Rotating machine lightweight fault diagnosis method and system capable of being deployed on edge device in real time

Abstract

The invention relates to a rotating machine lightweight fault diagnosis method and system capable of being deployed on edge equipment in real time, and the method comprises the steps: collecting a vibration acceleration original time sequence signal of a rotating machine, storing the vibration acceleration original time sequence signal to an edge end, carrying out the sliding window segmentation and preprocessing, and constructing a vibration fragment data set; building a lightweight fault diagnosis model with a built-in adaptive residual enhancement module and a multi-group coordinate attention module at the edge end; training a lightweight fault diagnosis model by using the vibration fragment data set; performing fault diagnosis by using the trained lightweight fault diagnosis model to obtain a diagnosis result; and performing fault level judgment according to a diagnosis result, and performing real-time feedback on the fault by the edge end equipment. According to the invention, the fault diagnosis model can be deployed at the edge terminal for local operation in the form of low storage, low computing power and low time delay without depending on a high-performance server or cloud reasoning, the problem that the existing method is difficult to apply in real time at the industrial field equipment end is solved, and the light-weight fault diagnosis of the rotating machinery is realized.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis and intelligent manufacturing technology, and in particular to a lightweight fault diagnosis method and system for rotating machinery that can be deployed in real time on edge devices. Background Technology

[0002] With the continuous development of industrial automation and intelligent manufacturing, rotating machinery is being used more and more widely in various industrial production processes. The health of key components in rotating machinery, such as wind turbines, aircraft engines, pumps, and compressors, directly affects the safe operation and production efficiency of the equipment. During operation, rotating machinery often experiences component failures due to bearing wear, gear damage, and other malfunctions. If these failures are not detected in time, they can lead to equipment downtime, production accidents, and even safety hazards.

[0003] Currently, fault diagnosis in rotating machinery primarily relies on signals collected by sensors, such as vibration, temperature, and pressure. These signals are then processed and analyzed to extract potential fault characteristics. However, traditional fault diagnosis methods face the challenge of high computational complexity and difficulty in real-time processing. These methods typically depend on high-performance computing platforms for complex signal processing and feature extraction, making deployment on resource-constrained edge devices extremely difficult. Edge computing, as an emerging technology, can effectively reduce data transmission latency by moving data processing tasks closer to the data source, but it also faces the challenge of limited computing resources.

[0004] To achieve real-time fault diagnosis on edge devices, lightweight model design is crucial. Existing high-precision models are objectively unable to run directly on resource-constrained edge devices (millions of FLOPs) due to their high computational complexity (quadratic level) and large parameter redundancy.

[0005] Furthermore, with the advancement of intelligent manufacturing, real-time monitoring of equipment health status and fault prediction have become key tasks in production management. Therefore, how to efficiently diagnose faults on edge devices and address noise interference in the field environment has become an urgent problem to be solved in the context of intelligent manufacturing and Industry 4.0. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a lightweight fault diagnosis method and system for rotating machinery that can be deployed in real time on edge devices. It overcomes the limitations of traditional fault diagnosis methods, such as high computational complexity and difficulty in real-time processing. The invention employs an adaptive residual enhancement module and a multi-group coordinate attention module to build a lightweight fault diagnosis model. It preprocesses multi-source vibration signal data collected by sensors as input to extract efficient fault features.

[0007] To achieve the above technical objectives, the present invention provides the following technical solution: a lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices, comprising the following steps: Vibration sensors are installed on key parts of the rotating machinery to be monitored to collect the raw time-series signals of vibration acceleration during operation and store them at the edge. The original time-series vibration acceleration signal is segmented and preprocessed at the edge to obtain the vibration segment tensor, and then combined with the corresponding real fault category label to construct the vibration segment dataset. Sliding window segments are divided according to fixed window length and step size to obtain multiple vibration segment signals. All vibration segment signals are organized into tensors according to channels and amplitude normalization and noise reduction are performed at the edge. A lightweight fault diagnosis model is built at the edge; the lightweight fault diagnosis model includes an adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier; The vibration segment dataset is input into the adaptive residual enhancement module to extract temporal fault features, and then the temporal fault features are sent to the multi-group coordinate attention module to obtain weighted temporal fault features; the adaptive residual enhancement module includes parallel depthwise separable convolutional branches and inverse residual branches, as well as gated fusion units; The weighted temporal fault features are fed into a lightweight classifier to determine the fault category and obtain the classification prediction probability. A lightweight fault diagnosis model is trained by calculating the total loss based on the classification prediction probability; the total loss includes classification loss and reconstruction loss. Vibration signals are collected in real time at the edge, and a trained lightweight fault diagnosis model is used to perform fault diagnosis and obtain the diagnosis results; the fault level is determined based on the diagnosis results. Edge devices transmit fault levels to the host computer via audible and visual alarms and information push notifications, or record fault data locally to provide real-time feedback, enabling lightweight fault diagnosis of rotating machinery.

