Generator set defect detection method based on deep learning

Abstract

The invention discloses a generator set defect detection method based on deep learning, and belongs to the technical field of generator sets, and the method comprises the following steps: S1, collecting and marking multi-source heterogeneous data; s2, data preprocessing and feature engineering; s3, constructing a deep neural network model: designing and constructing a composite deep learning model architecture fused with multi-modal data processing capability; s4, performing model training and verification; s5, optimizing and lightening the model; s6, edge computing system deployment and real-time detection; s7, carrying out online feedback and model iteration; and S8, generating a visual diagnosis interface and a report. By integrating various heterogeneous data sources such as the vibration sensor, the acoustic sensor and the thermal infrared imager, multi-angle real-time monitoring of the running state of the generator set is realized, and in the acquisition stage, the synchronous acquisition card is utilized to ensure the time precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generator set, and particularly relates to a generator set defect detection method based on deep learning. BACKGROUND

[0002] In the field of generator set state monitoring and fault diagnosis, traditional methods mainly rely on single sensor data or simple threshold alarm mechanism to realize defect recognition. Such methods are difficult to effectively process multi-source heterogeneous data generated in the operation process of the unit, including high-frequency vibration signals, acoustic spectrum, thermal imaging and electrical parameters and various modal information. The existing technology lacks robustness to noise interference and baseline drift, the feature extraction process relies on artificial experience and has limited discriminability, which makes it difficult to accurately capture complex defects early. At the same time, the common mode is usually designed for a specific working condition, and the generalization ability is significantly reduced under variable load, aging equipment and complex environmental interference. The detection result is often lagging and has a high false positive rate, which cannot meet the strict requirements of modern power systems on unit reliability and predictive maintenance.

[0003] Although the current deep learning-based solution has made progress in some scenarios, it still faces the key contradiction between model complexity and edge deployment capability. The existing neural network architecture has a heavy computational burden when processing multi-modal heterogeneous data fusion, and it is difficult to achieve real-time response on resource-constrained industrial edge devices. At the same time, the model training process lacks systematic control over the diversity of field conditions and the quality of data labeling, and is prone to overfitting due to sample imbalance. The traditional deployment mode relies on cloud computing, which has data transmission delay and network reliability risk, and cannot guarantee the timeliness of defect judgment. In addition, the system lacks a closed-loop feedback mechanism, and the model cannot dynamically evolve with device aging, operation feedback and new fault modes. A systematic solution that integrates high-precision multi-source perception, lightweight cross-modal modeling, edge real-time inference and continuous self-optimization is urgently needed to break through the technical bottleneck of generator set intelligent diagnosis; therefore we propose a generator set defect detection method based on deep learning to solve this problem. SUMMARY

[0004] The present application aims to provide a generator set defect detection method based on deep learning to solve the problems raised in the background art.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: A generator set defect detection method based on deep learning, comprising the following steps: S1, multi-source heterogeneous data collection and labeling: through the vibration sensors, acoustic sensors, infrared thermography, current and voltage transformers installed in key parts of the generator set and the operation parameter monitoring system SCADA, real-time synchronous collection of multi-dimensional operating state data, including high-frequency vibration signals, sound spectrum, temperature distribution map, three-phase current and voltage waveform, and working condition parameters such as power, speed, oil pressure, etc.; S2, data preprocessing and feature engineering: strict preprocessing operation is performed on the original collected signals, first wavelet transform or adaptive filtering technology is used to remove environmental noise, electromagnetic interference and baseline drift of the sensor itself; S3, deep neural network model construction: design and build a composite deep learning model architecture that integrates multi-modal data processing capabilities; S4, model training and verification: divide the labeled data set into training set, verification set and test set according to the preset proportion; S5, model optimization and lightweight: in view of the possible overfitting or high computational complexity of the model, a series of optimization measures are implemented; S6, edge computing system deployment and real-time detection: deploy the optimized lightweight model to the high-performance industrial gateway or embedded AI acceleration card installed on the generator set site; S7, online feedback and model iteration: establish a continuous online feedback closed-loop mechanism, the system automatically records all detection results, alarm events, and subsequent operation and maintenance personnel's on-site inspection confirmation and maintenance work orders; S8, visual diagnosis interface and report generation: develop a graphical user interface based on Web and integrate it into the power plant monitoring system; the interface displays the health status overview of the generator set, sensor data flow, model detection results including defect type, location, confidence heat map, historical alarm records and trend analysis.

