A non-contact fault diagnosis method and system for key components of rotating equipment

Abstract

This invention relates to the field of rotating equipment fault diagnosis technology, and discloses a non-contact fault diagnosis method and system for key components of rotating equipment. The method includes: acquiring acoustic signals, surface temperature signals, and displacement vibration signals; constructing a signal dataset and performing data cleaning; extracting Mel-frequency cepstral coefficients from the acoustic signals to form an acoustic feature set; extracting the dynamic mean and trend slope of the sliding window from the temperature signals to form a temperature feature set; and extracting the rotational frequency component and time-domain statistical features from the vibration signals to form a vibration feature set; inputting the three types of features into a feature weight adaptive module to generate a multi-dimensional feature vector; and then inputting the vector into a lightweight model that fuses a lightweight convolutional neural network and a long short-term memory network to output the fault type and the degree of slight, moderate, or severe wear of the rotating equipment, triggering an early warning, and generating a fault report. This invention achieves a more refined health assessment, providing quantitative references for maintenance personnel.

Description

Technical Field

[0001] This invention relates to the field of rotating equipment fault diagnosis technology, and more specifically, to a non-contact fault diagnosis method and system for key components of rotating equipment. Background Technology

[0002] Rotating equipment is widely used in industrial production, including power generation, metallurgy, petrochemicals, shipbuilding, and transportation. The health status of its key components (such as bearings, gears, and shafts) directly affects the overall operational safety and economy. To achieve condition monitoring and fault diagnosis of rotating equipment, existing technologies generally employ intelligent diagnostic methods based on multi-source sensor data. By analyzing multimodal signals collected during equipment operation, the types and severity of faults can be identified.

[0003] For example, existing patent CN120849872A proposes a fault diagnosis method for rotating equipment based on multimodal sensor data and a large model. This method collects multimodal sensor data such as sound waves, temperature, pressure, and vibration during the operation of the rotating equipment. After time alignment, noise suppression, and normalization encoding, a unified data set is formed. Subsequently, the multimodal data undergoes quality assessment and correlation analysis to calculate the perceptual redundancy and heterogeneity between modes. A modal screening mechanism and a feature suppression mechanism are constructed, and key modes are weighted and redundant information is suppressed. The screening weight parameters and redundant feature suppression parameters are input into a fusion-based large model. Fault feature expression is achieved through feature embedding and fusion extraction, and fault classification and diagnostic confidence output are completed based on feature matching degree. This technology has certain advantages in multimodal signal fusion and intelligent recognition. However, shortcomings remain in non-contact fault diagnosis scenarios for critical components of rotating equipment: existing multimodal diagnostic systems mainly rely on traditional contact sensors such as vibration acceleration and temperature resistance sensors, which are difficult to adapt to the long-term detection needs in high-speed rotation or sealed environments. The fusion application of non-contact signals (such as sound waves, infrared temperature, and laser displacement) is still imperfect. Emphasis is placed on global weight allocation and feature suppression between modes, without fully considering the physical characteristics of non-contact signals (sound wave spectral envelope, dynamic changes in surface temperature, displacement vibration frequency components, etc.), resulting in insufficient sensitivity to minor wear or latent faults. While multimodal fusion frameworks based on large models have high accuracy, their large network structure and numerous parameters make them unsuitable for real-time deployment and operation on field edge computing devices. They often only output fault categories, making it difficult to quantitatively distinguish fault severity (such as slight wear, moderate wear, and severe wear), and thus failing to provide fine-grained basis for equipment maintenance decisions.

[0004] Therefore, it is necessary to design a non-contact fault diagnosis method and system for key components of rotating equipment to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a non-contact fault diagnosis method and system for key components of rotating equipment, aiming to solve the problems of inability to acquire signals non-contactly, insufficient targeting of feature extraction, complex model structure and large amount of computation, and lack of hierarchical expression of diagnostic results.

[0006] In one aspect, the present invention proposes a non-contact fault diagnosis method for key components of rotating equipment, comprising: Raw signal data is collected to construct a signal dataset, which includes acoustic signals, surface temperature signals, and displacement vibration signals; and the signal dataset is then cleaned. Based on the cleaned signal dataset, Mel frequency cepstral coefficient features are extracted from the acoustic signal to form an acoustic feature set; dynamic mean features within a sliding time window and trend slope features obtained from temperature change trend analysis are extracted from the surface temperature signal to form a temperature feature set; vibration frequency components of the equipment rotation frequency are extracted from the displacement vibration signal, and several time-domain statistical features are calculated to form a vibration feature set; the acoustic feature set, temperature feature set, and vibration feature set are input into a feature weight adaptive module to generate a multidimensional feature vector. The multidimensional feature vector is input into a lightweight fusion model; the lightweight fusion model is formed by fusing a lightweight convolutional neural network model and a long short-term memory network model; it outputs frequency domain feature representation and temporal feature representation, and performs feature-level fusion to generate fault types and fault degrees; the fault degrees include slight wear, moderate wear and severe wear; When the output fault level is moderate or severe wear, an early warning is triggered, and a fault report is generated. The fault report includes the faulty component, the cause of the fault, and the recommended maintenance cycle.

[0007] Furthermore, data cleaning of the signal dataset includes: The acoustic signal is subjected to wavelet transform noise reduction processing, the surface temperature signal is subjected to outlier removal, and the displacement vibration signal is subjected to Fourier transform to generate frequency domain data.

[0008] Furthermore, when extracting Mel-frequency cepstral coefficient features from the acoustic signal to form an acoustic feature set, the following are included: The acoustic signal is segmented into frames, and the Mel spectrum of each frame is calculated; a logarithmic operation is performed on the Mel spectrum to obtain a logarithmic sequence; a discrete cosine transform is performed on the logarithmic sequence to generate a cepstral coefficient sequence. Among them, the low-order coefficients of the cepstral coefficient sequence represent the shape of the sound spectrum envelope, and the high-order coefficients represent the sound detail features; The low-order coefficients in the cepstral coefficient sequence are extracted as Mel frequency cepstral coefficient features to form the acoustic feature set.

[0009] Furthermore, when extracting the dynamic mean feature within the sliding time window and the trend slope feature obtained based on temperature change trend analysis from the surface temperature signal to form a temperature feature set, the following are included: Temperature data within a preset time-limited sliding window is collected in real time, and the arithmetic mean of the temperature data is calculated as a dynamic mean feature. The time series curve of the temperature data within the sliding window is fitted using the least squares method, and the slope of the time series curve is calculated as a trend slope feature. The dynamic mean feature and the trend slope feature are combined to form the temperature feature set.

[0010] Furthermore, when extracting the vibration frequency component of the equipment rotation frequency from the displacement vibration signal and calculating several time-domain statistical features to form a vibration feature set, the following steps are included: In the frequency domain data, frequency components with amplitudes greater than a preset baseline level and within a preset multiple range of the device's switching frequency are selected, and the amplitudes of the frequency components are extracted as frequency domain features. The peak value, root mean square value, kurtosis, and waveform factor of the displacement vibration signal are calculated as time-domain statistical features. The frequency domain features and time domain statistical features are combined to form the vibration feature set.

[0011] Furthermore, when inputting the acoustic feature set, temperature feature set, and vibration feature set into the feature weight adaptive module to generate a multidimensional feature vector, the process includes: Based on the cleaned signal dataset, the signal-to-noise ratio (SNR) improvement of the acoustic signal is calculated, and the SNR improvement is normalized to the acoustic sensor confidence level. Calculate the standard deviation of the data within the sliding window of the surface temperature signal, and normalize the reciprocal of the standard deviation to obtain the confidence level of the temperature sensor; Calculate the proportion of the amplitude of the frequency component in the displacement vibration signal whose amplitude is greater than the preset baseline level and is within a preset multiple of the device frequency to the total amplitude, and normalize the proportion to the vibration sensor confidence level. Based on the confidence levels of the acoustic sensor, temperature sensor, and vibration sensor, feature weight coefficients are assigned, and the acoustic feature set, temperature feature set, and vibration feature set are weighted and fused; the multidimensional feature vector is generated by feature splicing.