[0008] Optionally, the original vibration acceleration time-series signal is an acceleration amplitude sequence acquired at the sampling frequency; The original vibration acceleration timing signal is transmitted in real time to the local cache of the edge diagnostic unit and stored therein, without being uploaded to a remote server. The edge diagnostic unit is a resource-constrained terminal that can be deployed in an industrial field and has a million-level floating-point computing power, which can support the execution of subsequent processing steps of the present invention.

[0009] Optionally, the step of performing sliding window segmentation and preprocessing on the original time-series vibration acceleration signal at the edge to obtain the vibration segment tensor includes: The sliding window is segmented according to a fixed window length and step size to obtain multiple vibration segment signals; All vibration segment signals are organized into tensors according to channels, and amplitude normalization and noise reduction are performed at the edges to obtain vibration segment tensors.

[0010] Optionally, the step of inputting the vibration segment dataset into the adaptive residual enhancement module to extract time-series fault features includes: The vibration segment tensors in the vibration segment dataset are input into the depthwise separable convolution branch and the inverse residual branch, respectively. The depthwise separable convolution branch performs depthwise separable convolution on the vibration segment signal to generate depthwise separable convolution features. The inverse residual branch is an inverse residual structure that extracts inverse residual features from the vibration segment signal. The depthwise separable convolutional features and the inverse residual features are fed into the gated fusion unit. After global average pooling, they are concatenated at the vector level, and then subjected to learnable linear weighting and sigmoid activation to obtain the gated weight vector. Then, based on the gated weight vector, the depthwise separable convolutional features and the inverse residual features are fused channel by channel to obtain the temporal fault features.

[0011] Optionally, the step of feeding the temporal fault features into the multi-group coordinate attention module to obtain weighted temporal fault features includes: The time-series fault features are divided into multiple sub-group features according to the channel dimension. The features of each sub-group are averaged and fused along the channel dimension at each time step to obtain the time response vector of each sub-group feature. Lightweight convolutions are applied to the temporal response vectors of each subgroup feature to generate temporal attention weights for each subgroup feature; The features of each subgroup are weighted at each time step according to the temporal attention weight to obtain the weighted subgroup features; By concatenating all weighted subgroup features along the channel dimension, weighted temporal fault features are obtained.

[0012] Optionally, the step of inputting the weighted temporal fault features into a lightweight classifier for fault category determination and obtaining the classification prediction probability includes: The weighted temporal fault features are averaged and compressed over the time dimension, and then passed through a dropout layer, a linear classification layer, and a softmax activation layer to obtain the classification prediction probability.

[0013] Optionally, the classification loss is the cross-entropy loss calculated between the classification prediction probability output by the lightweight fault diagnosis model and the true fault category label.

[0014] Optionally, the edge end acquires vibration signals in real time, uses a trained lightweight fault diagnosis model to perform fault diagnosis, and obtains diagnostic results, including: The collected vibration signals are normalized, denoised, and preprocessed. The preprocessed vibration signal is fed into the trained lightweight fault diagnosis model for inference to obtain the classification prediction probability of each vibration signal. Based on the classification prediction probability, the category with the highest probability is selected as the final diagnostic result for the vibration signal.