[0006] Preferably, the S1 comprises the following steps: S101, sensor network deployment and calibration: according to the structural characteristics and common fault modes of the generator set, high-precision accelerometers are scientifically arranged at key positions to measure vibration, acoustic emission sensors to measure abnormal sound, infrared thermography to measure temperature field distribution, current clamps to measure current, voltage probes to measure voltage, and oil pressure and water temperature sensors; all sensors must be calibrated strictly on site after installation to ensure that the range, sensitivity and sampling frequency meet the detection requirements; the vibration signal sampling frequency needs to be greater than 10kHz, and the current signal needs to be greater than 2kHz; and through a synchronous acquisition card, the time synchronization accuracy of all channel data is ensured to be in the microsecond level, laying a foundation for subsequent multi-modal data fusion analysis; S102, Multi-source data synchronous acquisition and storage: Develop or configure a dedicated data acquisition system to synchronously acquire raw signals from multiple sensors through a high-speed data acquisition card according to a unified time reference; Set reasonable sampling parameters to ensure the capture of fault characteristics such as bearing fault characteristic frequencies; Store the acquired raw data waveform data and image frames together with the time stamp device ID operating parameters such as load speed in a structured format in a high-capacity and high-reliability storage device to ensure data integrity and traceability; S103, Expert collaborative data labeling and quality control: Organize a team composed of experienced equipment engineers and experts to use professional signal analysis software and image analysis tools in combination with equipment operation logs and maintenance history records to conduct a detailed review and labeling of the data; The labeling content not only includes defect categories such as inner ring damage and rotor breakage, but also should label the severity level and occurrence position; Establish a strict quality control process including cross-checking and sampling review to ensure the accuracy and consistency of the labeled data.

[0007] Preferably, the S2 comprises the following steps: S201, Signal denoising and outlier processing: Advanced signal processing techniques are used to denoise the ubiquitous noise interference in the original signal; For stationary noise, digital filters such as Butterworth Chebyshev low-pass and band-pass filters are applied; For non-stationary noise, wavelet threshold denoising or empirical mode decomposition is used for separation; At the same time, outliers in the signal such as wild points caused by sensor transient faults are detected and removed to avoid negative effects; S202, Multi-dimensional feature extraction and fusion: Extract rich discriminative features from the pre-processed data of various types; For time series signals, calculate time domain statistical features such as mean, variance, kurtosis, margin factor, frequency domain features such as spectral peak, center frequency, sideband energy, and time-frequency domain features such as wavelet packet energy, entropy, and Hilbert spectrum; For infrared thermal images, extract temperature statistical features such as maximum temperature, average temperature, temperature difference, texture features such as gray level co-occurrence matrix, and key region temperature distribution features; Use feature fusion techniques such as feature splicing to combine heterogeneous features into a comprehensive feature vector; S203, Feature standardization and dimensionality reduction: Since the extracted features have different dimensions and numerical ranges, they must be standardized, such as Z-score standardization; Then, for high-dimensional feature redundancy, apply feature selection methods such as tree model-based feature importance ranking or feature dimensionality reduction methods such as principal component analysis and linear discriminant analysis; The goal is to retain the most informative features, reduce data dimensionality, and improve model efficiency and performance.