[0012] Furthermore, before inputting the multidimensional feature vector into the lightweight fusion model, the following steps are included: A training dataset is constructed based on historical fault data of key components of rotating equipment. The training data in the training dataset are labeled according to fault type as gearbox gear wear, bearing inner ring spalling, bearing outer ring spalling, and bearing rolling element wear. Each training data is assigned a continuous label value representing the degree of fault. The training dataset is divided into a multi-scale time window sequence, and the multi-dimensional feature vectors in each time window are aligned in the time and frequency domains to make different fault types comparable in the frequency domain feature space. The lightweight fusion model is trained using a joint optimization strategy: the cross-entropy loss function is used to optimize the fault type classification task, the mean squared error loss function is used to optimize the fault degree regression task, and the training process of the two tasks is balanced through a task weight adaptive mechanism. During training, redundant neuron connections are dynamically pruned to preserve network paths sensitive to fault characteristics, resulting in a lightweight fusion model.

[0013] Furthermore, when inputting the multidimensional feature vector into the lightweight fusion model, the following steps are included: The lightweight convolutional neural network model performs channel-wise spatial filtering on the input multidimensional feature vector using deep convolutional layers to capture the local patterns of sound wave spectrum envelope and vibration spectrum energy distribution; it then fuses cross-channel features using point-wise convolutional layers to generate an initial frequency domain feature representation with spatial invariance; finally, it compresses the feature dimension through a global average pooling layer to output the frequency domain feature representation. The Long Short-Term Memory (LSTM) network model captures historical state information by processing the input multidimensional feature vector in chronological order through a forward LSM layer; it captures future state information by processing the input multidimensional feature vector in reverse chronological order through a backward LSM layer; the bidirectional output-input attention mechanism layer characterizes the key time points of fault evolution by calculating the correlation weights between time steps; and it outputs a temporal feature representation through a fully connected layer.

[0014] Furthermore, when generating fault types and fault severity, the following are included: The frequency domain feature representation and the time series feature representation are fused at the feature level. The fault type is output through the classification layer, and the continuous label value representing the degree of fault is output through the regression layer.

[0015] On the other hand, this application also provides a non-contact fault diagnosis system for critical components of rotating equipment, applied to the aforementioned non-contact fault diagnosis method for critical components of rotating equipment, comprising: The multi-source non-contact sensing module integrates a sound sensor, an infrared temperature sensor, and a laser vibration sensor to synchronously acquire sound wave signals, surface temperature signals, and displacement vibration signals of key components of rotating equipment. An edge computing module, equipped with an embedded processor, connects to the multi-source non-contact sensing module via the EtherCAT protocol. It receives acoustic signals, surface temperature signals, and displacement vibration signals, constructs a signal dataset, and performs data cleaning. Based on the cleaned signal dataset, it extracts Mel-frequency cepstral coefficient features from the acoustic signals to form an acoustic feature set; it extracts dynamic mean features within a sliding time window and trend slope features obtained from temperature change trend analysis from the surface temperature signals to form a temperature feature set; and it extracts the vibration frequency components of the device's rotational frequency from the displacement vibration signals and calculates several time-domain statistical features to form a vibration... The system generates a feature set by inputting the acoustic feature set, temperature feature set, and vibration feature set into a feature weight adaptive module to generate a multi-dimensional feature vector. This multi-dimensional feature vector is then input into a lightweight fusion model, which is formed by fusing a lightweight convolutional neural network model and a long short-term memory network model. The model outputs frequency domain feature representations and temporal feature representations, which are then fused at the feature level to generate fault types and fault degrees. The fault degrees include minor wear, moderate wear, and severe wear. When the output fault degree is moderate or severe wear, an early warning is triggered, and a fault report is generated, which includes the faulty component, the cause of the fault, and a recommended maintenance cycle. The interface and communication module includes a USB-C debugging interface, an RJ45 network port, and an RS485 serial port, used to transmit the acoustic wave signals, surface temperature signals, and displacement vibration signals of key components of the rotating equipment synchronously to the edge computing module; and to send a shutdown request to the rotating equipment controller via the RS485 serial port with the output fault type and fault degree.

[0016] Compared with existing technologies, the advantages of this invention are as follows: The sound sensor, infrared temperature sensor, and laser vibration sensor work together to achieve non-contact signal acquisition of key components of rotating equipment, avoiding the problems of complex installation, easy wear and tear, and interference with equipment operation caused by traditional contact sensors. The introduction of multi-source signals enables the diagnostic process to simultaneously perceive acoustic, thermal, and vibration characteristics, comprehensively reflecting the operational health status of components from multiple dimensions. By calculating the confidence level of each sensor signal (including signal-to-noise ratio improvement, temperature fluctuation stability, and vibration spectrum proportion), the weights of sound wave, temperature, and vibration features are dynamically adjusted, achieving adaptive weighted fusion at the feature level. This suppresses the impact of single sensor channel anomalies or noise interference on the diagnostic results and improves the stability of the overall feature representation. The fusion model combines a lightweight convolutional neural network (CNN) and a long short-term memory network (LSTM) dual structure, which can extract both the spatial spectral features of sound wave and vibration signals and capture the temporal evolution features of equipment operation status. Through joint optimization and network pruning strategies, the number of model parameters and computational complexity are reduced, enabling the model to run efficiently on edge computing devices and achieving real-time diagnosis. Through a joint optimization mechanism of classification and regression, the system can simultaneously output fault types (such as gear wear, bearing spalling, etc.) and continuous fault severity indicators (minor wear, moderate wear, severe wear). Compared to traditional systems that can only perform binary judgments, this system enables more refined health assessments, providing quantitative references for maintenance personnel. The edge computing module performs local computation and judgment based on an embedded processor, enabling feature extraction and fault identification without uploading data to the cloud, thus shortening response time. Simultaneously, it links with the device controller via EtherCAT bus and RS485 communication interface, automatically triggering shutdown and alarms when moderate or severe wear is detected, ensuring safe equipment operation. When the hardware is first connected, the software automatically completes device identification (reading the device controller's model information), sensor calibration, and model parameter loading. The entire process requires no manual configuration, achieving "plug and connect to work immediately," reducing deployment time from the usual 2 hours to 10 minutes. When the software diagnoses a "serious fault," it automatically sends a shutdown request to the device controller via the hardware's RS485 serial port (if authorized by the customer), and simultaneously triggers the hardware's buzzer alarm, forming a closed loop of "diagnosis-early warning-emergency handling," reducing the risk of fault escalation. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1A flowchart of a non-contact fault diagnosis method for key components of rotating equipment provided in an embodiment of the present invention; Figure 2 A functional block diagram of a non-contact fault diagnosis system for key components of rotating equipment provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Current technologies for multimodal data acquisition still primarily rely on contact sensors, such as accelerometers and resistance temperature sensors, making it difficult to fully utilize non-contact signals (sound waves, infrared temperature, laser displacement, etc.). This results in the inability to achieve long-term, stable health monitoring in high-speed rotating or sealed equipment scenarios. Emphasis is placed on global weight allocation and feature suppression between modes, without fully integrating the physical characteristics of non-contact signals, such as the acoustic wave spectral envelope, surface temperature change trends, or harmonic components of displacement vibrations. This leads to insufficient sensitivity in diagnosing minor wear or latent faults. While multimodal fusion frameworks based on large models can improve diagnostic accuracy, the large number of network parameters and high computational overhead make it difficult to implement real-time fault detection and early warning on edge computing devices in industrial settings.

[0020] For example, in the monitoring of bearings in wind turbine generators, traditional accelerometers need to be mounted on the bearing housing, which are prone to loosening or failure due to vibration and ambient temperature. Non-contact acoustic sensors and infrared temperature sensors, if not fully utilized, have limited ability to detect early-stage bearing micro-cracks or minor wear. In high-speed gear transmission systems, when initial minor wear appears on the gear surface, the vibration signal changes are weak. If diagnostic methods do not extract and fuse the acoustic spectrum and tooth surface temperature rise characteristics, they often cannot accurately identify this wear in its early stages, potentially leading to serious gear damage later. Using large-scale fusion deep learning models for multimodal data processing requires transmitting data to a remote high-performance server for computation. This not only increases network latency but may also cause real-time warnings to fail due to communication anomalies, affecting the safe operation of the equipment.