[0015] This invention also provides a lightweight fault diagnosis system for rotating machinery that can be deployed in real time on edge devices, for applying the aforementioned lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices, including: The data acquisition and edge preprocessing module is used to acquire the raw time-series signal of vibration acceleration during the operation of rotating machinery, cache it to the edge, perform sliding window segmentation, normalize and preprocess it to construct a vibration segment dataset. The lightweight fault diagnosis model module is used to build and train a lightweight fault diagnosis model, and to use the trained lightweight fault diagnosis model for inference to obtain diagnostic results. The lightweight fault diagnosis model includes an adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier. The edge warning module is used to determine the fault level based on the diagnostic results, enabling edge devices to transmit the fault level to the host computer through audible and visual alarms and information push, or to record fault data locally, providing real-time feedback and realizing lightweight fault diagnosis of rotating machinery. If the edge warning module diagnoses a serious fault, the system will trigger a high-priority alarm; if the diagnosis result is an early fault, it will trigger a medium-priority alarm and provide a warning; if the diagnosis result is normal, no alarm will be triggered.

[0016] The present invention also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the described method for lightweight fault diagnosis of rotating machinery that can be deployed in real time on an edge device.

[0017] By employing the above technical solution, the present invention provides a lightweight fault diagnosis method and system for rotating machinery that can be deployed in real time on edge devices, and has at least the following beneficial effects: (1) By deploying the fault diagnosis task directly on the edge computing device, the present invention significantly reduces the data transmission latency and avoids the latency problem caused by data uploading to the remote server in the traditional centralized computing method. This enables fault diagnosis to be carried out in a real-time environment, ensuring the immediate fault warning and response of rotating machinery under high load and complex working conditions, and greatly improving the safety and stability of the production process. (2) In view of the fact that existing high-precision models are often large in size and cannot be directly implanted, resulting in limited storage and computing power of edge hardware, this invention constructs an adaptive residual enhancement module, abandons the dependence of traditional deep networks on a large number of parameters, and adopts a combination of depth-separable convolution and inverse residual structure, which solves the physical problem that complex diagnostic models cannot be loaded under extremely low hardware configuration, so that low-cost terminals with millions of floating-point computing power can also run high-precision diagnostics. (3) In view of the problem that traditional lightweight models sacrifice noise resistance due to structural simplification in industrial sites under strong noise and variable working conditions, resulting in failure under complex electromagnetic interference and mechanical vibration, this invention designs a multi-group coordinate attention module. Under the premise of limited computing power, it maximizes the use of limited computing resources to lock the key fault frequency band, thereby overcoming environmental noise interference locally without uploading data to the cloud for cleaning, and achieving accurate diagnosis. (4) The present invention adopts a lightweight classifier structure. The lightweight classifier simplifies the network hierarchy, removes redundant calculations, and reduces the memory occupation and computational burden of the model. Compared with the traditional classifier, it adopts a more efficient activation function and fewer parameters, thereby realizing low-latency real-time inference. At the same time, the multi-group coordinate attention module uses the group attention mechanism to enhance the feature discrimination ability with linear computational complexity, while reducing the computational complexity of the traditional self-attention network. This makes the model more suitable for edge deployment, so that the entire fault diagnosis process can run with low power consumption and high efficiency on edge devices. Compared with the traditional deep learning model, the above lightweight design significantly reduces computational complexity and storage requirements, ensuring that the rotating machinery fault diagnosis system can realize real-time inference on edge devices with limited computing resources. (5) For shutdown protection scenarios with extremely high real-time requirements in the fault diagnosis process of rotating machinery, the present invention has a targeted lightweight design, which eliminates the uncertainty of network latency caused by data uploading to the cloud, and ensures that reasoning and feedback are completed in milliseconds, thus meeting the real-time early warning needs of rotating machinery under sudden faults. (6) This invention effectively extracts key features of rotating machinery faults through the depth-separable convolution branch and the inverse residual branch in the adaptive residual enhancement module, as well as the group attention mechanism in the multi-group coordinate attention module, and maintains high diagnostic accuracy under complex working conditions and noise interference. Specifically, the depth-separable convolution branch captures rapidly changing local details and high-frequency impact components in the vibration signal, and the inverse residual branch extracts global context information and low-frequency periodic components through channel expansion and compression. The gated fusion unit performs global average pooling and linear mapping on the outputs of the above two branches, so that the fault diagnosis process can dynamically adjust the attention to high-frequency or low-frequency features according to the spectral characteristics of the current vibration signal without manual intervention, realizing adaptive diagnosis under complex working conditions. At the same time, it effectively suppresses the influence of irrelevant signals such as background noise and power frequency interference, improves the robustness and stability of fault diagnosis, accurately predicts fault type and health status, and can still ensure high prediction accuracy even when the data is incomplete. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is an overall flowchart of a lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices according to the present invention. Figure 2 The curve showing the change of average training loss with training rounds during the training process of the lightweight fault diagnosis model built in this invention. Figure 3 The curve shows the change in average training accuracy with training rounds during the training process of the lightweight fault diagnosis model built in this invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding and implementation of how the present application uses technical means to solve technical problems and achieve technical effects.

