the original distribution.

[0023] The coal mill fault detection method based on intelligent deep learning technology includes the following steps:

[0024] S1, data collection and preprocessing: The operation data of the coal mill is collected through sensors, including vibration signals and PLC data. First, the collected vibration signals are processed by wavelet filtering to remove noise interference; then the frequency domain features are extracted by fast Fourier transform (FFT), and the data with different sampling frequencies are aligned to generate multi-source feature data suitable for fault detection model training.

[0025] S2, Fault data generation: To solve the problem of insufficient actual fault data, an improved diffusion model is used to generate synthetic fault data. The diffusion model generates high-quality synthetic data similar to the actual data distribution through the process of forward addition of Gaussian noise and reverse denoising and reconstruction, so as to enhance the generalization ability and robustness of the fault detection model.

[0026] S3, Fault classification model training: Based on the generated fault data and preprocessed feature data, a deep residual network (ResNet) is used to build a fault classification model. Through feature extraction and pattern recognition of the data, the classification model can accurately identify various types of coal mill faults, including shaft imbalance, bearing damage, coupling damage, etc.

[0027] S4, real-time monitoring and early warning: During the operation of the equipment, the system processes the collected data in real time and judges the operating status of the equipment through the classification model. When a fault is detected, the system issues an alarm message through the early warning module and pushes it to the operation and maintenance personnel to ensure that the fault can be responded to and handled in a timely manner.

[0028] As an improvement of the present invention, the data preprocessing in step S1 specifically includes:

[0029] S11: Vibration signal denoising. Wavelet filtering technology is used to denoise the vibration signal and remove the noise components in the signal, thereby extracting clearer equipment operation status information and providing an accurate signal basis for subsequent fault diagnosis.

[0030] S12: Frequency domain feature extraction. The denoised vibration signal is converted into the frequency domain using the fast Fourier transform (FFT) to extract key features such as frequency components and amplitude to more comprehensively reflect the operating characteristics of the equipment.

[0031] S13: Data alignment. Since the sampling frequencies of different PLC sensors are inconsistent, direct use may lead to data asynchrony. To this end, the present invention uses an interpolation method to align multi-source data to ensure that all types of data are synchronized in the same time dimension, providing consistent data support for model training.

[0032] As another improvement of the present invention, the fault data generation process in step S2 includes the following steps:

[0033] S21: Generate fault data. To solve the problem of scarcity of fault data in industrial sites, the present invention uses a diffusion model to generate diverse fault data samples to enrich the training data set. These synthetic data simulate different operating scenarios, allowing the model to adapt to a variety of operating conditions, thereby improving the robustness of fault detection.

[0034] S22: Data authenticity verification. In order to ensure that the generated fault data has a high degree of authenticity, the present invention adopts a multi-step verification method to ensure data availability and authenticity:

[0035] (1) Sample consistency verification: Randomly extract synthetic data samples and actual data samples to input into the fault detection model and compare the prediction results of the two. If the prediction distribution is significantly different, fine-tune the data generation model to ensure that the synthetic data accurately reflects the characteristics of various faults.

[0036] (2) Model accuracy verification: Combine synthetic data and real data into a mixed training set and evaluate the prediction accuracy of the model on the actual test set. If the accuracy improves or stabilizes, it indicates that the synthetic data has good authenticity; if the accuracy decreases, the data generation process needs to be optimized.

[0037] (3) Model robustness verification: Train the model multiple times and observe its performance stability under different data combinations. If the training results fluctuate little and the classification accuracy is high, it means that the synthetic data enhances the robustness of the model; otherwise, the data generation strategy needs to be adjusted.

[0038] (4) Domain expert review: The feature distribution of synthetic data and the actual data distribution are submitted to experts for review. Experts judge their rationality based on their experience and provide feedback to further optimize the diffusion model parameters to ensure that the synthetic data is authentic and reliable.

[0039] As another improvement of the present invention, the structure of the fault classification model in step S3 is specifically as follows:

[0040] S31: ResNet structure. The fault classification model of the present invention is designed based on the ResNet structure, uses multi-layer residual blocks for feature extraction, and captures complex patterns and detailed features in vibration signals through a deep learning model. Finally, accurate fault classification is achieved by outputting the probability distribution of various faults.

[0041] S32: Model parameter optimization. To further improve the accuracy and robustness of fault detection, the present invention uses cross-validation and parameter tuning techniques in the model training process to ensure that the prediction accuracy of the classification model is not less than 85%, which can meet the high-precision detection requirements of industrial sites.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) Enhanced data generation capability: The diffusion model is used to generate synthetic fault data, which solves the problem of insufficient actual fault data and improves the model's training effect and recognition accuracy.