[0021] For this, please refer to Figure 1 As shown, this application proposes a non-contact fault diagnosis method for key components of rotating equipment, including: S100: Collect raw signal data to construct a signal dataset, which includes acoustic signals, surface temperature signals, and displacement vibration signals; and perform data cleaning on the signal dataset. S200: Based on the cleaned signal dataset, extract the Mel frequency cepstral coefficient features from the acoustic signal to form an acoustic feature set; extract the dynamic mean features within the sliding time window and the trend slope features obtained based on temperature change trend analysis from the surface temperature signal to form a temperature feature set; extract the vibration frequency components of the equipment rotation frequency from the displacement vibration signal and calculate several time-domain statistical features to form a vibration feature set; input the acoustic feature set, temperature feature set, and vibration feature set into the feature weight adaptive module to generate a multi-dimensional feature vector; S300: Inputs multi-dimensional feature vectors into a lightweight fusion model; the lightweight fusion model is formed by fusing a lightweight convolutional neural network model and a long short-term memory network model; outputs frequency domain feature representation and temporal feature representation, and performs feature-level fusion to generate fault types and fault degrees; fault degrees include slight wear, moderate wear and severe wear; S400: When the output fault level is moderate or severe wear, an early warning is triggered, and a fault report is generated. The fault report includes the faulty component, the cause of the fault, and the recommended maintenance cycle.

[0022] Specifically, three types of raw signals are synchronously acquired through non-contact sensing units: sound waves (acoustic array), surface temperature (infrared thermometry), and displacement / vibration (laser displacement or laser Doppler). All sensors should be timestamped and clock-synchronized (e.g., hardware triggering or precise NTP / PTP). The sampling frequency differences are resampled to obtain a unified time reference. Then, data cleaning is performed on the raw data, including removing breakpoints and lost data, removing outliers based on median / percentile, applying wavelet or spectral domain filtering to reduce noise in the sound wave signal, smoothing / interpolating the temperature sequence, and removing DC and bandpass filtering from the displacement and vibration signals to generate a frequency domain representation. Modal features are extracted from the cleaned data: the acoustic path is first framed (typical frame length 20–40ms, 50% overlap), the power spectrum is calculated and mapped to the Mel scale (e.g., 40 filters), the logarithm is taken and a discrete cosine transform is performed to obtain several MFCC coefficients (usually 13-dimensional and can be extended to include first- and second-order difference features); the temperature path calculates the dynamic mean within a preset sliding window (window length is configurable, such as 1–10 minutes or related to the operating conditions), and fits the time series using the least squares method to obtain the trend slope; the vibration path identifies the equipment rotation frequency (converted from RPM / 60) and its several harmonic components in the frequency domain, selects frequency points with amplitudes higher than the baseline threshold as frequency domain features, and calculates time domain statistics (peak value, root mean square value, kurtosis, waveform factor, etc.) to form a vibration feature set. Three types of features are input into the feature weight adaptive module: This module calculates modal confidence based on the quality indicators of each modality (such as the improvement in acoustic signal-to-noise ratio, the reciprocal of the standard deviation of the temperature sequence, the proportion of target frequency band energy in vibration, etc.) and generates the final weights through normalization (such as softmax or proportional normalization), which are used to weight and concatenate the feature vectors of each modality to obtain a multi-dimensional feature vector. This multi-dimensional vector is input into a lightweight fusion model—the model is composed of a lightweight convolutional network (used to capture local spectral envelopes and cross-channel spatial modes) and a bidirectional LSTM (used to capture forward and reverse temporal dependencies), supplemented by attention mechanisms or channel-wise pointwise convolutions to enhance key modal channels. After global average pooling, classification (fault type) and regression (continuous fault severity score) tasks are completed in parallel through fully connected layers; during training, joint loss (cross-entropy + mean squared error) is used, along with lightweight methods such as pruning / quantization for edge deployment. The model output includes the fault type (such as gear wear, bearing inner / outer ring damage, rolling element wear, etc.) and the continuous fault severity score (e.g., 0–1), and is mapped to "slight / moderate / severe" according to the threshold (example threshold: <0.33 is slight, 0.33–0.66 is moderate, >0.66 is severe, which can be adjusted according to equipment calibration).When the fault severity falls into the moderate or severe range, an early warning is triggered: a structured fault report is automatically generated (including detection time, faulty component, fault type determination, feature matching basis, diagnostic confidence level, possible root cause analysis, and recommended maintenance cycle / response measures - such as "moderate wear: it is recommended to arrange maintenance and increase the frequency of inspections within 24–72 hours", "severe wear: it is recommended to immediately stop the machine for maintenance and replace the component"). This report can be sent to the controller, sent to maintenance personnel, and archived to the log via the field edge interface (RS485 / Ethernet / USB-C). It supports PDF export and cloud storage for post-event analysis and model retraining; it also supports manual review and feedback of repair results for closed-loop optimization.

[0023] This invention solves the problem of non-contact acquisition of sound waves, surface temperature, and displacement vibration signals, followed by data cleaning, denoising, and normalization. It overcomes the limitations of traditional diagnostic methods, which rely on contact sensors, are highly susceptible to installation location and environmental influences, and suffer from high data noise. This enables stable and continuous acquisition even in high-speed rotation or sealed environments. By extracting the Mel-frequency cepstral coefficients of sound waves, the dynamic mean and trend slope of temperature, and the rotational frequency components and temporal statistical features of vibration, and inputting these into a feature weight adaptive module to generate multi-dimensional feature vectors, it effectively addresses the issues of information redundancy, weak correlation interference, and differences in the dimensions of features across different modes in multimodal data, ensuring the prominent expression of fault-sensitive features. A lightweight model formed by fusing a lightweight convolutional neural network and a long short-term memory network achieves efficient fusion of multi-dimensional feature vectors and captures spatiotemporal features. This solves the problems of large models having high computational cost, difficulty in deployment on field edge devices, and inability to respond in real time. Simultaneously, it can output both the fault type and the degree of continuous fault, overcoming the limitations of traditional methods that only provide category judgments and cannot quantify fault severity. When the diagnostic results show moderate or severe wear, an early warning is automatically triggered and a detailed fault report is generated, providing information on the faulty component, the cause of the fault, and maintenance cycle suggestions. This effectively solves the problem of lacking rapid decision-making basis and being unable to prevent further damage to equipment in a timely manner during online monitoring.

[0024] The working process and principle of this application are as follows: Acoustic wave signals, surface temperature signals, and displacement vibration signals are simultaneously acquired using multimodal non-contact sensors. The acquired raw signals undergo data cleaning and preprocessing, including outlier removal, noise suppression, frequency domain filtering, and normalization, to improve data quality and consistency, providing a reliable foundation for subsequent feature extraction. Modal features are extracted from the cleaned signals: For acoustic waves, Mel frequency cepstral coefficients (MFCCs) are obtained through framing, Mel spectrum mapping, and discrete cosine transform, reflecting the envelope and detailed characteristics of the sound spectrum; for surface temperature signals, dynamic mean features are calculated within a sliding time window, and slope features are obtained by fitting temperature change trends, reflecting abnormal thermal behavior; for displacement vibration signals, frequency components of the device's rotational frequency and harmonics are extracted, and time-domain statistics such as peak value, root mean square value, kurtosis, and waveform factor are calculated to form a vibration feature set. The three types of features are input into a feature weight adaptive module, which calculates weights based on the signal-to-noise ratio and correlation of each modality, generating a multidimensional feature vector to enhance key modal features and suppress redundant weakly correlated features. Multidimensional feature vectors are input into a lightweight fusion model, which is composed of a lightweight convolutional neural network and a bidirectional long short-term memory network (Bi-LSTM). The convolutional network extracts local spatial features and spectral patterns of each modality, while the Bi-LSTM captures forward and backward state information in the time series. An attention mechanism highlights key time step features, and finally, a fully connected layer simultaneously outputs the fault type and the degree of continuous fault. An early warning is triggered based on the fault degree. When the diagnostic result is moderate or severe wear, a structured report containing the faulty component, the cause of the fault, and the recommended maintenance cycle is automatically generated and can be sent to the field controller or maintenance personnel via an interface, enabling real-time online monitoring and preventative maintenance.