[0020] Those skilled in the art will understand that all or part of the steps in the implementation of the methods of the embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0021] Please refer to Figures 1-3 This illustration shows a specific implementation of the present embodiment. This embodiment collects the raw time-series vibration acceleration signals of rotating machinery, stores them at the edge terminal, and performs sliding window segmentation and preprocessing to construct a vibration segment dataset. A lightweight fault diagnosis model with a built-in adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier is built at the edge terminal. The lightweight fault diagnosis model is trained using the vibration segment dataset. Fault diagnosis is performed using the trained lightweight fault diagnosis model to obtain diagnostic results. The fault level is determined based on the diagnostic results, and the edge terminal device provides real-time feedback on the fault. While ensuring diagnostic accuracy, this significantly reduces the number of parameters and computational load, enabling the fault diagnosis model to be deployed and run locally on the edge terminal with low storage, low computing power, and low latency, without relying on high-performance servers or cloud inference. This overcomes the problem of existing methods being difficult to apply in real-time at industrial field equipment, achieving lightweight fault diagnosis for rotating machinery.

[0022] In this embodiment, the experiment uses the PT500Mini mechanical fault simulation platform to simulate the faults of rotating machinery and collect vibration signals. The experimental equipment is a rotating machinery testing system on the PT500Mini platform, equipped with a three-axis accelerometer and a three-component force sensor. The accelerometer is used to collect vibration signals from key parts of the rotating machinery, and the force sensor is used to measure the forces acting on the machinery at different speeds. The sampling frequency of all collected signals is 48 kHz. Stainless steel (HRC-52) is used as the cutting material in the experiment, and tests are conducted under various operating conditions.

[0023] During the experiment, different fault states were simulated, including normal state, inner ring fault, outer ring fault, and rolling element fault. A large amount of vibration signal data was collected under each fault state for subsequent fault diagnosis and prediction algorithm verification. The experimental equipment parameters and sensor configurations, vibration signal acquisition settings under different fault states, and sample size at each rotational speed are shown in Tables 1, 2, and 3. These settings ensured that signal acquisition at different rotational speeds could comprehensively reflect the working state of the rotating machinery under various operating conditions.

[0024] Table 1. Experimental equipment parameters and sensor configuration parameter numerical values Equipment Model PT500Mini Sensor type Triaxial accelerometer, three-component force sensor Sampling frequency 48 kHz Test materials Stainless steel (HRC-52) Measurement location Key components such as bearings and gearboxes Measurement signal Vibration acceleration signal, force signal Speed ​​range 500 r / min to 1500 r / min Table 2. Vibration signal acquisition settings under different fault conditions Fault status Signal acquisition time Rotational speed (r / min) Number of signals acquired (units) Sampling frequency Normal state 10 seconds 500, 1000, 1500 200 48 kHz Inner ring fault 10 seconds 500, 1000, 1500 200 48 kHz Outer ring fault 10 seconds 500, 1000, 1500 200 48 kHz Rolling element failure 10 seconds 500, 1000, 1500 200 48 kHz Table 3. Number of samples at each rotational speed Rotational speed (r / min) Normal state Inner ring fault Outer ring fault Rolling element failure 500 200 200 200 200 1000 200 200 200 200 1500 200 200 200 200 Based on the above experimental equipment, processing parameters, and sample quantity, please refer to... Figure 1This embodiment proposes a lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices. The method includes the following steps: S1. On-site vibration signal acquisition: Vibration sensors are installed on key parts of the rotating machinery to be monitored (bearings, gearboxes, coupling housings, etc.) to collect the raw time-series signals of vibration acceleration during operation, denoted as . ;in Indicates a time step. This indicates the number of time steps in signal acquisition (i.e., the time period from the start to the end of acquisition). The original time-series signal of vibration acceleration is also included. According to sampling frequency The acquired acceleration amplitude sequence. Original time-series signal of vibration acceleration. The data is transmitted in real time to the local cache of the edge diagnostic unit and stored therein, without being uploaded to a remote server, to ensure that subsequent processing is completed locally at the edge, thus achieving low-latency detection on the field side.