[0008] Preferably, in S3, the model architecture design needs to balance between accuracy and efficiency, especially considering edge device deployment; lightweight network modules such as depth separable convolution and efficient attention mechanism should be prioritized; attention-based feature fusion or cross-modal Transformer using complementary information should also be explored; in addition, a regression output head can be added to predict defect severity or a multi-task learning framework can be designed.

[0009] Preferably, in S4, training stability and generalization ability are crucial; special attention should be paid to the class imbalance problem, using oversampling techniques such as SMOTE or undersampling or loss function weighting such as Focal Loss; the validation set should contain difficult samples from different unit operating conditions, and the test set should cover various defect types and severity; industrial standard indicators must be used for comprehensive evaluation to avoid good performance on a single indicator.

[0010] Preferably, S5 includes the following steps: S501, structured pruning and sparse training: apply structured pruning techniques to remove entire filters or channels with smaller contributions; sparse training is required before pruning to introduce L1 or L2 regularization constraints to guide some weights to zero; after pruning, fine-tuning is used to restore accuracy; S502, quantization-aware training and low-bit inference: implement quantization-aware training to simulate quantization operations during forward propagation, while still using floating-point calculations to compute gradients during backpropagation; make the model adapt to quantization precision loss; select appropriate quantization schemes such as symmetric quantization and calibration strategies to achieve inference acceleration; S503, knowledge distillation and model compression: use knowledge distillation to transfer teacher model knowledge to smaller student models; student models learn original labels and probability distributions of teacher model outputs; by minimizing output differences, a deployment model with fast speed and small volume is obtained.

[0011] Preferably, S6 includes the following steps: S601, edge hardware selection and interface development: select appropriate edge hardware platforms such as industrial-grade embedded systems with acceleration modules based on computing requirements, site environment, and cost budget; develop data acquisition interface drivers compatible with sensor networks and industrial buses.

[0012] S602, edge inference engine deployment and optimization: deploy optimized models to edge devices; configure efficient inference engines such as TensorRT for hardware optimization operator fusion and memory utilization; set up a complete inference pipeline including data reception, preprocessing, model inference, and post-processing to ensure that inference delay is less than 100ms.

[0013] S603, Local alarm and cloud cooperation: Real-time defect judgment logic is realized, and when the detection result exceeds the threshold, the local sound and light alarm is triggered; at the same time, the key results and original data fragments are transmitted to the cloud through the industrial network for storage, deep analysis and report generation.

[0014] Preferably, in S7, an automatic data pipeline is designed to integrate alarm data, maintenance results and realize model continuous updating; concept drift, equipment aging and working condition changes are prevented during incremental training to prevent data distribution changes; and the update and deployment need to support a hot update mechanism to ensure the continuity of online detection services.