[0025] As a preferred embodiment, the specific implementation of this application is as follows: Taking an industrial high-speed centrifugal pump as an example, the practical application of this method can be specifically illustrated. The operating acoustic signals of the pump shaft and impeller are collected using a non-contact acoustic sensor. The temperature of the pump body surface and bearings is measured using an infrared thermometer. Simultaneously, a laser displacement sensor is used to acquire the vibration signal of the pump shaft. All signals undergo clock synchronization, resampling, outlier removal, and filtering to form a high-quality signal dataset. The acoustic signals are processed through framing, Mel-frequency spectrum mapping, and discrete cosine transform to obtain MFCC features, which reflect the acoustic characteristics of the impeller and bearing during operation. The temperature signal is used to calculate the dynamic mean within a 1-minute sliding window, and the trend slope is obtained through linear fitting, reflecting abnormal heating of the bearing or mechanical seal. The vibration signal extracts the frequency components of the pump shaft rotation frequency and its harmonics, and simultaneously calculates the peak value, root mean square value, kurtosis, and waveform factor to form a vibration feature set. Subsequently, the three types of features are input into a feature weight adaptive module, which assigns weights based on the acoustic signal-to-noise ratio, temperature fluctuation standard deviation, and vibration frequency band energy proportion to generate a multi-dimensional feature vector. The model integrates a lightweight convolutional neural network with a multidimensional vector input and a bidirectional long short-term memory network. The convolutional network captures local acoustic and vibration patterns, while the Bi-LSTM captures time-series state changes and outputs a score indicating the type and continuity of pump bearing wear. When the diagnostics indicate moderate or severe wear, an early warning is automatically triggered, and a detailed report is generated, including the cause of the damaged bearing, possible abnormal vibration due to impeller imbalance, abnormal temperature rise due to heat concentration, etc., and a suggested maintenance cycle is provided. For example, moderate wear can be scheduled for maintenance within 24 hours, while severe wear requires immediate shutdown and bearing replacement.

[0026] Through the above-mentioned scheme, the data cleaning and preprocessing of this application improves signal consistency and availability and reduces noise interference; by extracting and weighting adaptively processing sound wave, temperature and vibration features, fault-sensitive features are highlighted, while redundancy and weak related information are suppressed, thus improving the fusion quality of multimodal data; the fusion of lightweight convolutional neural network and long short-term memory network can capture spatial and temporal features simultaneously, achieving accurate prediction of fault type and continuous fault degree; it can automatically trigger early warning based on moderate or severe wear and generate a report containing faulty parts, causes and maintenance suggestions, supporting rapid decision-making and preventive maintenance.

[0027] This application further proposes methods for cleaning signal datasets, including: The acoustic signal is denoised by wavelet transform, the surface temperature signal is removed by outlier removal, and the displacement vibration signal is generated by Fourier transform to generate frequency domain data.

[0028] Specifically, during data cleaning of the signal dataset, wavelet transform is used for noise reduction of the acoustic signal. Wavelet transform decomposes the acoustic signal into coefficients of different frequency bands, suppressing high-frequency noise components while retaining effective low-frequency signal characteristics, thereby eliminating environmental noise and sensor interference. For the surface temperature signal, outliers are detected and transient peaks or measurement error points exceeding statistical thresholds are removed to ensure the continuity and reliability of the temperature data. The displacement vibration signal is converted into frequency domain data through Fourier transform, clearly presenting the energy distribution of different frequency components, facilitating subsequent extraction of equipment rotation frequency and harmonic characteristics, and enabling accurate analysis of vibration characteristic changes.

[0029] Through the above technical solution, this application can improve the quality and consistency of multimodal signals and reduce the interference of noise and outliers on subsequent feature extraction by using the above data cleaning method; the frequency domain data after Fourier transform is convenient for capturing the frequency switching characteristics of the equipment, ensuring that vibration anomalies can be accurately identified.

[0030] This application further proposes methods for extracting Mel-frequency cepstral coefficient features from acoustic signals to form an acoustic feature set, including: The acoustic signal is segmented into frames, and the Mel spectrum of each frame is calculated. Logarithmic operation is performed on the Mel spectrum to obtain a logarithmic sequence. Discrete cosine transform is performed on the logarithmic sequence to generate a cepstral coefficient sequence. Among them, the lower-order coefficients of the cepstral coefficient sequence represent the shape of the sound spectrum envelope, while the higher-order coefficients represent the sound detail features. The lower-order coefficients in the cepstral coefficient sequence are extracted as the characteristics of the Mel frequency cepstral coefficients to form the acoustic wave feature set.

[0031] Specifically, when extracting Mel-frequency cepstral coefficients (MFCC) features from acoustic signals of key components of rotating equipment, the continuously acquired acoustic signals are preprocessed, including removing DC components and normalizing amplitudes to eliminate the influence of sensor offset and amplitude differences. The signal is then framed, dividing the long, continuous acoustic signal into several short, partially overlapping frames of fixed length (typically 20-40 milliseconds in length, with a frame shift of 10-20 milliseconds), ensuring that the signal within each frame is approximately stable, which is beneficial for short-time frequency analysis. For each frame, its spectrum is calculated using a Fast Fourier Transform (FFT), and the resulting linear frequency spectrum is then mapped to a Mel-frequency scale to form the Mel spectrum. The Mel-frequency scale mimics the sensitivity of the human ear to different frequencies, resulting in higher resolution of low-frequency components and helping to capture low-frequency vibrations or friction anomalies during equipment operation. Finally, a logarithmic operation is performed on the Mel spectrum to convert the amplitude information to a logarithmic scale, thereby enhancing weak frequency components while compressing the amplitude range and improving the stability of the features. Then, a Discrete Cosine Transform (DCT) is performed on the log-Mel spectrum to obtain a cepstral coefficient sequence. In this sequence, the lower-order coefficients mainly characterize the overall spectral envelope of the sound, reflecting the energy distribution trend and main vibration modes of the sound wave; the higher-order coefficients reflect minute details, such as high-frequency anomalies like local friction and intermittent collisions. To reduce the feature dimensionality and highlight key fault information, the lower-order cepstral coefficients are usually truncated to form the final MFCC feature set. This feature set can preserve the overall acoustic characteristics of the rotating component's operation while removing high-frequency noise and irrelevant details.

[0032] In Mel-frequency cepstral coefficients (MFCC) features, the distinction between low-order and high-order coefficients is primarily based on the cepstral sequence index and the spectral information they reflect. Typically, the first few coefficients of the cepstral sequence (e.g., c1 to c12 or c13) are defined as low-order coefficients. These coefficients mainly characterize the overall envelope shape of the sound wave spectrum and can reflect the macroscopic vibration state and main fault characteristics of rotating equipment. Higher-order coefficients in the sequence are high-order coefficients, reflecting subtle details of the sound, such as local friction or impact sounds. The method for determining low-order and high-order coefficients usually employs a fixed truncation method. This involves setting a truncation index based on experience or experimentation, retaining the first part of the cepstral sequence as low-order coefficients for feature set construction, while high-order coefficients can be ignored as noise or auxiliary information.

[0033] As a preferred embodiment, the solution of this application is implemented as follows: Taking the main shaft bearing of a wind turbine generator as an example, the bearing acoustic signal collected during operation includes high-frequency friction noise and low-frequency rotational noise. After dividing the signal into frames, the Mel spectrum of each frame is calculated, and the energy peaks generated by the wear of the rolling elements can be clearly captured in the low-frequency band. After taking the logarithm of the Mel spectrum and performing a discrete cosine transform, the resulting low-order cepstral coefficients can accurately characterize the overall envelope change of the bearing vibration, so that the abnormal friction sound characteristics are preserved and highlighted, forming an acoustic feature set that can be used for diagnosis.

[0034] Through the above technical solution, this application can effectively compress complex acoustic signals into low-dimensional feature vectors by using the MFCC feature extraction method, while retaining the key spectral envelope information related to the fault, making the acoustic anomaly features more prominent; it enhances the sensitivity to early wear or local defects, and improves the accuracy and robustness of fault detection.

[0035] This application further proposes extracting dynamic mean features within a sliding time window from surface temperature signals and trend slope features obtained based on temperature change trend analysis, forming a temperature feature set including: Temperature data is collected in real time within a preset time sliding window, and the arithmetic mean of the temperature data is calculated as a dynamic mean feature. The time series curve of temperature data within a sliding window is fitted using the least squares method, and the slope of the time series curve is calculated as a trend slope feature. The dynamic mean feature and the trend slope feature are combined to form a temperature feature set.