[0025] More specifically, the edge diagnostic unit is a resource-constrained terminal that can be deployed in an industrial field. Its processing power only supports online inference with a capacity of millions of floating-point operations (FLOPs), but it can still execute the subsequent steps of this invention.

[0026] S2, Edge-side sliding window segmentation and data packaging: The original time-series signal of vibration acceleration Fixed window length at the edge. and step length By performing sliding window segmentation, multiple vibration segment signals are obtained, denoted as: ; in, The first in the acquisition timing Vibration segment signals divided into windows , This represents the total number of vibration segment signals available in the current time period (i.e., the total number of windows). This indicates the transpose operation. Represents dimensional space. Indicates by and The calculated first The starting time step of each window. This segmentation operation transforms continuous long-series signals into standard input blocks that can be used for real-time inference at the edge, avoiding the need to push entire long-series signals to remote computing.

[0027] Organize all vibration segment signals into tensors according to channels, denoted as: ; in, The tensor representing the original vibrational segments formed by the tissue serves as the basic input structure for the subsequent feature extraction and diagnostic network batch processing.

[0028] S3. Normalization and denoising preprocessing: For each vibration segment signal Amplitude normalization and noise reduction are performed at the edge to weaken the dimensional changes caused by differences in operating conditions and reduce on-site power frequency / background interference, thereby obtaining the vibration segment tensor and constructing a vibration segment dataset. Normalization can take the following form: ; Among them, the segment mean Fragment standard deviation , Indicates the first In the tensor of the original vibration segment, the first Time series amplitude of each sampling point; This represents the normalized value of the first digit. The first vibration segment signal The amplitude of each sampling point; This represents the index of the sampling point in the window, with a value of ; To prevent the use of tiny constants with a denominator of zero, their values ​​are limited to (0, 1 × 10⁻⁶). -3 The normalized segment signal is obtained. .

[0029] Optionally, for the normalized segment signal Bandpass filtering, fixed bandwidth energy suppression, or empirical denoising are performed to obtain the noise-suppressed segment signal. This is used as a vibration segment tensor. All vibration segment tensors are concatenated to form a vibration segment dataset. The mathematical representation is as follows: .

[0030] After this step, the signal is represented with a uniform amplitude scale and comparable noise level, forming a "vibration segment dataset" for direct input into the subsequent lightweight fault diagnosis model at the edge. Steps S1-S3 realize data acquisition and edge preprocessing.

[0031] S4. Build a lightweight fault diagnosis model at the edge: A lightweight fault diagnosis model is built at the edge. The lightweight fault diagnosis model includes an adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier.

[0032] S5, Lightweight Feature Extraction of the Adaptive Residual Enhancement Module: To address the limitations of edge hardware storage and computing power (such as embedded chip memory being only in the MB range) caused by the large size of existing high-precision models, which prevent direct embedding, this invention constructs an adaptive residual enhancement module. The vibration segment dataset obtained in step S3 is then used... An adaptive residual enhancement module at the input edge is used to extract robust temporal fault features. This adaptive residual enhancement module includes parallel depthwise separable convolutional branches and inverse residual branches, as well as a gated fusion unit.