[0015] The beneficial effects of the present application are: 1、The power generating set defect detection method based on deep learning, through the integration of vibration sensors, acoustic sensors, infrared thermographs and other various heterogeneous data sources, realizes multi-angle real-time monitoring of the power generating set operation state, in the collection stage, the time precision is ensured by using a synchronous acquisition card, and the data is labeled by an expert team, covering defect types and severity levels, ensuring data richness and reliability, advanced noise reduction technology and feature extraction strategy are used in the pretreatment link, effectively separating noise and extracting key time sequence features and image features, providing high-quality training data for subsequent model input, this multi-modal fusion capability enables the system to identify multiple defect modes such as inner ring damage and rotor broken bar, thereby improving the early fault detection rate, compared with the traditional single-source detection method, this method reduces the false alarm rate and enhances the comprehensive diagnosis, providing a more reliable decision basis for power plant operation and maintenance; 2、The power generating set defect detection method based on deep learning, by adopting lightweight modules such as deep separable convolution and efficient attention mechanism, combining attention-based feature fusion strategy, balancing precision and computational efficiency, in the optimization stage, through knowledge distillation and quantization training technology, the model volume is compressed and the inference process is accelerated, making it suitable for edge device deployment, in the deployment link, the industrial-grade hardware platform is selected and matched, and the compatible interface is developed, and the complete flow line is configured with the efficient inference engine, realizing fast data reception and processing, this design ensures that the inference delay remains at a low level, meeting the real-time detection demand, avoiding the shutdown risk caused by the delay problem, finally, the system stably runs on the high-performance industrial gateway or embedded device, greatly improving the on-site response speed, maintaining the continuity and safety of the power generating set operation; 3、The method comprises the following steps: in the preprocessing stage, an advanced noise reduction method such as wavelet transform and empirical mode decomposition is adopted, different types of noise are processed in a targeted manner, and abnormal values are removed, so that the signal purity is ensured; in the feature engineering, multi-dimensional features are extracted and fused into a comprehensive vector; subsequently, the core information is retained through standardization and dimension reduction technology to prevent data redundancy from affecting the performance of the model; in the training stage, the verification set covers difficult samples in different working conditions, the test set contains comprehensive defect types, and a loss function weighting strategy is used to cope with the class imbalance problem; this strict quality control and industrial standard index evaluation prevent overfitting and enhance the generalization ability of the model in a variable field environment, therefore, the system can adapt to various load changes and aging factors, provide stable and reliable detection results, and reduce the misjudgment phenomenon caused by working condition fluctuations; 4、The method comprises the following steps: an automatic feedback pipeline is established, alarm events, operation and maintenance confirmation and maintenance data are integrated, real-time analysis of the detection results is performed, and model incremental updating is triggered; in the iteration process, the concept drift problem is prevented, the incremental training is adapted to equipment aging and working condition changes, the data distribution change does not affect the performance, the updating adopts a hot deployment strategy to ensure the continuity of online services without shutdown maintenance; this closed-loop design enables the system to self-correct and learn new fault modes, continuously improves the detection accuracy over time, reduces the need for manual intervention, and combines a visual interface to intuitively display historical alarm trends and health status; operation and maintenance personnel can conveniently track the evolution track of the system to realize intelligent optimization of the operation and maintenance process, and finally prolong the service life of the generator set; 5、The method comprises the following steps: the real-time display of the unit health status, sensor flow, defect confidence heat map and trend analysis is integrated into the power plant monitoring platform, the interface provides intuitive data visualization such as temperature distribution map and historical alarm record, which facilitates quick positioning of problems, and the report generation function is used to automatically summarize the detection results and subsequent operation and maintenance feedback to form traceable documents; this integrated design simplifies manual operation, operation and maintenance personnel can efficiently handle alarm events through a one-key interface, reduce manual inspection time, and overall, it optimizes the informatization management of the power plant, reduces maintenance cost and labor input, ensures long-term reliable operation of the generator set, and improves the overall production efficiency of the power plant. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a defect detection method for a generator set based on deep learning is provided. DETAILED DESCRIPTION

[0017] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application.

[0018] With reference to Figure 1 A generator set defect detection method based on deep learning includes the following steps: S1, multi-source heterogeneous data acquisition and labeling: through the vibration sensor, acoustic sensor, infrared thermal imager, current-voltage transformer and operation parameter monitoring system SCADA installed at the key parts of the generator set, multi-dimensional operation state data are synchronously collected in real time, including high-frequency vibration signal, sound spectrum, temperature distribution map, three-phase current and voltage waveform and working condition parameters such as power, speed and oil pressure; S2, data preprocessing and feature engineering: strict preprocessing operation is performed on the original collected signal, first, wavelet transform or adaptive filtering technology is used to remove environmental noise, electromagnetic interference and baseline drift of the sensor itself; S3, deep neural network model construction: a composite deep learning model architecture integrating multi-modal data processing capability is designed and built; S4, model training and verification: the labeled data set is divided into training set, verification set and test set according to the preset proportion; S5, model optimization and lightening: in view of the possible overfitting or high computational complexity of the model, a series of optimization measures are implemented; S6, edge computing system deployment and real-time detection: the optimized lightened model is deployed on the high-performance industrial gateway or embedded AI acceleration card installed on the generator set site; S7, online feedback and model iteration: a continuous online feedback closed-loop mechanism is established, the system automatically records all detection results, alarm events and subsequent operation and maintenance personnel's on-site maintenance confirmation and repair work order; S8, visual diagnosis interface and report generation: a graphical user interface based on Web is developed and integrated into the power plant monitoring system; the interface displays the health status overview of the generator set, each sensor data stream, model detection results including defect type position confidence heat map, historical alarm record and trend analysis in real time.

