Oilless air compressor fault detection method based on acoustic emission signal and attention mechanism

Abstract

The invention belongs to the technical field of air compressor state monitoring and intelligent operation and maintenance, particularly relates to an oilless air compressor fault detection method based on acoustic emission signals and an attention mechanism, and aims to realize high-sensitivity and high-reliability detection of early faults of an oilless air compressor. Comprising the steps that acoustic emission signals and operation condition information of key parts of the oilless air compressor are collected, and a structured acoustic emission fault data set with fault labels is constructed; the key parts at least comprise an area nearby a bearing seat, an air valve, an area nearby a raft plate, a machine body shell high-stress area and a leakage suspected area. Preprocessing acoustic emission signals in the structured acoustic emission fault data set, and extracting time-frequency features; converting the time-frequency features into a feature sequence, adding position codes, and retaining time sequence and frequency band structure information; and modeling the feature sequence by adopting a Transform network containing a multi-head self-attention mechanism, and outputting a fault identification result.

Description

Technical Field

[0001] This invention relates to the field of air compressor condition monitoring and intelligent operation and maintenance technology, and in particular to a fault detection method for oil-free air compressors based on acoustic emission signals and attention mechanisms. Background Technology

[0002] Oil-free air compressors, capable of providing clean, oil-free compressed air, have become core equipment in industrial settings with stringent requirements for gas cleanliness, such as food processing, pharmaceutical production, and electronics manufacturing. Their operational stability directly impacts the continuity of production processes and product quality. However, because they do not use lubricating oil during operation, traditional fault diagnosis methods relying on oil analysis are completely inapplicable. Furthermore, early faults in oil-free air compressors often manifest as weak, short-duration impacts or friction phenomena, which conventional vibration and temperature monitoring methods struggle to effectively capture. This leads to early faults being easily overlooked, potentially resulting in equipment downtime or even production accidents.

[0003] Acoustic emission (AE) technology, as a dynamic non-destructive testing technique, can accurately capture high-frequency elastic wave signals generated by internal crack initiation, surface impact, and micro-friction during equipment operation. It exhibits unique advantages in early-stage fault detection, providing a new technical approach for condition monitoring of oil-free air compressors. However, in actual industrial scenarios, the operating conditions of oil-free air compressors are complex and variable. Acoustic emission signals are easily affected by load fluctuations, temperature changes, and other operating conditions, and the signals themselves exhibit strong non-stationarity. Traditional feature extraction methods and simple classification models struggle to effectively distinguish fault signals from operating noise, making stable and reliable fault identification impossible.

[0004] While existing technologies have attempted to apply acoustic emission technology to equipment fault detection, they generally suffer from poor adaptability to the unique operating characteristics of oil-free air compressors, insufficient sensitivity to early, subtle fault features, and weak resistance to operational interference. Especially under complex operating conditions, traditional models struggle to effectively model the temporal and frequency band characteristics of acoustic emission signals, resulting in low fault detection accuracy and insufficient robustness, failing to meet the practical needs of industrial sites for accurate early warning of faults in oil-free air compressors. Summary of the Invention

[0005] The purpose of this invention is to provide a fault detection method for oil-free air compressors based on acoustic emission signals and attention mechanisms, aiming to achieve highly sensitive and reliable detection of early faults in oil-free air compressors.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a fault detection method for oil-free air compressors based on acoustic emission signals and an attention mechanism, comprising: S1, collecting acoustic emission signals and operating condition information from key parts of the oil-free air compressor to construct a structured acoustic emission fault dataset with fault labels; wherein, the key parts include at least the area near the bearing housing, the area near the valve and the raft plate, the high-stress area of ​​the casing, and the suspected leakage area; S2, preprocessing the acoustic emission signals in the structured acoustic emission fault dataset to extract time-frequency features; S3, converting the time-frequency features into feature sequences and adding positional encoding, while retaining the temporal and frequency band structure information; S4, using a Transformer network with a multi-head self-attention mechanism to model the feature sequences and output the fault identification results.