[0044] (2) Optimize the data processing process: Vibration signals and PLC data are effectively preprocessed through wavelet filtering, FFT feature extraction, and data alignment technology to ensure the quality and consistency of model input data.

[0045] (3) Improve the accuracy and real-time performance of fault diagnosis: By utilizing the deep structure of the ResNet classification model, the system can accurately identify various types of coal mill faults based on multi-frequency features, meeting the real-time and accuracy requirements of industrial sites.

[0046] (4) Realize intelligent early warning function: When a fault is detected, the system issues early warning information in real time, providing support for predictive maintenance of equipment, significantly reducing downtime and lowering equipment operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a system architecture diagram of the coal mill fault diagnosis system and its data generation and classification method of the present invention.

[0048] Figure 2 The present invention is a flowchart of the coal mill fault diagnosis system and its data generation and classification method.

[0049] Figure 3 Flowchart for working with the improved diffusion model. DETAILED DESCRIPTION

[0050] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0051] like Figure 1As shown, the coal mill fault diagnosis system described in the present invention includes the following main modules: data acquisition module, data preprocessing and feature extraction module, core algorithm module and front-end interface module. The data acquisition module obtains vibration and PLC data from the coal mill sensor, and after being processed by the preprocessing module, it is input into the core algorithm module for fault detection and classification, and finally the detection results and warning information are displayed in the front-end interface module:

[0052] System structure and function:

[0053] 1. Data acquisition module

[0054] Vibration sensors and temperature sensors are installed at key parts of the coal mill (such as drive motors, bearings, etc.), and parameters such as motor current, bearing temperature and regulating valve opening are obtained through the PLC system. The vibration signal and PLC data are transmitted to the data processing center in real time through wireless transmission or wired interface.

[0055] 2. Data preprocessing and feature extraction module

[0056] This module preprocesses the collected raw data, mainly including vibration signal denoising, frequency domain feature extraction and multi-source data fusion processing.

[0057] Vibration signal denoising: The vibration signal is denoised through wavelet filtering method to remove high-frequency noise interference, thereby obtaining a clearer equipment operation signal.

[0058] Frequency domain feature extraction: Use Fast Fourier Transform (FFT) to convert the time domain signal to the frequency domain and extract key frequency features, including the frequency domain root mean square value (RMS) and the energy value of the main frequency band.

[0059] Multi-source data fusion: interpolation and alignment of PLC data with different sampling frequencies to ensure the consistency of all data in the time dimension and provide complete data feature input for subsequent fault detection models.

[0060] In the system, the traditional diffusion model is improved to make it more efficient and interpretable in generating high-quality time series data. Specifically, the reverse denoising process of the model is modified to allow the generator network to perform interpretable decomposition while learning time series data, and can decompose time series data into three parts: trend, periodic change, and non-periodic change. Figure 3 As shown in Figure 2, this improvement not only enhances the model’s ability to understand the data generation process, but also improves the interpretability of the generated data. Through this process, the system can effectively generate time series data with multi-dimensional features, providing more diverse data inputs for the fault detection model, thereby enhancing its robustness and generalization capabilities.

[0061] The process of adding noise in the forward direction converts the original data x0 into the Pure Gaussian noise

[0062]

[0063] (1) The basis of the forward noise addition process is a random process that gradually transforms the original data into Gaussian noise. At each step, the data is added with noise based on a Gaussian distribution. The system adds noise to three components - trend, cyclical changes, and non-cyclical changes - according to the characteristics of the data. This decomposition allows the generation process to handle different dynamic characteristics in the time series in a more detailed manner.

[0064] (2) For trends, the addition of noise will simulate the possible offset or change of data during long-term operation; the addition of noise to the periodic change part is used to simulate the periodic fluctuations in equipment operation, such as the vibration frequency of the machine under different load conditions; and the addition of noise to the non-periodic change part is mainly used to simulate those unpredictable random events, which may come from environmental changes or instantaneous abnormalities of the equipment.

[0065] (3) Parameter adjustment is also an important step in adding forward noise. In practical applications, the noise intensity and frequency can be adjusted to simulate data under different working conditions. For example, increasing the noise intensity can simulate the state of high-load operation, while reducing the noise intensity can simulate the state of the equipment in relatively stable operation. This flexible adjustment capability enables the generated data to cover a wider range of application scenarios and enhances the adaptability of the model in different environments.