[0019] In the embodiment, S1 includes the following steps: S101, sensor network deployment and calibration: according to the structure characteristics and common failure modes of the generator set, high-precision accelerometers are scientifically arranged at key positions to measure vibration, acoustic emission sensors are arranged to measure abnormal sound, infrared thermographs are arranged to measure temperature field distribution, current clamps are arranged to measure current, voltage probes are arranged to measure voltage, and oil pressure and water temperature sensors are arranged; all sensors must be strictly calibrated on site after installation to ensure that the range, sensitivity and sampling frequency meet the detection requirements; the vibration signal sampling frequency needs to be greater than 10 kHz, and the current signal needs to be greater than 2 kHz; and through a synchronous acquisition card, the time synchronization accuracy of all channel data is ensured to be in the microsecond level, laying a foundation for subsequent multi-modal data fusion analysis; S102, multi-source data synchronous acquisition and storage: develop or configure a dedicated data acquisition system to synchronously acquire original signals from multiple sensors according to a unified time reference through a high-speed data acquisition card; set reasonable sampling parameters to ensure that fault characteristics such as bearing fault characteristic frequencies are captured; store the acquired original data waveforms, data and image frames together with time stamps, device IDs and working condition parameters such as load speed in a structured format in a high-capacity and high-reliability storage device to ensure data integrity and traceability; S103, expert collaborative data labeling and quality control: organize a team composed of experienced equipment engineers and experts to use professional signal analysis software and image analysis tools to combine with equipment operation logs and maintenance history records to carefully review and label the data; the labeling content not only includes defect categories such as inner ring damage and rotor breakage, but also should label the severity level and occurrence position; a strict quality control process including cross-checking and sampling review is established to ensure the accuracy and consistency of the labeled data.

[0020] In this embodiment, S2 includes the following steps: S201, signal noise reduction and outlier processing: advanced signal processing techniques are used to reduce noise interference in the original signal; for stationary noise, digital filters such as Butterworth and Chebyshev low-pass and band-pass filters are applied; for non-stationary noise, wavelet threshold noise reduction or empirical mode decomposition is used for separation; at the same time, outliers in the signal such as sensor transient faults are detected and removed to avoid negative effects; S202, multi-dimensional feature extraction and fusion: rich discriminative features are extracted from the preprocessed data of various types; for time series signals, time domain statistical features, mean, variance, kurtosis, margin factor, frequency domain features, spectral peak, center of gravity frequency, sideband energy, and time-frequency domain features, wavelet packet energy, entropy, and Hilbert spectrum are calculated; for infrared thermal images, temperature statistical features, maximum temperature, average temperature, temperature difference, texture features, gray level co-occurrence matrix features, and key region temperature distribution features are extracted; heterogeneous features are combined into a comprehensive feature vector using feature fusion techniques such as feature splicing; S203, feature standardization and dimensionality reduction: the extracted features have different dimensions and numerical ranges and must be standardized, such as Z-score standardization; then, feature selection is applied to high-dimensional feature redundancy, such as tree model-based feature importance sorting or feature dimensionality reduction methods such as principal component analysis linear discriminant analysis; the goal is to retain the most informative features, reduce data dimensionality, and improve model efficiency and performance.