[0007] In step S1, sensors are deployed at key locations to collect raw acoustic emission signals; the raw signals are segmented according to fixed time windows or fixed revolution windows, and equipment information and operating condition information are recorded synchronously; based on maintenance records and expert experience, the segmented signals are labeled and structured to form a structured acoustic emission fault dataset.

[0008] Operating condition information includes at least one or more of the following: load rate, exhaust pressure, exhaust temperature, intake temperature, ambient temperature, start / stop status, and control mode parameters. Equipment information includes model type, rated power, rated exhaust volume, rated speed, number of stages, cooling method, and pipeline pressure range. Tags include basic information, status information, and severity information. Basic information includes at least: normal, abnormal valve plate impact, minor bearing damage, minor leakage, rubbing, and performance degradation. Status information indicates the time of the fault occurrence; severity information indicates the severity of the fault.

[0009] The preprocessing in step S2 includes: denoising the acoustic emission signal using one or more of bandpass filtering, spectral subtraction, or wavelet thresholding; performing amplitude normalization or standardization on the denoised signal; and removing or interpolating abnormal spikes and lost packets in the signal after amplitude normalization or standardization.

[0010] Time-frequency feature extraction employs one or more of short-time Fourier transform, wavelet transform, or empirical mode decomposition to obtain time-spectrum maps or multi-scale frequency band energy features.

[0011] In step S3, the time-frequency features are segmented into feature subsequences according to the time dimension; the feature subsequences are divided into feature blocks or the multi-scale frequency band energy features are grouped and mapped into a unified dimension vector; position encoding is added to the vector to form a feature sequence that meets the input requirements of the Transformer network.

[0012] In step S3, the correlation between elements in the feature sequence is analyzed through a multi-head self-attention mechanism, and weights are assigned to features of different time segments and frequency bands. The corresponding features in the feature sequence are weighted and fused according to the assigned weights to highlight fault-related features. After nonlinear transformation and dimensional adjustment, the weighted and fused features are processed by residual connection and normalization. The processed feature sequence is globally modeled to output a high-level fault feature representation.

[0013] The multi-head self-attention mechanism divides attention heads into low-frequency, mid-frequency, and high-frequency bands, assigns priority weights to high-frequency band attention heads or optimizes attention calculation logic, and strengthens high-frequency characteristic responses to capture high-frequency impact characteristics generated by faults.

[0014] The fault identification results in step S4 include fault type probability or operational anomaly score. When the result exceeds a preset threshold, an early fault warning is triggered.

[0015] Early failures include one or more of the following: abnormal valve plate impact, minor bearing damage, slight gas leakage, rubbing of moving parts, and performance degradation.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application provides a fault detection method for oil-free air compressors based on acoustic emission signals and attention mechanisms. It employs acoustic emission technology, placing sensors near fault-sensitive key locations such as bearing housings and valves / valve plates to directly capture high-frequency elastic wave signals generated by faults, ensuring the integrity of early fault information from the signal source. Simultaneously, it combines bandpass filtering, wavelet threshold denoising, and other preprocessing methods to eliminate operating noise and signal interference, significantly improving the signal-to-noise ratio of weak fault signals. This solves the problem of traditional technologies failing to detect early faults, accurately identifying early fault types such as abnormal valve plate impact and minor bearing damage, thus providing sufficient time for equipment maintenance.

[0017] 2. This application provides a fault detection method for oil-free air compressors based on acoustic emission signals and an attention mechanism. In the feature processing stage, various time-frequency feature extraction methods, such as short-time Fourier transform and wavelet transform, are used to comprehensively mine the time-domain and frequency-domain correlation information of the acoustic emission signal. After time-dimensional segmentation, feature block mapping, and position encoding, the temporal relationship and frequency band structure of the signal are fully preserved, avoiding the information loss problem in traditional feature extraction. At the modeling level, a Transformer network with a multi-head self-attention mechanism analyzes the correlation of feature sequence elements and dynamically assigns weights to features in different time segments and frequency bands. In particular, by dividing the attention heads according to low, medium, and high frequency bands and strengthening the high-frequency response, it can accurately focus on the high-frequency impact features generated by the fault, effectively suppressing noise interference caused by operating condition fluctuations. Simultaneously, the nonlinear transformation, residual connection, and normalization processing after weighted fusion further enhance the identification of fault features, significantly improving the model's sensitivity to weak faults. The detection accuracy and robustness far exceed traditional methods. Attached Figure Description