[0066] Reverse data reconstruction, based on the following formula, gradually learns to reversely remove the noise generated by the forward process and constructs a new time series approximating the original data x0

[0067]

[0068] (1) In the forward process, noise is gradually added to the data, and finally it is transformed into complete Gaussian noise. The goal of reverse data reconstruction is to gradually remove the added noise from the noise, and finally reconstruct an approximation of the original data. The basis of this process is an inverse process model, which gradually recovers the data distribution close to the original from the noise by learning denoising.

[0069] (2) During the reverse reconstruction process, the system uses a generative network to learn how to extract useful features from noisy samples and gradually remove the noise through multiple iterations. This process is not just a simple noise removal, but also involves how to recombine the decomposed trends, periodic changes, and non-periodic changes to form a complete time series.

[0070] (3) In order to ensure the consistency of the distribution of the reconstructed data with the original data, the system uses the Fourier loss function during the training process. The introduction of the Fourier loss function can more accurately capture the differences in the frequency domain of the reconstructed data, which is very important for processing data with obvious frequency characteristics such as vibration signals. Through backpropagation and parameter optimization, the model continuously improves itself, so that it can efficiently reconstruct data similar to the original distribution.

[0071] 3. Core algorithm module

[0072] The core algorithm modules include data generation model and fault detection classification model.

[0073] Data generation model: Generate synthetic fault data based on the diffusion model to address the problem of scarce fault data. The model generates diversified fault data by adding Gaussian noise and reconstructing it, making the training data set richer and more diverse, thereby improving the robustness of the fault detection model.

[0074] Fault detection classification model: This paper adopts a deep neural network based on ResNet as the classification model, extracts features through residual blocks, and captures complex features in vibration signals. The model finally outputs the probability distribution of various faults and accurately identifies various fault types of coal mills.

[0075] 4. Front-end interface module

[0076] The front-end interface module is used to display the equipment operation status and fault detection results in real time, including equipment operation data monitoring, fault warning, historical fault record query and fault data management. Operation and maintenance personnel can view real-time equipment status and alarm information through the interface, which facilitates timely maintenance decisions.

[0077] The workflow of the data generation and classification method of the coal mill fault diagnosis system of the present invention is as follows: Figure 2 As shown in the figure, assuming that the fault detection system is deployed on a coal mill in a steel manufacturing plant, the specific workflow of the system is as follows:

[0078] Step S1: Field data collection

[0079] Vibration and temperature data collection, vibration sensors and temperature sensors are installed at key positions of the coal mill, such as motor bearings, casing, etc. The sensors monitor the vibration and temperature status of the equipment in real time and collect data every 5 seconds.

[0080] PLC controls data acquisition, and the PLC control system is connected to the main components of the coal mill to monitor key parameters such as motor current, power, valve opening, etc. Data is collected at intervals of 10 seconds and transmitted to the data preprocessing module through the industrial bus.

[0081] Data transmission, all collected data are transmitted to the central data processing center via a wired network for further analysis.

[0082] Step S2: Data preprocessing and feature extraction

[0083] Vibration signal denoising: The vibration signal received by the system usually contains noise components. Wavelet filtering technology is used to denoise the data and extract the key signal components that reflect the status of the equipment. The denoised signal enters the next step of feature extraction.

[0084] Frequency domain feature extraction: Fast Fourier Transform (FFT) is applied to the denoised vibration signal to convert the signal from the time domain to the frequency domain and extract the frequency features. The system focuses on extracting features such as frequency components, frequency domain root mean square value (RMS) and signal energy.

[0085] Multi-source data fusion and alignment: For data with different sampling frequencies (for example, vibration signals are sampled every 5 seconds, and PLC data are sampled every 10 seconds), data alignment is performed through interpolation algorithms to keep data from different sources consistent on the time axis. The fused data set contains multi-dimensional features such as vibration signals, temperature data, current, and power, which are used for model training and real-time detection.

[0086] Step S3: Fault data generation

[0087] Synthetic fault data generation,In order to solve the problem of insufficient fault data, the system uses a diffusion model to generate synthetic fault data.,The generated data samples include different fault scenarios, such as bearing damage,,coupling looseness, foundation looseness, etc.

[0088] Data quality verification, the generated synthetic data undergoes multiple steps of verification to ensure that it is similar in characteristics to the actual fault data in order to enhance the robustness of the model; specifically as follows:

[0089] (1) Sample consistency verification: Randomly extract synthetic data samples and actual data samples to input into the fault detection model and compare the prediction results of the two. If the prediction distribution is significantly different, fine-tune the data generation model to ensure that the synthetic data accurately reflects the characteristics of various faults.