[0021] In this embodiment, in S3, the model architecture design needs to balance between accuracy and efficiency, especially considering edge device deployment; lightweight network modules should be used first, such as depth separable convolution, efficient attention mechanism, lightweight attention; at the same time, attention-based feature fusion or cross-modal Transformer is explored to utilize complementary information; in addition, a regression output head can be added to predict defect severity or a multi-task learning framework can be designed.

[0022] In this embodiment, in S4, training stability and generalization ability are crucial; special attention should be paid to the class imbalance problem, and oversampling techniques such as SMOTE undersampling or loss function weighting such as Focal Loss should be used; the validation set should contain difficult samples from different machine operating conditions, and the test set should cover various defect types and severity; industrial standard indicators must be used for comprehensive evaluation to avoid good performance on a single indicator.

[0023] In this embodiment, S5 includes the following steps: S501, structured pruning and sparse training: structured pruning techniques are applied to remove entire filters or channels with smaller contributions; sparse training is required before pruning to introduce L1 or L2 regularization constraints to guide some weights to zero; after pruning, fine-tuning is used to restore accuracy; S502, quantization-aware training and low-bit inference: quantization-aware training is implemented to simulate quantization operations during forward propagation, while still using floating-point numbers to calculate gradients during backpropagation; the model is adapted to quantization precision loss; select appropriate quantization schemes such as symmetric quantization and calibration strategies to achieve inference acceleration; S503, knowledge distillation and model compression: knowledge distillation is used to transfer teacher model knowledge to a smaller student model; the student model learns the original label and the probability distribution of the teacher model output; by minimizing the output difference, a deployment model with fast speed and small volume is obtained.

[0024] In this embodiment, S6 includes the following steps: S601, edge hardware selection and interface development: according to the computing requirements, on-site environment, and cost budget, select appropriate edge hardware platforms such as industrial-grade embedded systems equipped with acceleration modules; develop data acquisition interface drivers compatible with sensor networks and industrial buses.

[0025] S602, Edge inference engine deployment and optimization: deploy the optimization model on the edge device; configure an efficient inference engine such as TensorRT to optimize the hardware operator fusion memory utilization; set up a complete inference pipeline including data reception, pre-processing, model inference, post-processing, and ensure that the inference delay is less than 100ms.

[0026] S603, Local alarm and cloud cooperation: realize real-time defect judgment logic, trigger local sound and light alarm when the detection result exceeds the threshold; at the same time, transmit the key results and original data segments to the cloud through the industrial network for storage, deep analysis and report generation.

[0027] In this embodiment, in S7, an automatic data pipeline is designed to integrate alarm data and maintenance results to realize continuous model updating; concept drift caused by equipment aging and working condition changes is prevented during incremental training; and an update and deployment mechanism is required to support hot updating to ensure the continuity of online detection services.

[0028] In this embodiment, by integrating vibration sensors, acoustic sensors, infrared thermographs and other heterogeneous data sources, multi-angle real-time monitoring of the generator set operating state is realized. In the acquisition stage, a synchronous acquisition card is used to ensure time accuracy, and an expert team is used to label data, covering defect types and severity levels, to ensure data richness and reliability. Advanced noise reduction technology and feature extraction strategies are used in the preprocessing link to effectively separate noise and extract key time sequence features and image features, providing high-quality training data for subsequent model input. This multi-modal fusion capability enables the system to identify multiple defect modes such as inner ring damage and rotor breakage, thereby improving early fault detection rate. Compared with traditional single-source detection methods, this method reduces false positives and enhances comprehensive diagnosis, providing more reliable decision-making basis for power plant operation and maintenance.

[0029] The above describes in detail a generator set defect detection method based on deep learning. The principles and implementation methods of the present application are described in this paper using specific examples. The above examples are only used to help understand the method and its core idea. It should be noted that for those skilled in the art, without departing from the principles of the present application, the present application can be improved and modified in several ways, and these improvements and modifications also fall within the scope of the claims of the present application.