[0036] Specifically, when extracting features from the surface temperature signals of key components of rotating equipment, a sliding time window is set, typically ranging from tens of seconds to several minutes in length, to ensure that the temperature data within the window reflects the short-term operating status of the equipment. Temperature data within the window is acquired in real time and preprocessed, including outlier removal and normalization, to eliminate the influence of sensor measurement errors and environmental interference. The arithmetic mean of the temperature data within the window is calculated to obtain a dynamic mean feature, which reflects the overall thermal state of the equipment over a short period. The temperature time series within the sliding window is linearly fitted using the least squares method, and the slope of the fitted curve is used as a trend slope feature to quantify the temperature change trend over time, such as the rate of temperature increase or decrease. The dynamic mean feature and the trend slope feature are combined to form a complete temperature feature set.

[0037] As a preferred embodiment, the solution of this application is implemented as follows: For example, in the monitoring of spindle bearing temperature, the sliding time window is set to 60 seconds, and a temperature signal is collected once per second, resulting in 60 temperature data points. Outliers, such as instantaneous measurement noise or sensor failure data, are removed, and the arithmetic mean of these 60 data points is calculated to obtain the dynamic mean characteristic of the current window, such as a mean of 75℃. Least squares linear fitting is performed on these 60 data points, and the slope of the fitted curve is 0.1℃ / s, indicating that the temperature is slowly rising. Combining the mean of 75℃ with the slope of 0.1℃ / s forms the temperature feature vector of this window, which is used to determine whether the bearing is in normal operating condition or has a potential overheating risk.

[0038] Through the above technical solution, this application, by employing dynamic mean and trend slope features within a sliding window, can capture the instantaneous level and trend of equipment temperature, making the temperature features more sensitive to faults. The dynamic mean feature reflects the average thermal state of the equipment in a short period of time, avoiding interference from instantaneous abnormal fluctuations; the trend slope feature reveals the temperature change trend, identifying potential overheating or cooling anomalies.

[0039] This application further proposes extracting the vibration frequency component of the equipment rotation frequency from the displacement vibration signal and calculating several time-domain statistical features to form a vibration feature set, including: In the frequency domain data, select frequency components whose amplitude is greater than the preset baseline level and is within a preset multiple of the device's switching frequency, and extract the amplitude of the frequency components as frequency domain features. The peak value, root mean square value, kurtosis, and waveform factor of the displacement vibration signal are calculated as time-domain statistical features. The frequency domain features and time domain statistical features are combined to form a vibration feature set.

[0040] Specifically, when extracting features from the displacement vibration signals of rotating equipment, a Fourier transform is performed on the original vibration signal to convert the time-domain signal into frequency-domain data, revealing the energy distribution of different frequency components. Frequency components with amplitudes greater than a preset baseline level and located within the equipment's rotational frequency and its preset multiples are selected from the frequency-domain data. These frequency components typically correspond to the periodic vibration characteristics of key components such as bearings, gears, or rotors. Their amplitudes are extracted as frequency-domain features to characterize the equipment's rotational state and potential anomalies. Several statistical features are calculated on the time-domain signal, including peak value (reflecting instantaneous impact or abnormal vibration), root mean square value (reflecting the overall energy of the signal), kurtosis (reflecting the intensity of sharp vibration events), and waveform factor (reflecting the smoothness or abrupt changes of the signal waveform), quantifying the vibration characteristics from different perspectives. The frequency-domain features are combined with the time-domain statistical features to form a complete vibration feature set.

[0041] When extracting vibration features from displacement vibration signals, the preset baseline level is usually set based on the steady-state vibration amplitude or no-load vibration amplitude of the equipment's historical operating conditions. The mean value plus a certain number of standard deviations can be calculated through statistical analysis to use as a threshold to distinguish between normal and abnormal vibration components. The preset multiple range of the equipment's rotation frequency is set based on the working structural characteristics and mechanical transmission relationship of key rotating components. For example, the rotation frequency of components such as bearings, gears, or rotors and their integer multiples are usually closely related to fault characteristics. Therefore, a range of 1 to 5 times the rotation frequency can be selected as the screening interval to ensure that the extracted frequency components can cover the periodic vibration characteristics of key components.

[0042] As a preferred embodiment, the solution of this application is implemented as follows: For example, in bearing vibration monitoring, the displacement vibration signal of the bearing is acquired through an accelerometer. After performing a Fourier transform on the acquired signal, a spectrum of 0–500 Hz is obtained. The device frequency is 50 Hz, and the preset multiple range is 1–5 times. Components with amplitudes greater than the baseline threshold at frequencies such as 50 Hz, 100 Hz, and 150 Hz are selected, yielding corresponding amplitudes such as 0.02g, 0.015g, and 0.01g, which are used as frequency domain features. The peak value of the original time domain signal is calculated to be 0.05g, the root mean square value is 0.018g, the kurtosis is 5.2, and the waveform factor is 1.4. These time domain statistical features reflect the overall energy and instantaneous impact characteristics of the vibration signal. The frequency domain features are combined with the time domain statistical features to form a vibration feature vector for that time period, which is used to determine whether the bearing has faults such as wear, loosening, or imbalance.

[0043] Through the above technical solution, this application can capture the vibration characteristics of rotating equipment by combining frequency domain and time domain features. The frequency domain features highlight the abnormal signals of key components corresponding to the equipment rotation frequency and its harmonic components, while the time domain statistical features reflect the overall energy, impact, and waveform characteristics of the vibration signal. This enhances the sensitivity to early faults and helps to detect potential problems in advance before the vibration signal reaches an obvious abnormal level.

[0044] This application further proposes an adaptive feature weighting module that inputs acoustic wave feature sets, temperature feature sets, and vibration feature sets into the feature weighting module to generate multidimensional feature vectors, including: Based on the cleaned signal dataset, the signal-to-noise ratio (SNR) improvement of the acoustic signal is calculated, and the SNR improvement is normalized to the confidence level of the acoustic sensor. Calculate the standard deviation of the data within the sliding window of the surface temperature signal, and normalize the reciprocal of the standard deviation to obtain the confidence level of the temperature sensor; In the displacement vibration signal, the proportion of the amplitude of the frequency component whose amplitude is greater than the preset baseline level and is within a preset multiple of the equipment frequency is calculated, and the proportion is normalized to the confidence level of the vibration sensor. Based on the confidence levels of the acoustic sensor, temperature sensor, and vibration sensor, feature weight coefficients are assigned, and the acoustic feature set, temperature feature set, and vibration feature set are weighted and fused; a multidimensional feature vector is generated by feature splicing.

[0045] Specifically, when inputting the acoustic feature set, temperature feature set, and vibration feature set into the feature weight adaptive module, this module performs reliability and sensitivity assessments on the signal data for each mode to dynamically allocate weights during the fusion process, thereby improving the effectiveness and accuracy of diagnostic features. For acoustic signals, the module quantifies signal quality by calculating the signal-to-noise ratio (SNR) boost value. A higher SNR indicates a larger proportion of effective fault information and less noise interference in the signal. This SNR value is normalized to obtain the acoustic sensor confidence score, which is used to reflect its importance in fault diagnosis during feature weighting. For temperature signals, the module calculates the standard deviation of temperature data within a sliding time window. A smaller standard deviation indicates higher temperature data stability and reflects the reliability of equipment surface temperature measurement. Its reciprocal is normalized to obtain the temperature sensor confidence score, so that highly stable temperature features are given higher weights during the fusion process. For displacement vibration signals, the module statistically analyzes the proportion of frequency components with amplitudes greater than a preset baseline and within the range of multiples of the equipment's rotational frequency. This proportion characterizes the vibration sensitivity of key rotating components. This proportion is normalized to obtain the vibration sensor confidence score, thereby highlighting the vibration features most valuable for fault identification during feature fusion. Based on the confidence levels of the three types of sensors, the feature weight coefficients of each sensor are dynamically calculated. The sound wave, temperature and vibration features are then weighted and fused according to their weights. A multi-dimensional feature vector of uniform length is generated by feature splicing. This vector not only integrates the main fault information of the multi-modal signals, but also fully considers the reliability differences of each mode.