[0090] (2) Model accuracy verification: Combine synthetic data and real data into a mixed training set and evaluate the prediction accuracy of the model on the actual test set. If the accuracy improves or stabilizes, it indicates that the synthetic data has good authenticity; if the accuracy decreases, the data generation process needs to be optimized.

[0091] (3) Model robustness verification: Train the model multiple times and observe its performance stability under different data combinations. If the training results fluctuate little and the classification accuracy is high, it means that the synthetic data enhances the robustness of the model; otherwise, the data generation strategy needs to be adjusted.

[0092] (4) Domain expert review: The feature distribution of synthetic data and the actual data distribution are submitted to experts for review. Experts judge their rationality based on their experience and provide feedback to further optimize the diffusion model parameters to ensure that the synthetic data is authentic and reliable.

[0093] Step S4: Fault detection and classification

[0094] Data input, the pre-processed vibration signal and PLC data are input into the ResNet-based deep learning model. The ResNet model extracts complex features in the data through multiple layers of residual blocks.

[0095] Fault identification,The ResNet model analyzes real-time data by learning the normal and fault modes of equipment operation, and outputs the equipment's fault type (such as bearing failure, shaft imbalance, etc.) and its corresponding confidence score.

[0096] Processing of classification results: If the confidence score exceeds the set fault threshold, the model determines that the device may be faulty and triggers the early warning mechanism.

[0097] Step S5: Fault warning and information display

[0098] Fault alarm: Once a potential fault is detected, the system will immediately generate an alarm message. The alarm message includes the device number, fault type, detection time and fault description, and is pushed to the operation and maintenance personnel via SMS or email.

[0099] Front-end interface display: On the front-end interface, the system displays the equipment's operating status, alarm information, fault type and fault location in real time. Operation and maintenance personnel can view detailed historical fault records through the interface to analyze fault frequency and fault causes.

[0100] Step S6: Historical data recording and system optimization

[0101] Historical data records: The system automatically stores the results of each fault detection and sensor data in the database, generating detailed historical records. Operation and maintenance personnel can query historical data at any time to make equipment health management and maintenance decisions.

[0102] Model optimization: The system regularly uses historical fault data and newly added fault samples to retrain the fault detection model to continuously improve the accuracy and adaptability of the model. By using the latest data, the system can better cope with changes in equipment failures under different loads and environments.

[0103] Workflow Summary

[0104] In actual application, the overall process of the system is: real-time data collection → data preprocessing and feature extraction → fault data generation → data input and fault detection → fault warning and display → historical records and system optimization. Through the above process, the system can realize continuous and real-time monitoring of coal mills and immediately alarm when a fault occurs, thereby helping factories reduce equipment downtime, reduce maintenance costs, and improve the operating efficiency of production lines.

[0105] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A coal mill fault detection system based on intelligent deep learning technology includes a data acquisition module, a data preprocessing and feature extraction module, a core algorithm module and a front-end interface module, and is characterized by: The data acquisition module is used to collect the operation data of the coal mill in real time, including vibration signals and PLC data; The algorithm module includes a data preprocessing unit, a data generation model and a classifier model, wherein: The data preprocessing unit denoises the vibration signal through wavelet filtering, extracts frequency domain features using fast Fourier transform, and aligns PLC data with different sampling frequencies; The data generation model generates synthetic fault data based on the diffusion model to solve the problem of insufficient actual fault data; The classifier model uses a ResNet-based structure to identify fault types on preprocessed data; The front-end interface module is used to display the operating status, alarm information, fault data management and historical fault data viewing of the equipment.

2. The coal mill fault detection system based on intelligent deep learning technology according to claim 1 is characterized in that: The data generation model generates fault data by a process of forward addition of Gaussian noise and reverse denoising to simulate various equipment working conditions, thereby improving the robustness of the classifier model and the accuracy of fault identification.

3. The coal mill fault detection system based on intelligent deep learning technology according to claim 1 is characterized in that: The classifier model adopts ResNet with a multi-layer residual block structure, and classifies the drive motor shaft imbalance, loose installation, and bearing damage faults through feature extraction and pattern recognition.