[0018] Figure 1 This is a flowchart of a fault detection method for an oil-free air compressor based on acoustic emission signals and attention mechanisms, provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 This application provides a fault detection method for oil-free air compressors based on acoustic emission signals and attention mechanisms, including: S1. Collect acoustic emission signals and operating condition information from key components of the oil-free air compressor to construct a structured acoustic emission fault dataset with fault labels. Key components include at least the area near the bearing housing, the area near the valves and raft plate, the high-stress area of ​​the compressor casing, and areas suspected of leakage.

[0021] For example, in step S1, sensors are arranged at key locations to collect raw acoustic emission signals; the raw signals are segmented according to fixed time windows or fixed revolution windows, and operating condition information is recorded synchronously; the segmented signals are labeled based on maintenance records and expert experience, and structured storage is used to form a structured acoustic emission fault dataset.

[0022] The operating condition information includes at least one or more of the following: load rate, exhaust pressure, exhaust temperature, intake temperature, ambient temperature, start / stop status, and control mode parameters. Equipment information includes model type (oil-free screw / oil-free scroll / oil-free reciprocating), rated power, rated displacement, rated speed, number of stages, cooling method, and pipeline pressure range. Tags include basic information, status information, and severity information. Basic information includes at least: normal, abnormal valve plate impact, minor bearing damage, minor leakage, rubbing, and performance degradation; status information distinguishes between early, mid, and late-stage faults; severity information distinguishes between minor, moderate, and severe faults.

[0023] Specifically, the continuously sampled acoustic emission signals are denoted as... The signal is segmented using a sliding window of length L and a step size of S. The k-th segment is defined as follows: Where k = 1, 2, ..., K, and K is the number of segments. Segmenting the original signal according to a fixed time window or a fixed revolution window is used to convert the acoustic emission signal during operation into a set of processable segments.

[0024] S2. Preprocess the acoustic emission signals in the structured acoustic emission fault dataset and extract time-frequency features.

[0025] As one possible implementation, the preprocessing in step S2 includes: denoising the acoustic emission signal using one or more of bandpass filtering, spectral subtraction, or wavelet thresholding; performing amplitude normalization or standardization on the denoised signal; and removing or interpolating abnormal spikes and lost packets in the signal after amplitude normalization or standardization.

[0026] For example, each acoustic emission signal is normalized to ensure consistent amplitude scale under different operating conditions, such as by using zero-mean unit variance standardization: in, and and are the mean and standard deviation of the k-th segment of the signal, respectively.

[0027] The preprocessed acoustic emission signal can be denoted as: Where D(·) represents the denoising operator, which can be implemented by bandpass filtering, wavelet thresholding, or spectral subtraction.

[0028] Perform time-frequency transformation on the preprocessed acoustic emission signal to obtain time-frequency features. As a possible implementation method, time-frequency feature extraction can be performed using one or more of the following: Short Time Fourier Transform (STFT), Wavelet Transform (CWT), or Empirical Mode Decomposition (EMD) to obtain the time spectrum or multi-scale frequency band energy features of the acoustic emission signal.

[0029] For example, the time-frequency representation obtained by using the short-time Fourier transform is as follows: in (·) is a window function. For time indexing, For frequency indexing. | |As a time-frequency amplitude / energy characteristic, it is also known as a time-frequency characteristic.

[0030] S3. Convert the time-frequency features into feature sequences and add position coding to preserve the time sequence and frequency band structure information.

[0031] As one possible implementation, in step S3, the time-frequency features are segmented along the time dimension to form a feature subsequence arranged in chronological order. The feature subsequences are then divided into feature blocks, or the multi-scale frequency band energy features are grouped and mapped to a unified-dimensional vector. In some embodiments, the obtained time-frequency features are time-spectrum maps of acoustic emission signals. In these embodiments, the time-spectrum maps are divided into several feature blocks of a fixed size and mapped to a unified-dimensional vector. Position encoding is added to the vectors to form a feature sequence that meets the input requirements of the Transformer network. In other embodiments, the obtained time-frequency features are multi-scale frequency band energy features. In these embodiments, the multi-scale frequency band energy features are grouped along the frequency band dimension and mapped to a unified-dimensional vector. Position encoding is added to the vectors to form a feature sequence that meets the input requirements of the Transformer network.