[0046] As a preferred embodiment, the solution of this application is implemented as follows: For example, for monitoring a high-speed rotary pump, the bearing operating noise collected by the acoustic sensor is processed by wavelet denoising, and the signal-to-noise ratio improvement is calculated to be 12dB, which is normalized to a confidence level of 0.85; the temperature sensor measures a temperature standard deviation of 0.6℃ within a 10-second sliding window, and its reciprocal is normalized to a confidence level of 0.90; the vibration sensor measures a key rotational frequency and its harmonic amplitude as a proportion of the total amplitude of 0.65, which is normalized to a confidence level of 0.75. The confidence levels of the three sensors are used to allocate their respective feature weights, and the acoustic, temperature, and vibration features are weighted and fused. A 30-dimensional multi-dimensional feature vector with uniform length and rich information is generated through feature concatenation.

[0047] Through the above technical solution, this application, via a feature weight adaptive module, can dynamically reflect the reliability and sensitivity of different sensor signals, assigning greater weight to high-confidence features and suppressing the influence of low-confidence or interfering features. This improves the feature representation quality after multimodal data fusion and enhances the accuracy of fault diagnosis for rotating equipment. This application further proposes inputting multi-dimensional feature vectors into a lightweight fusion model, including: A training dataset is constructed based on historical fault data of key components of rotating equipment. The training data in the training dataset are labeled according to fault type as gearbox gear wear, bearing inner ring spalling, bearing outer ring spalling, and bearing rolling element wear. Each training data is assigned a continuous label value representing the degree of fault. The training dataset is divided into a multi-scale time window sequence, and the multi-dimensional feature vectors in each time window are aligned in the time and frequency domains to make different fault types comparable in the frequency domain feature space. A joint optimization strategy is adopted to train a lightweight fusion model: the cross-entropy loss function is used to optimize the fault type classification task, the mean squared error loss function is used to optimize the fault degree regression task, and the training process of the two tasks is balanced through a task weight adaptive mechanism. During training, redundant neuron connections are dynamically pruned to preserve network paths sensitive to fault characteristics, resulting in a lightweight fusion model.

[0048] Specifically, before inputting the multi-dimensional feature vectors into the lightweight fusion model, a comprehensive training dataset needs to be constructed based on historical fault data of key components of rotating equipment. This dataset covers different types of fault samples, such as gearbox gear wear, bearing inner ring spalling, bearing outer ring spalling, and bearing rolling element wear. Each sample is not only labeled with the fault type but also assigned a continuous fault severity label to accurately depict the evolution process from slight wear to severe damage. To enhance the model's understanding of time-series features, the training data is divided into multi-scale time window sequences. The multi-dimensional feature vectors within each time window are aligned in the time-frequency domain, making different fault types comparable in the frequency domain feature space and avoiding biases caused by sampling differences and signal length differences. A joint optimization strategy is used to train the lightweight fusion model, where the cross-entropy loss function is used to optimize the fault type classification task, and the mean squared error loss function is used to optimize the fault severity regression task. A task weight adaptive mechanism is used to balance the training process of the two tasks, thereby simultaneously improving classification accuracy and regression accuracy. During training, the model dynamically prunes redundant neuron connections, preserving network paths sensitive to key fault features.

[0049] Through the above technical solution, this application employs a joint optimization strategy to train a lightweight fusion model, enabling the model to simultaneously perform fault type classification and fault severity regression tasks. The cross-entropy loss function optimizes the classification task, while the mean squared error loss function optimizes the regression task. An adaptive task weight mechanism balances the training process for both tasks, thereby improving diagnostic accuracy. During training, redundant neuron connections are dynamically pruned, preserving network paths sensitive to fault characteristics. This not only reduces the model's computational load and storage requirements but also improves its operational efficiency in real-world online monitoring environments.

[0050] This application further proposes methods for inputting multidimensional feature vectors into a lightweight fusion model, including: The lightweight convolutional neural network model performs channel-wise spatial filtering on the input multidimensional feature vectors using deep convolutional layers to capture local patterns of sound wave spectrum envelope and vibration spectrum energy distribution; it then fuses cross-channel features using point-wise convolutional layers to generate an initial frequency domain feature representation with spatial invariance; finally, it compresses the feature dimension through a global average pooling layer to output the frequency domain feature representation. The Long Short-Term Memory (LSTM) network model captures historical state information by processing the input multidimensional feature vectors in chronological order through a forward LSM layer; it captures future state information by processing the input multidimensional feature vectors in reverse chronological order through a backward LSM layer; the bidirectional output-input attention mechanism layer characterizes the key time points of fault evolution by calculating the correlation weights between time steps; and it outputs the temporal feature representation through a fully connected layer.

[0051] Specifically, when inputting multi-dimensional feature vectors into the lightweight fusion model, a lightweight convolutional neural network (CNN) is used to perform channel-by-channel filtering of the spatial dimensions of the multi-dimensional feature vectors. Deep convolutional layers can effectively capture local patterns such as the acoustic spectral envelope and vibration spectral energy distribution. By fusing cross-channel features through point-by-point convolution, an initial frequency domain feature representation with spatial invariance is generated, thereby reducing interference caused by sensor position differences or signal amplitude variations. After global average pooling layers compress the features, the output frequency domain feature representation is obtained, reducing the computational load of the model while retaining key fault-sensitive features, thus improving the model's adaptability to complex operating conditions. For example, for a wind turbine in operation, after processing acoustic and vibration features through a CNN, local spectral patterns of abnormal gear meshing or bearing rolling element damage can be clearly identified. When processing time-series features, the lightweight fusion model employs a bidirectional long short-term memory (BiLSTM) network. The forward LLSTM layer processes multidimensional feature vectors sequentially, capturing historical state information and identifying early signals of fault evolution. The backward LLSTM layer processes data in reverse chronological order to obtain future trend information, thus combining historical and future states to model key fault points. The bidirectional output input attention mechanism layer emphasizes the importance of key time points for fault diagnosis by calculating correlation weights between time steps, enabling the model to focus on specific stages of abnormal signal occurrence. For example, in bearing damage monitoring, BiLSTM can identify early minor vibration anomalies and, combined with trend information, determine whether they have progressed to moderate or severe wear. The fully connected layers map the time-series feature representation into high-dimensional feature outputs that can be used for classification and regression, achieving fault type identification and fault severity assessment.

[0052] Through the above technical solution, this application inputs multi-dimensional feature vectors into a lightweight fusion model, effectively extracts the local patterns of acoustic wave spectrum envelope and vibration spectrum energy distribution through a lightweight convolutional neural network, and achieves spatially invariant initial frequency domain feature representation by fusing cross-channel features through pointwise convolution. At the same time, global average pooling compresses feature dimensions and improves computational efficiency. Long short-term memory network further captures historical and future state information, and combines input attention mechanism to focus on key time points of fault evolution, achieving accurate characterization of fault development trends.

[0053] This application further proposes output frequency domain feature representation and time-series feature representation, and performs feature-level fusion to generate fault types and fault degrees, including: The frequency domain feature representation and the time series feature representation are fused at the feature level. The fault type is output through the classification layer, and the continuous label value representing the degree of fault is output through the regression layer.

[0054] Specifically, the frequency domain feature representation extracted by a lightweight convolutional neural network is fused with the temporal feature representation obtained by a long short-term memory network at the feature level. Spatial and temporal information is integrated through feature concatenation or weighted fusion to form a more comprehensive and representative multidimensional feature representation. These fused features include not only the spectral envelope changes of acoustic signals and the local patterns of vibration frequency energy distribution, but also the dynamic trends of temperature and vibration signals over time, accurately characterizing the operating status and potential anomalies of key components of rotating equipment. The fused features are input into the classification layer, which outputs predicted probabilities for various fault types, such as gearbox gear wear, bearing inner ring spalling, bearing outer ring spalling, and bearing rolling element wear, through softmax or similar activation functions. The final fault type is determined by the maximum probability or a threshold. The fused features are input into the regression layer, which outputs continuous values ​​representing the degree of fault through mean square error optimization or continuous label mapping. These values ​​can be classified into slight wear, moderate wear, and severe wear.

[0055] As a preferred embodiment, the solution of this application is implemented as follows: For example, in the monitoring of the main shaft bearing of a wind turbine generator set, a lightweight convolutional neural network is used to obtain the frequency domain characteristics of the bearing vibration signal, including the amplitude distribution of the main frequency components within the equipment speed and its octave range, as well as the Mel frequency cepstral coefficient characteristics of the acoustic signal; a long short-term memory network captures the dynamic change trends and historical state information of temperature, vibration, and acoustic signals during continuous operation. After feature-level fusion of these frequency domain features and time-series features, the classification layer outputs the predicted fault type, such as rolling element wear or inner ring spalling, and calculates the probability value of each type to determine the most likely fault category; the regression layer outputs continuous fault severity values, for example, a bearing wear severity of 0.65 corresponds to a moderate wear level.