[0033] Specifically, depthwise separable convolutional features are extracted from the vibration segment tensors in the vibration segment dataset by a depthwise separable convolutional branch, denoted as... It is used to extract local fault-sensitive features such as impact components and periodic modulation components with extremely low computational cost; The inverse residual features of the vibration segment signal are extracted by the inverse residual branch, denoted as... By using the inverse residual structure of "channel expansion → nonlinear mapping → channel compression", higher-level robust semantic features are extracted, maintaining discriminativeness while reducing redundant parameters; in , They are respectively and The number of channels and timing length.

[0034] The adaptive residual enhancement module also includes a gated fusion unit for adaptively selecting information from two branches, specifically: Depthwise separable convolutional features and inverse residual characteristics After global average pooling, the vectors are concatenated, then subjected to learnable linear weighting and sigmoid activation to obtain the gated weight vector, mathematically represented as follows: ; in, Represents the gating weight vector. It is the Sigmoid activation function. This indicates global average pooling, used to average the temporal dimensions of features to obtain the channel description vector. This indicates vector-level concatenation. , These are the learnable weights and biases, respectively, which are learnable linearly weighted.

[0035] Based on this gating weight vector The outputs of the two branches are then weighted and fused channel by channel to obtain the enhanced feature representation. This is the characteristic of timing faults; the mathematical representation of this process is as follows: ; In the above formula This represents the Hadamard product across channels and time points. This represents the broadcast of an all-one vector over time. Therefore, the adaptive residual enhancement module can adaptively emphasize more robust features based on noise intensity and operating condition differences, avoiding single-path failure.

[0036] The timing fault features generated in this step It possesses both high-resolution local details and a globally robust mode, and has a low number of model parameters, making it suitable for real-time deployment on resource-constrained edge hardware.

[0037] The adaptive residual enhancement module constructed in this embodiment abandons the dependence of traditional deep networks on a large number of parameters and adopts a combination of depthwise separable convolution and inverse residual structure. It solves the physical problem of not being able to load complex diagnostic models under extremely low hardware configuration, enabling low-cost terminals with millions of floating-point computing power to run high-precision diagnostics.

[0038] S6, Lightweight discrimination enhancement for multi-group coordinate attention module: To address the problem that traditional lightweight models, due to structural simplification, sacrifice noise immunity in industrial environments with high noise and varying operating conditions, leading to failure under complex electromagnetic interference and mechanical vibration, this invention designs a multi-group coordinate attention module. The timing fault features obtained in step S5 are then processed... The input multi-group coordinate attention module is used to enhance the discriminative ability of features through a group attention mechanism with low computational complexity.

[0039] The core steps of the multi-group coordinate attention module are as follows: S61, Channel Grouping: Characteristics of timing faults Classified by channel dimension Subgroups, obtained Subgroup characteristics: ; Each subgroup feature Include One channel, For subgroup indexes, This represents the original number of channels. By dividing the time-series fault features into multiple subgroups, subsequent calculations can be performed independently within each subgroup, avoiding the high complexity of traditional self-attention methods that calculate correlations across all channels.

[0040] S62. Calculation of Time Response Vector: The features of each subgroup are averaged and fused along the channel dimension at each time step to obtain the time response vector of each subgroup feature. : ; in, For subgroups The Middle Each channel at time step Subgroup features.

[0041] S63. Intra-group attention calculation: The time response vector for each subgroup feature Lightweight convolutions are applied to generate temporal attention weights for each subgroup of features, mathematically represented as follows: ; in It is the Sigmoid activation function. This is a one-dimensional convolution operation; the temporal attention weights are learned in this step. This weight describes the importance of subgroup features in the time dimension.

[0042] S64. Intra-group weighting and reorganization: Based on temporal attention weights The features of each subgroup are weighted at each time step: ; in, These are the weighted features of the subgroups.

[0043] Finally, all weighted subgroup features are concatenated along the channel dimension to obtain the weighted temporal fault features: .