Claims

1. A deep learning-based generator set defect detection method, characterized in that, Comprise the following steps: S1, multi-source heterogeneous data collection and labeling: through the vibration sensor, acoustic sensor, infrared thermal imager, current and voltage transformer and operation parameter monitoring system SCADA installed in the key parts of the generator set, real-time synchronous collection of multi-dimensional running state data, including high-frequency vibration signal, sound spectrum, temperature distribution map, three-phase current and voltage waveform and power, speed, oil pressure and other working condition parameters; S2, data preprocessing and feature engineering: strict preprocessing operation is performed on the original collected signal, first wavelet transform or adaptive filtering technology is used to remove environmental noise, electromagnetic interference and baseline drift of the sensor itself; S3, deep neural network model construction: design and build a composite deep learning model architecture that integrates multi-modal data processing capabilities; S4, model training and verification: the labeled data set is divided into training set, verification set and test set according to the preset proportion; S5, model optimization and lightening: in view of the possible overfitting or high calculation complexity of the model, a series of optimization measures are implemented; S6, edge computing system deployment and real-time detection: the optimized lightened model is deployed to the high-performance industrial gateway or embedded AI acceleration card installed on the generator set site; S7, online feedback and model iteration: a continuous online feedback closed-loop mechanism is established, and the system automatically records all detection results, alarm events and subsequent operation and maintenance personnel's on-site maintenance confirmation and maintenance work order; S8, visual diagnosis interface and report generation: develop a graphical user interface based on Web and integrate it into the power plant monitoring system; The interface displays the health status overview of the generator set, sensor data flow, model detection results including defect type, position confidence map, historical alarm record and trend analysis. 2.The deep learning-based generator set defect detection method of claim 1, wherein, The S1 comprises the following steps: S101, sensor network deployment and calibration: according to the structural characteristics and common fault modes of the generator set, high-precision accelerometers are arranged at key positions to measure vibration, acoustic emission sensors to measure abnormal sound, infrared thermal imagers to measure temperature field distribution, current clamps to measure current, voltage probes to measure voltage, and oil pressure and water temperature sensors; after installation, all sensors must be calibrated strictly to ensure that the range, sensitivity and sampling frequency meet the detection requirements, the vibration signal sampling frequency should be greater than 10kHz, and the current signal should be greater than 2kHz; and through the synchronous acquisition card, the time synchronization accuracy of all channel data is ensured to be in the microsecond level, laying a foundation for subsequent multi-modal data fusion analysis; S102, multi-source data synchronous acquisition and storage: develop or configure a dedicated data acquisition system to synchronize the acquisition of original signals from multiple sensors through a high-speed data acquisition card according to a unified time reference; set reasonable sampling parameters to ensure the capture of fault characteristics such as bearing fault characteristic frequency; Store the collected original waveform data and image frames together with the time stamp, equipment ID and working condition parameters such as load speed in a structured format in a high-capacity and highly reliable storage device to ensure data integrity and traceability; S103, Expert collaborative data labeling and quality control: a team composed of experienced equipment engineers and experts uses professional signal analysis software and image analysis tools combined with equipment operation log maintenance history records to conduct detailed review and labeling of data; the labeling content not only includes defect categories such as inner ring damage and rotor broken bar, but also should label severity level and occurrence position; a strict quality control process is established, including cross-checking and sampling review, to ensure the accuracy and consistency of the labeled data. 3.The deep learning-based generator set defect detection method of claim 1, wherein, The S2 includes the following steps: S201, Signal denoising and outlier processing: advanced signal processing techniques are used to denoise the noise interference commonly existing in the original signal; for stationary noise, digital filters such as Butterworth Chebyshev low-pass band-pass filters are applied; for non-stationary noise, wavelet threshold denoising or empirical mode decomposition is used for separation; at the same time, outliers in the signal, such as wild points caused by sensor transient faults, are detected and removed to avoid negative effects; S202, Multi-dimensional feature extraction and fusion: rich discriminative features