4. A coal mill fault detection method based on intelligent deep learning technology, characterized in that: The following steps are involved: Data collection steps: collect the vibration signal and PLC data of the coal mill in real time through the data collection module; Data preprocessing steps: Use wavelet filtering to denoise the vibration signal, and use fast Fourier transform to extract frequency domain features to complete the alignment of PLC data; Fault data generation step: Generate synthetic fault data using diffusion model; Fault classification step: Identify the fault type based on the ResNet classifier model; Fault warning step: When a fault is detected, a warning message is generated through the fault warning module and pushed to the operation and maintenance personnel; Logging step: During system operation, the log module records relevant information about model iterations, fault alarms, and fault data updates.

5. A coal mill fault detection method based on intelligent deep learning technology according to claim 4, characterized in that: In the fault data generation step, the diffusion model generates diversified fault data through the process of forward noise addition and reverse denoising and reconstruction, thereby improving the training effect of the classification model and the accuracy of fault identification; The process of adding noise in the forward direction converts the original data x0 into the Pure Gaussian noise (1) The basis of the forward noise addition process is a random process that gradually transforms the original data into Gaussian noise. At each step, noise based on Gaussian distribution is added to the data. The system adds noise to three components - trend, cyclical change and non-cyclical change - according to the characteristics of the data. This decomposition method enables the generation process to handle different dynamic characteristics in the time series in a more detailed manner. (2) For trends, the addition of noise will simulate the possible offset or change of data during long-term operation; the addition of noise to the periodic change part is used to simulate the periodic fluctuations in equipment operation, while the addition of noise to the non-periodic change part is mainly used to simulate those unpredictable random events, which may come from environmental changes or transient abnormalities of equipment; (3) Parameter adjustment is also an important step in adding forward noise; by adjusting the noise intensity and frequency, data under different working conditions can be simulated; Reverse data reconstruction, based on the following formula, gradually learns to reversely remove the noise generated by the forward process and constructs a new time series approximating the original data x0 (1) In the forward process, noise is gradually added to the data, and finally it is transformed into complete Gaussian noise; the goal of reverse data reconstruction is to gradually remove the added noise from the noise, and finally reconstruct an approximation of the original data; the basis of this process is an inverse process model, which gradually recovers the data distribution close to the original from the noise by learning denoising; (2) During the reverse reconstruction process, the system uses a generative network to learn how to extract useful features from noisy samples and gradually remove the noise through multiple iterations; This process is not just a simple noise removal, but also involves how to recombine the decomposed trends, periodic changes and non-periodic changes to form a complete time series; (3) In order to ensure the consistency of the distribution of the reconstructed data with the original data, the system uses the Fourier loss function during the training process. The introduction of the Fourier loss function can more accurately capture the differences in the frequency domain of the reconstructed data, which is very important for processing data with obvious frequency characteristics such as vibration signals. Through back propagation and parameter optimization, the model continuously improves itself, so that it can efficiently reconstruct data similar to the original distribution.

6. A coal mill fault detection method based on intelligent deep learning technology according to claim 4, characterized in that: The fault data generation step includes the following steps: S21: Generate fault data: Generate diverse fault data samples using the diffusion model to enrich the training data set; these synthetic data simulate different operating scenarios, allowing the model to adapt to a variety of operating conditions, thereby improving the robustness of fault detection; S22: Data authenticity verification: In order to ensure that the generated fault data has a high degree of authenticity, the following verification methods are used to ensure data availability and authenticity: (1) Sample consistency verification: Randomly extract synthetic data samples and actual data samples and input them into the fault detection model to compare the prediction results of the two. If the prediction distribution is significantly different, fine-tune the data generation model to ensure that the synthetic data accurately reflects the characteristics of various faults. (2) Model accuracy verification: Combine synthetic data and real data into a mixed training set and evaluate the prediction accuracy of the model on the actual test set. If the accuracy improves or stabilizes, it indicates that the synthetic data has good authenticity. If the accuracy decreases, the data generation process needs to be optimized. (3) Model robustness verification: Train the model multiple times and observe its performance stability under different data combinations. If the training results fluctuate little and the classification accuracy is high, it means that the synthetic data enhances the robustness of the model. Otherwise, the data generation strategy needs to be adjusted. (4) Domain expert review: The feature distribution of synthetic data and the actual data distribution are submitted to experts for review. Experts judge their rationality based on their experience and provide feedback to further optimize the diffusion model parameters to ensure that the synthetic data is authentic and reliable.

7. A coal mill fault detection method based on intelligent deep learning technology according to claim 4, characterized in that: In the fault classification step, the ResNet model performs feature extraction and pattern recognition through multi-layer residual blocks, and classifies and identifies multiple faults based on the multi-frequency features in the vibration signal.

Images