[0032] For example, time-frequency features Divide into N feature subsequences arranged in chronological order: Mapped to a vector of uniform dimension: Add positional encoding to the vector to form a feature sequence that meets the input requirements of the Transformer network: in, It is a positional encoding vector used to preserve sequence order and structural information.

[0033] As one possible implementation, when using multiple acoustic emission sensors for signal acquisition, the feature vectors corresponding to different sensors are spliced ​​or weighted and fused in the channel dimension to form a multi-channel acoustic emission feature sequence. The constructed acoustic emission feature sequence is then length-aligned or padded to meet the input format requirements of the subsequent attention-based Transformer network.

[0034] The raw signals collected by each sensor are preprocessed, extracted in the time domain, and mapped, transforming them into vectors of the same dimension. This ensures that the features from different sensors are comparable and can be integrated. The "channel dimension" is also the sensor dimension; each sensor corresponds to a "data channel," and multi-channel features are the set of feature vectors output by each sensor. Concatenating the feature vectors along the channel dimension involves joining the feature vectors from each sensor end-to-end in channel order, resulting in a feature vector with strong complementarity among the signals collected by the sensors. Weighted fusion assigns different weights to the feature vectors of different sensors and then performs a weighted sum. Weighted fusion uses the sensor deployment location to strategically set weights, highlighting fault signals at critical locations while reducing noise interference from sensors in irrelevant locations.

[0035] As one possible implementation, the Transformer network includes a 4-layer encoder structure, which includes at least a multi-head self-attention module, a feedforward network module, and a residual connection and normalization module.

[0036] For example, in step S3, the correlation between elements in the feature sequence is analyzed by a multi-head self-attention mechanism, and weights are assigned to features of different time segments and frequency bands; the corresponding features in the feature sequence are weighted and fused according to the assigned weights to highlight fault-related features; after nonlinear transformation and dimensional adjustment of the weighted and fused features, residual connection processing and normalization processing are performed; the processed feature sequence is globally modeled to output a high-level fault feature representation.

[0037] The multi-head self-attention module analyzes the correlation of elements in the feature sequence and assigns weights to features in different time segments and frequency bands to improve sensitivity to early, weak fault features. For example, the multi-head self-attention mechanism divides attention heads into low-frequency, mid-frequency, and high-frequency bands, assigns priority weights to high-frequency band attention heads or optimizes attention calculation logic to enhance the response of high-frequency features and capture the high-frequency impact features generated by faults.

[0038] The multi-head self-attention module performs weighted fusion of corresponding features in the feature sequence according to the assigned weights. To avoid overly restricting the specific implementation, the weighted fusion process is represented in an abstract form in this embodiment: For sequence elements Calculate weights And form a weighted feature representation: in, This represents the attention weight coefficient related to the fault. Through the above weighted fusion, key time-frequency segment features related to early faults can be highlighted, while non-critical features caused by changes in operating conditions can be suppressed. The weight coefficients satisfy non-negativity and normalization constraints, and are used to represent the degree of contribution of each time-frequency segment to fault identification.

[0039] S4. A Transformer network with a multi-head self-attention mechanism is used to model the feature sequence and output the fault identification results.

[0040] The fault identification results in step S4 include fault type probability or operational anomaly score. When the result exceeds a preset threshold, an early fault warning is triggered.

[0041] As one possible implementation, a classification header or an anomaly scoring header can be set at the output of the Transformer network. The classification header is used to output the probability distribution of fault types, and the anomaly scoring header is used to output the anomaly score or fault severity.

[0042] Examples of early failures include one or more of the following: abnormal valve plate impact, micro-damage to bearings, minor gas leaks, rubbing of moving parts, and performance degradation.