are extracted from the pre-processed data; for time series signals, time domain statistical features, mean, variance, kurtosis, margin factor, frequency domain features, spectral peak, barycenter frequency, sideband energy, and time-frequency domain features, wavelet packet energy, entropy, and Hilbert spectrum are calculated; for infrared thermal images, temperature statistical features, maximum temperature, average temperature, temperature difference, texture features, gray level co-occurrence matrix features, and key region temperature distribution features are extracted; heterogeneous features are combined into a comprehensive feature vector using feature fusion techniques such as feature splicing; S203, Feature standardization and dimensionality reduction: the extracted features have different dimensions and numerical ranges, so standardization processing such as Z-score standardization is necessary; then, for high-dimensional feature redundancy, feature selection such as tree model-based feature importance sorting or feature dimensionality reduction methods such as principal component analysis linear discriminant analysis are applied; the goal is to retain the most informative features, reduce data dimensionality, and improve model efficiency and performance. 4.The deep learning-based generator set defect detection method of claim 1, wherein, In the S3, model architecture design needs to balance between accuracy and efficiency, especially considering edge device deployment; lightweight network modules such as depth separable convolution, efficient attention mechanism, and lightweight attention should be prioritized; at the same time, attention-based feature fusion or cross-modal Transformer using complementary information is explored; in addition, a regression output head can be added to predict defect severity or a multi-task learning framework can be designed. 5.The deep learning-based generator set defect detection method of claim 1, wherein, In the S4, training stability and generalization ability are crucial; special attention should be paid to the class imbalance problem, using oversampling techniques such as SMOTE, undersampling, or loss function weighting such as Focal Loss; the validation set should contain difficult samples from different unit operating conditions, and the test set should cover various defect types and severity levels; industrial standard indicators must be used for comprehensive evaluation to avoid good performance on a single indicator. 6.The deep learning-based generator set defect detection method of claim 1, wherein, The S5 includes the following steps: S501, Structured pruning and sparse training: structured pruning techniques are applied to remove entire filters or channels with smaller contributions; sparse training is required before pruning to introduce L1 or L2 regularization constraints to guide some weights to zero; after pruning, fine-tuning is used to restore accuracy; S502, Quantization-aware training and low-bit inference: Implementing quantization-aware training simulates quantization operations in the forward propagation process while still using floating-point number calculations to calculate gradients in the backward propagation; adapting the model to the loss of quantization precision; selecting a suitable quantization scheme such as symmetric quantization and calibration strategy to achieve inference acceleration; S503, Knowledge distillation and model compression: Using knowledge distillation to migrate the knowledge of the teacher model to a smaller student model; the student model learns the original label and the probability distribution output by the teacher model; by minimizing the output difference, a deployment model with fast speed and small volume is obtained. 7.The deep learning-based generator set defect detection method of claim 1, wherein, The S6 includes the following steps: S601, Edge hardware selection and interface development: Selecting appropriate edge hardware platforms such as industrial-grade embedded systems equipped with acceleration modules according to computing requirements, on-site environment, and cost budget; developing data acquisition interface drivers compatible with sensor networks and industrial buses. 8.S602, Edge inference engine deployment and optimization: Deploying the optimized model to the edge device; configuring an efficient inference engine such as TensorRT to optimize operators for hardware and memory utilization; setting up a complete inference pipeline including data reception, preprocessing, model inference, post-processing to ensure that the inference delay is less than 100ms. 9.S603, Local alarm and cloud collaboration: Realizing real-time defect judgment logic to trigger local sound and light alarms when the detection result exceeds the threshold; at the same time, key results and original data segments are transmitted to the cloud through industrial networks for storage, deep analysis, and report generation. 10.The deep learning-based generator set defect detection method of claim 1, wherein, In the S7, an automatic data pipeline needs to be designed to integrate alarm data and maintenance results to realize continuous model updates; prevent concept drift and equipment aging caused by changes in data distribution during incremental training; support hot update mechanism to ensure the continuity of online detection services.

Images