[0044] Through this process, the multi-group coordinate attention module utilizes a group attention mechanism to enhance the discriminative ability of features with linear computational complexity, while simultaneously reducing the computational complexity of traditional self-attention networks. Specifically, traditional self-attention networks or full-channel attention mechanisms typically require calculating the correlation matrix between all pairs of channels (of size 1). Therefore, the computational complexity is usually O(n). ,in , Let G and G represent the number of channels and the time series length, respectively. The computational complexity increases quadratically with the number of channels. However, the multi-group coordinate attention module of this invention employs a grouped computation strategy, avoiding the construction of a full-channel correlation matrix. This module only calculates the time response vector for each of the G subgroups and performs a one-dimensional convolution, resulting in a computational complexity approximately O(G). (Or, when the convolution kernel is small, it can be simplified to linear complexity) ),in Indicates the number of groups. ,because The value is very small, making the computational complexity from The quadratic relationship is reduced to a linear relationship, thus significantly reducing the computational load and making it suitable for execution on edge computing devices. This module can retain key information and enhance diagnostic accuracy with lower computational overhead, making it suitable for deploying lightweight, low-latency real-time inference in rotating machinery fault detection tasks.

[0045] The final output is the weighted timing fault characteristics. This allows for subsequent lightweight classifiers to determine fault categories and assess health status. In this way, the model achieves efficient feature extraction and low-latency fault diagnosis.

[0046] This step suppresses irrelevant frequency bands, background vibrations, and frame noise, highlighting the local impulse response and modulation envelope related to the fault mode, so that the output characteristics remain stable under strong noise, variable speed, and variable load conditions.

[0047] The multi-group coordinate attention module constructed in this invention maximizes the use of limited computing resources to lock onto key fault frequency bands under the premise of limited computing power, thereby overcoming environmental noise interference locally and achieving accurate diagnosis without uploading data to the cloud for cleaning.

[0048] On the other hand, through the depthwise separable convolution branch and the inverse residual branch in the adaptive residual enhancement module, and the group attention mechanism in the multi-group coordinate attention module, this invention can effectively extract the key features of rotating machinery faults and maintain high diagnostic accuracy under complex working conditions and noise interference. Specifically, the depthwise separable convolution branch captures rapidly changing local details and high-frequency impact components in the vibration signal, while the inverse residual branch extracts global context information and low-frequency periodic components through channel expansion and compression. The gated fusion unit performs global average pooling and linear mapping on the outputs of the above two branches, so that the fault diagnosis process can dynamically adjust the attention to high-frequency or low-frequency features according to the spectral characteristics of the current vibration signal without manual intervention, realizing adaptive diagnosis under complex working conditions. At the same time, it effectively suppresses the influence of irrelevant signals such as background noise and power frequency interference, improves the robustness and stability of fault diagnosis, and accurately predicts the fault type and health status. Even when the data is incomplete, it can still ensure high prediction accuracy.

[0049] S7. Perform model training: The time-series fault features obtained in step S6 are fed into a lightweight classifier to determine the fault category and obtain the classification prediction probability; the total loss is calculated based on the classification prediction probability, and the lightweight fault diagnosis model is trained.

[0050] The lightweight classifier has the following specific structure: The global average pooling layer is used to average and compress the weighted temporal fault features along the time dimension, reducing the dimensionality from... Dimensional reduction This significantly reduces the number of parameters in subsequent calculations while preserving the key features of the channel. The Dropout layer is used to randomly disconnect neuron connections during training to prevent the model from overfitting under small sample conditions. In this embodiment, the dropout rate is set to 0.2. The linear classification layer, consisting of a fully connected layer, has the number of nodes corresponding to the total number of fault categories, and is used to map feature vectors to the label space. The softmax activation layer uses the softmax activation function to transform the output of the linear classification layer into a probability distribution, and finally outputs the predicted probability of each fault category as the classification prediction probability.

[0051] This invention employs a lightweight classifier structure. By simplifying the network hierarchy and eliminating redundant computations, the lightweight classifier reduces the model's memory footprint and computational burden. Compared to traditional classifiers, it uses a more efficient activation function and fewer parameters, thereby achieving low-latency real-time inference. Simultaneously, the multi-group coordinate attention module utilizes a group attention mechanism to enhance the discriminative ability of features with linear computational complexity, while reducing the computational complexity of traditional self-attention networks. This makes the model more suitable for edge deployment, enabling the entire fault diagnosis process to run efficiently and with low power consumption on edge devices. Compared to traditional deep learning models, the aforementioned lightweight design significantly reduces computational complexity and storage requirements, ensuring that the rotating machinery fault diagnosis system can achieve real-time inference on edge devices with limited computing resources.