[0043] In step S4, the model needs to be trained and validated. For example, the structured acoustic emission fault dataset is divided into a training set and a test set, and supervised training is used for training. The model performance is evaluated by metrics such as accuracy, recall, F1 score or AUC. Optionally, different operating conditions (different loads / pressures / temperatures) can be evaluated separately to verify the robustness of the model.

[0044] The trained model is deployed on an edge computing device or industrial control computer. The real-time acquired acoustic emission signals are processed sequentially according to steps S2 and S3 and then input into the model for inference. An alarm is triggered when the fault type probability or anomaly score exceeds a threshold. The alarm level can be determined jointly by the fault probability, anomaly score, and operating condition risk. The results are output in conjunction with operation and maintenance, outputting the fault type, confidence level, possible fault location, and suggested inspection items. The alarm events are associated with and stored with equipment operating parameters and acoustic emission segment indexes to support subsequent traceability analysis and closed-loop maintenance.

[0045] The overall process of the method provided in this application will be described below using a specific embodiment: An oil-free air compressor is used as the monitored object. The oil-free air compressor can be any one of an oil-free screw air compressor, an oil-free scroll air compressor, or an oil-free reciprocating air compressor.

[0046] First, basic information about the oil-free air compressor (model type, rated power, rated discharge volume, etc.) and operating condition information (load rate, discharge pressure, discharge temperature, etc.) are stored, and raw acoustic emission waveform signals during equipment operation are collected simultaneously. Next, the raw acoustic emission signals are segmented using a fixed time window or a fixed speed window to obtain a set of acoustic emission signal segments. Based on historical maintenance records, expert experience, and operational events, the acoustic emission signal segments are labeled. The labels include basic information, status information, and severity information. Basic information includes at least: normal, abnormal valve plate impact, minor bearing damage, minor leakage, rubbing, and performance degradation; status information indicates the period of fault occurrence (early, middle, late); severity information indicates the severity of the fault. The signal segments, operating condition information, equipment information, and label information are uniformly formatted to generate a structured acoustic emission fault dataset.

[0047] The acoustic emission signal segments are sequentially denoised, normalized, and subjected to the removal or interpolation of abnormal spikes and lost segments. One or more of the following methods—Short Time Fourier Transform (STFT), Wavelet Transform (CWT), or Empirical Mode Decomposition (EMD)—are used to extract the time-frequency features of the acoustic emission signal. These time-frequency features are then segmented along the time dimension to form a time-ordered feature subsequence. The time-spectrum is divided into feature blocks or grouped by frequency band energy features and mapped to a unified-dimensional feature vector. The feature vectors from multiple channels are concatenated or weighted and fused. After nonlinear transformation and dimensional adjustment (including length alignment and padding of the feature sequence), residual concatenation and normalization are performed. A positional encoding vector is added to meet the input requirements of the Transformer network.

[0048] The processed feature sequence is modeled globally. The positional encoding and global feature modeling adopt "time-frequency joint positional encoding". The time dimension retains the sine-cosine encoding to maintain the temporal relationship, and the frequency dimension introduces a learnable encoding vector to highlight the difference in fault contribution in different frequency bands. The feature sequence is modeled globally using a Transformer network to output a high-level feature representation that focuses on early faults.

[0049] The encoder structure, which constructs a Transformer network adapted to acoustic emission signals, verifies the feasibility of a 4-layer encoder stacking design through a Pareto experiment of "computing power-accuracy," reducing the number of model parameters and inference latency, and adapting to the deployment requirements of industrial edge devices. Each layer of the encoder structure consists of an improved multi-head self-attention module (MHSA), an adaptive feedforward network module (FFN), and a residual connection and normalization module.

[0050] An improved multi-head self-attention module design. Unlike the traditional Transformer network that distributes attention equally, it adds "frequency band-time dual-dimensional attention": attention heads are divided into low, medium and high frequency bands to specifically capture high-frequency impact features of faults; a fault-sensitive weight factor is introduced to automatically increase the weight of fault-related time segments, suppress operating noise interference, and solve the problem that traditional models are not sensitive to weak fault features.