[0056] Through the above technical solutions, the fault types output by the classification layer of this application can not only accurately identify specific fault categories such as rolling element wear, bearing inner ring spalling, or gear wear, but also provide corresponding probability or confidence level indicators, improving the reliability and interpretability of diagnosis. The continuous fault degree labels output by the regression layer can accurately quantify the severity of wear or damage, from slight to moderate to severe wear, providing a quantitative basis for maintenance decisions.

[0057] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a non-contact fault diagnosis system for critical components of rotating equipment, applied to the aforementioned non-contact fault diagnosis method for critical components of rotating equipment, including: The multi-source non-contact sensing module integrates a sound sensor, an infrared temperature sensor, and a laser vibration sensor to synchronously acquire sound wave signals, surface temperature signals, and displacement vibration signals of key components of rotating equipment. The edge computing module, equipped with an embedded processor, connects to the multi-source non-contact sensing module via the EtherCAT protocol. It receives acoustic signals, surface temperature signals, and displacement vibration signals, constructs a signal dataset, and performs data cleaning. Based on the cleaned signal dataset, it extracts Mel-frequency cepstral coefficient features from the acoustic signals to form an acoustic feature set; it extracts dynamic mean features within a sliding time window and trend slope features obtained from temperature change trend analysis from the surface temperature signals to form a temperature feature set; and it extracts the vibration frequency components of the equipment's rotational frequency from the displacement vibration signals and calculates several time-domain statistical features to form a vibration feature set. Feature sets: The acoustic wave feature set, temperature feature set, and vibration feature set are input into the feature weight adaptive module to generate a multi-dimensional feature vector. The multi-dimensional feature vector is then input into a lightweight fusion model. The lightweight fusion model is formed by fusing a lightweight convolutional neural network model and a long short-term memory network model. The output is a frequency domain feature representation and a time series feature representation, which are then fused at the feature level to generate the fault type and fault degree. The fault degree includes slight wear, moderate wear, and severe wear. When the output fault degree is moderate wear or severe wear, an early warning is triggered, and a fault report is generated. The fault report includes the faulty component, the cause of the fault, and the recommended maintenance cycle. The interface and communication module includes a USB-C debugging interface, an RJ45 network port, and an RS485 serial port. It is used to transmit the acoustic wave signals, surface temperature signals, and displacement vibration signals of key components of the rotating equipment to the edge computing module synchronously; and to send the output fault type and fault degree to the rotating equipment controller via the RS485 serial port to send a shutdown request.

[0058] Specifically, this invention achieves high-precision real-time monitoring and intelligent diagnosis of the operating status of key components through the organic integration of a multi-source non-contact sensing module, an edge computing module, and an interface and communication module. The multi-source non-contact sensing module integrates an acoustic wave sensor, an infrared temperature sensor, and a laser vibration sensor, enabling simultaneous acquisition of acoustic wave signals, surface temperature signals, and displacement vibration signals from key components of rotating equipment, ensuring that the acquired data comprehensively reflects the equipment's operating characteristics. It connects to the edge computing module via an EtherCAT high-speed bus, guaranteeing low latency and high reliability in signal transmission. In the edge computing module, the embedded processor first cleans the acquired signal data, removing noise and outliers. It then extracts Mel-frequency cepstral coefficients from the acoustic wave signal to form an acoustic feature set, extracts sliding window dynamic mean features and trend slope features from the temperature signal to form a temperature feature set, and extracts the equipment rotation frequency component and several time-domain statistical features from the vibration signal to form a vibration feature set. The feature weight adaptive module calculates feature weight coefficients based on the signal-to-noise ratio of each sensor, data volatility, and the proportion of key frequency components. It then weights and fuses the acoustic wave, temperature, and vibration features to generate a multi-dimensional feature vector, which is input into a lightweight fusion model for analysis. The lightweight fusion model captures local frequency domain features through a convolutional neural network and models temporal features through a long short-term memory network, ultimately outputting the fault type and severity, which covers minor wear, moderate wear, and severe wear. When moderate or severe wear is detected, the system triggers an early warning and generates a detailed fault report containing the faulty component, the cause of the fault, and the recommended maintenance cycle. The interface and communication module provides a USB-C debugging interface, an RJ45 network port, and an RS485 serial port to enable signal acquisition, data transmission, and sending shutdown requests of diagnostic results to the rotating equipment controller, ensuring that the equipment can take timely protective measures in abnormal conditions.

[0059] Through the above technical solution, this application achieves comprehensive, real-time monitoring and intelligent fault diagnosis of key components of rotating equipment by integrating multi-source sensing, edge computing, and high-efficiency communication interfaces. The multi-source non-contact sensing module simultaneously collects acoustic signals, surface temperature signals, and displacement vibration signals from key components using acoustic sensors, infrared temperature sensors, and laser vibration sensors, ensuring comprehensive and high-precision data acquisition and capturing minute abnormal features during equipment operation. The edge computing module utilizes an embedded processor to perform real-time data cleaning, feature extraction, and multi-dimensional vector generation on the collected data. It then uses a lightweight model fused with a lightweight convolutional neural network and a long short-term memory network to identify fault types and assess fault severity, achieving lightweight, fast, and accurate diagnostic analysis. When moderate or severe wear is detected, an early warning is triggered immediately, generating a detailed fault report containing the faulty component, the cause of the fault, and the recommended maintenance cycle, providing a scientific basis for maintenance decisions. The interface and communication module transmits data and diagnostic results via USB-C, RJ45, and RS485, supporting the sending of shutdown requests to the rotating equipment controller, ensuring that the equipment can quickly take protective measures in abnormal situations.