[0051] Adaptive feedforward network and operating condition fusion optimization. The feedforward network dynamically adjusts the hidden layer parameters through adaptive gating units to adapt to the feature dimension changes under different operating conditions; an operating condition feature fusion interface is added to the second layer of the encoder structure to map the standardized operating condition parameters into a consistent dimension vector, which is then fused with the acoustic emission features in a weighted manner to explicitly model the impact of operating condition fluctuations, thus making up for the shortcomings of traditional Transformer networks in dealing with changes in industrial operating conditions.

[0052] Set a classification header (outputs the probability distribution of fault types) and an anomaly scoring header (outputs the severity of faults), and generate early fault alarm judgments based on the output results.

[0053] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0054] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fault detection method for oil-free air compressors based on acoustic emission signals and attention mechanisms, characterized in that, include: S1. Collect acoustic emission signals and operating condition information from key components of the oil-free air compressor to construct a structured acoustic emission fault dataset with fault labels. Key components include at least the area near the bearing housing, the area near the valve and raft plate, the high-stress area of ​​the casing, and the suspected leakage area. S2. Preprocess the acoustic emission signals in the structured acoustic emission fault dataset to extract time-frequency features. S3. Convert the time-frequency features into feature sequences and add positional encoding, preserving temporal and frequency band structure information. S4. Use a Transformer network with a multi-head self-attention mechanism to model the feature sequences and output fault identification results.

2. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 1, characterized in that: In step S1, sensors are deployed at key locations to collect raw acoustic emission signals; the raw signals are segmented according to fixed time windows or fixed revolution windows, and equipment information and operating condition information are recorded synchronously; based on maintenance records and expert experience, the segmented signals are labeled and structured to form a structured acoustic emission fault dataset.

3. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 2, characterized in that: The operating condition information includes at least one or more of the following: load rate, exhaust pressure, exhaust temperature, intake temperature, ambient temperature, start / stop status, and control mode parameters; equipment information includes model type, rated power, rated exhaust volume, rated speed, number of stages, cooling method, and pipeline pressure range; tags include basic information, status information, and severity information. Basic information includes at least: normal, abnormal valve plate impact, minor bearing damage, minor leakage, rubbing, and performance degradation; status information indicates the period when the fault occurred; severity information indicates the severity of the fault.

4. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 1, characterized in that, The preprocessing in step S2 includes: denoising the acoustic emission signal using one or more of bandpass filtering, spectral subtraction, or wavelet thresholding; performing amplitude normalization or standardization on the denoised signal; and removing or interpolating abnormal spikes and lost packets in the signal after amplitude normalization or standardization.

5. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 1 or 4, characterized in that, The time-frequency feature extraction employs one or more of short-time Fourier transform, wavelet transform, or empirical mode decomposition to obtain time-spectrum maps or multi-scale frequency band energy features.

6. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 1, characterized in that, In step S3, the time-frequency features are segmented according to the time dimension to form feature subsequences arranged in time order; the feature subsequences are divided into feature blocks or multi-scale frequency band energy features are grouped and mapped into a unified dimension vector; position encoding is added to the vector to form a feature sequence that meets the input requirements of the Transformer network.

7. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 1, characterized in that, In step S3, the correlation between elements in the feature sequence is analyzed through a multi-head self-attention mechanism, and weights are assigned to features of different time segments and frequency bands. The corresponding features in the feature sequence are weighted and fused according to the assigned weights to highlight fault-related features. After nonlinear transformation and dimensional adjustment, the weighted and fused features are processed by residual connection and normalization. The processed feature sequence is globally modeled to output a high-level fault feature representation.

8. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 7, characterized in that, The multi-head self-attention mechanism divides attention heads into low-frequency, mid-frequency, and high-frequency bands, assigns priority weights to high-frequency band attention heads or optimizes attention calculation logic, and strengthens high-frequency characteristic responses to capture high-frequency impact characteristics generated by faults.

9. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 1, characterized in that, The fault identification results in step S4 include fault type probability or operational anomaly score. When the result exceeds a preset threshold, an early fault warning is triggered.

10. The oil-free air compressor fault detection method based on acoustic emission signals and attention mechanisms according to claim 9, characterized in that, The early failures include one or more of the following: abnormal valve plate impact, minor bearing damage, slight gas leakage, rubbing of moving parts, and performance degradation.

Images