[0060] In summary, the collaborative operation of sound sensors, infrared temperature sensors, and laser vibration sensors enables non-contact signal acquisition of key components in rotating equipment, avoiding the problems of complex installation, easy wear and tear, and interference with equipment operation caused by traditional contact sensors. The introduction of multi-source signals allows the diagnostic process to simultaneously perceive acoustic, thermal, and vibration characteristics, comprehensively reflecting the operational health status of components from multiple dimensions. By calculating the confidence level of each sensor signal (including signal-to-noise ratio improvement, temperature fluctuation stability, and vibration spectrum proportion), the weights of sound wave, temperature, and vibration features are dynamically adjusted, achieving adaptive weighted fusion at the feature level. This suppresses the impact of single-sensor channel anomalies or noise interference on diagnostic results and improves the stability of the overall feature representation. The fusion model combines a lightweight convolutional neural network (CNN) and a long short-term memory network (LSTM) dual structure, capable of extracting both the spatial spectral features of sound wave and vibration signals and capturing the temporal evolution features of equipment operation. Through joint optimization and network pruning strategies, the number of model parameters and computational complexity are reduced, enabling the model to run efficiently on edge computing devices and achieving real-time diagnosis. Through a joint optimization mechanism of classification and regression, the system can simultaneously output fault types (such as gear wear, bearing spalling, etc.) and continuous fault severity indicators (minor wear, moderate wear, severe wear). Compared to traditional systems that can only perform binary judgments, this system enables more refined health assessments, providing quantitative references for maintenance personnel. The edge computing module performs local computation and judgment based on an embedded processor, enabling feature extraction and fault identification without uploading data to the cloud, thus shortening response time. Simultaneously, it links with the device controller via EtherCAT bus and RS485 communication interface, automatically triggering shutdown and alarms when moderate or severe wear is detected, ensuring safe equipment operation. When the hardware is first connected, the software automatically completes device identification (reading the device controller's model information), sensor calibration, and model parameter loading. The entire process requires no manual configuration, achieving "plug and connect to work immediately," reducing deployment time from the usual 2 hours to 10 minutes. When the software diagnoses a "serious fault," it automatically sends a shutdown request to the device controller via the hardware's RS485 serial port (if authorized by the customer), and simultaneously triggers the hardware's buzzer alarm, forming a closed loop of "diagnosis-early warning-emergency handling," reducing the risk of fault escalation.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A non-contact fault diagnosis method for key components of rotating equipment, characterized in that, include: Raw signal data is collected to construct a signal dataset, which includes acoustic signals, surface temperature signals, and displacement vibration signals; and the signal dataset is then cleaned. Based on the cleaned signal dataset, Mel frequency cepstral coefficient features are extracted from the acoustic signal to form an acoustic feature set; dynamic mean features within the sliding time window and trend slope features obtained based on temperature change trend analysis are extracted from the surface temperature signal to form a temperature feature set; vibration frequency components of the equipment rotation frequency are extracted from the displacement vibration signal, and several time-domain statistical features are calculated to form a vibration feature set. The acoustic feature set, temperature feature set, and vibration feature set are input into the feature weight adaptive module to generate a multidimensional feature vector. The multidimensional feature vector is input into a lightweight fusion model; the lightweight fusion model is formed by fusing a lightweight convolutional neural network model and a long short-term memory network model; it outputs frequency domain feature representation and temporal feature representation, and performs feature-level fusion to generate fault types and fault degrees; the fault degrees include slight wear, moderate wear and severe wear; When the output fault level is moderate or severe wear, an early warning is triggered, and a fault report is generated. The fault report includes the faulty component, the cause of the fault, and the recommended maintenance cycle. Before inputting the multidimensional feature vector into the lightweight fusion model, the following steps are included: A training dataset is constructed based on historical fault data of key components of rotating equipment. The training data in the training dataset are labeled according to fault type as gearbox gear wear, bearing inner ring spalling, bearing outer ring spalling, and bearing rolling element wear. Each training data is assigned a continuous label value representing the degree of fault. The training dataset is divided into a multi-scale time window sequence, and the multi-dimensional feature vectors in each time window are aligned in the time and frequency domains to make different fault types comparable in the frequency domain feature space. The lightweight fusion model is trained using a joint optimization strategy: the cross-entropy loss function is used to optimize the fault type classification task, the mean squared error loss function is used to optimize the fault degree regression task, and the training process of the two tasks is balanced through a task weight adaptive mechanism. During training, redundant neuron connections are dynamically pruned to preserve network paths that are sensitive to fault characteristics, thereby generating a lightweight fusion model. When the multidimensional feature vector is input into the lightweight fusion model, it includes: The lightweight convolutional neural network model performs channel-wise spatial filtering on the input multidimensional feature vector using deep convolutional layers to capture the local patterns of sound wave spectrum envelope and vibration spectrum energy distribution; it then fuses cross-channel features using point-wise convolutional layers to generate an initial frequency domain feature representation with spatial invariance; finally, it compresses the feature dimension through a global average pooling layer to output the frequency domain feature representation. The Long Short-Term Memory (LSTM) network model captures historical state information by processing the input multidimensional feature vector in chronological order through a forward LSM layer; it captures future state information by processing the input multidimensional feature vector in reverse chronological order through a backward LSM layer; the bidirectional output-input attention mechanism layer characterizes the key time points of fault evolution by calculating the correlation weights between time steps; and it outputs a temporal feature representation through a fully connected layer. When generating fault types and fault severity, the following are included: The frequency domain feature representation and the time series feature representation are fused at the feature level. The fault type is output through the classification layer, and the continuous label value representing the degree of fault is output through the regression layer. When the acoustic feature set, temperature feature set, and vibration feature set are input into the feature weight adaptive module to generate a multidimensional feature vector, the following steps are included: Based on the cleaned signal dataset, the signal-to-noise ratio (SNR) improvement of the acoustic signal is calculated, and the SNR improvement is normalized to the acoustic sensor confidence level. Calculate the standard deviation of the data within the sliding window of the surface temperature signal, and normalize the reciprocal of the standard deviation to obtain the confidence level of the temperature sensor; Calculate the proportion of the amplitude of the frequency component in the displacement vibration signal whose amplitude is greater than the preset baseline level and is within a preset multiple of the device frequency to the total amplitude, and normalize the proportion to the vibration sensor confidence level. Based on the confidence levels of the acoustic sensor, temperature sensor, and vibration sensor, feature weight coefficients are assigned, and the acoustic feature set, temperature feature set, and vibration feature set are weighted and fused; the multidimensional feature vector is generated by feature splicing.

2. The non-contact fault diagnosis method for key components of rotating equipment according to claim 1, characterized in that, Data cleaning of the signal dataset includes: The acoustic signal is subjected to wavelet transform noise reduction processing, the surface temperature signal is subjected to outlier removal, and the displacement vibration signal is subjected to Fourier transform to generate frequency domain data.

3. The non-contact fault diagnosis method for key components of rotating equipment according to claim 2, characterized in that, When extracting Mel-frequency cepstral coefficient features from an acoustic signal to form an acoustic feature set, the following features are included: The acoustic signal is segmented into frames, and the Mel spectrum of each frame is calculated; a logarithmic operation is performed on the Mel spectrum to obtain a logarithmic sequence; a discrete cosine transform is performed on the logarithmic sequence to generate a cepstral coefficient sequence. Among them, the low-order coefficients of the cepstral coefficient sequence represent the shape of the sound spectrum envelope, and the high-order coefficients represent the sound detail features; The low-order coefficients in the cepstral coefficient sequence are extracted as Mel frequency cepstral coefficient features to form the acoustic feature set.

4. The non-contact fault diagnosis method for key components of rotating equipment according to claim 3, characterized in that, When extracting the dynamic mean feature within the sliding time window and the trend slope feature obtained based on temperature change trend analysis from the surface temperature signal to form a temperature feature set, the following are included: Temperature data within a preset time-limited sliding window is collected in real time, and the arithmetic mean of the temperature data is calculated as a dynamic mean feature. The time series curve of the temperature data within the sliding window is fitted using the least squares method, and the slope of the time series curve is calculated as a trend slope feature. The dynamic mean feature and the trend slope feature are combined to form the temperature feature set.

5. The non-contact fault diagnosis method for key components of rotating equipment according to claim 4, characterized in that, When extracting the vibration frequency component of the equipment rotation frequency from the displacement vibration signal and calculating several time-domain statistical features to form a vibration feature set, the following are included: In the frequency domain data, frequency components with amplitudes greater than a preset baseline level and within a preset multiple range of the device's switching frequency are selected, and the amplitudes of the frequency components are extracted as frequency domain features. The peak value, root mean square value, kurtosis, and waveform factor of the displacement vibration signal are calculated as time-domain statistical features. The frequency domain features and time domain statistical features are combined to form the vibration feature set.

6. A non-contact fault diagnosis system for a key component of rotating equipment, applied to the non-contact fault diagnosis method for a key component of rotating equipment as described in any one of claims 1-5, characterized in that, include: The multi-source non-contact sensing module integrates a sound sensor, an infrared temperature sensor, and a laser vibration sensor to synchronously acquire sound wave signals, surface temperature signals, and displacement vibration signals of key components of rotating equipment. An edge computing module, equipped with an embedded processor, is connected to the multi-source non-contact sensing module via the EtherCAT protocol. It receives acoustic signals, surface temperature signals, and displacement vibration signals, constructs a signal dataset, and performs data cleaning. Based on the cleaned signal dataset, it extracts Mel-frequency cepstral coefficient features from the acoustic signals to form an acoustic feature set; it extracts dynamic mean features within a sliding time window and trend slope features obtained from temperature change trend analysis from the surface temperature signals to form a temperature feature set; and it extracts the vibration frequency components of the device's rotational frequency from the displacement vibration signals and calculates several time-domain statistical features to form a vibration feature set. The acoustic feature set, temperature feature set, and vibration feature set are input into the feature weight adaptive module to generate a multi-dimensional feature vector. The multi-dimensional feature vector is then input into a lightweight fusion model. The lightweight fusion model is formed by fusing a lightweight convolutional neural network model and a long short-term memory network model. It outputs frequency domain feature representation and time series feature representation, and performs feature-level fusion to generate fault type and fault degree. The fault degree includes slight wear, moderate wear, and severe wear. When the output fault degree is moderate wear or severe wear, an early warning is triggered, and a fault report is generated. The fault report includes the faulty component, the cause of the fault, and the recommended maintenance cycle. The interface and communication module includes a USB-C debugging interface, an RJ45 network port, and an RS485 serial port, used to transmit the acoustic wave signals, surface temperature signals, and displacement vibration signals of key components of the rotating equipment synchronously to the edge computing module; and to send a shutdown request to the rotating equipment controller via the RS485 serial port with the output fault type and fault degree.

Images