[0052] The training process of the model includes the following parts: S71. Training Data Preparation: The input data consists of vibration signals from rotating machinery, processed through steps S1 to S5 to obtain weighted time-series fault characteristics. This includes samples under various operating conditions and failure modes.

[0053] S72. Training Objective and Loss Function: The training objective is to minimize the classification loss. In conjunction with reconstruction losses Achieving a balance between feature extraction and diagnosis. The classification loss uses the cross-entropy loss function. ,in It is the first The actual fault category label corresponding to each sample The output of the lightweight fault diagnosis model is the first The classification prediction probability corresponding to each sample. Reconstruction Loss is used to optimize the signal recovery accuracy of the model, especially in the presence of noise and missing signals, ensuring that the model can effectively recover key fault information from the compressed features. Specifically, the reconstruction loss... The mean squared error (MSE) function is used for calculation, and its mathematical expression is as follows: ; in, Indicates the first The original (or pre-processed) vibration segment signal of the input is in the first... The amplitude at each time step; Indicates the first Weighted temporal fault features extracted from each sample; This indicates that the auxiliary reconstruction decoder (typically composed of lightweight deconvolutional layers or upsampling layers) is used to convert features... Map back to Dimensionally consistent reconstructed signal; This represents the square of the L2 norm.

[0054] The effect of introducing reconstruction loss during model training is that it acts as a regularization term, forcing the model to retain the global waveform structure information of the signal while extracting fault discrimination features. This effectively prevents the model from losing key physical signal features due to overfitting during extremely lightweight processes, thus ensuring the robustness and integrity of features even in the presence of noise interference or partial data loss at the edges.

[0055] S73, Optimization Strategy: The Adam optimizer is used for weight updates, and a dynamic learning rate adjustment strategy is employed. The total loss is minimized through backpropagation at each training epoch. ( (Weigh the loss term) and select the optimal model weights using the validation set.

[0056] S74, Training Techniques: To avoid overfitting, data augmentation (such as time window perturbation) and early stopping strategies (automatically stopping training based on validation set performance) are employed.

[0057] The curve showing the change in average training loss with training rounds during the training process can be found in [reference needed]. Figure 2 The curve showing the change in average training accuracy with the number of training rounds can be found in [reference needed]. Figure 3 .

[0058] Through the above steps, a trained lightweight fault diagnosis model is finally obtained, which can perform real-time inference at the edge.

[0059] S8. Reasoning and Feature Standardization: During the inference phase, a pre-trained lightweight fault diagnosis model is used to receive vibration signals collected in real time at the edge for fault diagnosis. To improve inference accuracy, the input vibration signals undergo standardization to ensure consistency of input characteristics under different operating conditions. Specifically: S81, Input Normalization: For each input vibration signal Normalization is performed: ; in, This represents the normalized vibration signal. and These represent the mean and standard deviation of the collected vibration signals, respectively. This represents the index of the sampling point within the current input vibration signal window during the inference phase, and its value range remains consistent with that during the training phase (i.e., (), used to locate specific data points in a time series.

[0060] S82, Noise Reduction and Preprocessing: On edge devices, bandpass filtering or fixed-bandwidth energy suppression is performed to mitigate the effects of high-frequency noise and ensure efficient feature extraction.

[0061] S83, Feature Input Model: The standardized signal is input into a lightweight fault diagnosis model for inference. This model can accurately extract key fault features from the signal and prepare for subsequent classification. The model outputs the classification prediction probability for each vibration signal, and selects the category with the highest probability as the final diagnostic result for that vibration signal.

[0062] Steps S4-S8 realize the construction and application of the lightweight fault diagnosis model.

[0063] S9. Fault Classification and Alarm Output: Based on the diagnostic results obtained during the inference phase, the edge device triggers the corresponding alarm mechanism. Specifically: S91, Alarm Triggered: The fault level is determined based on the diagnostic results. If the diagnostic result is "critical fault", the system will trigger a high-priority alarm; if it is "early fault", a medium-priority alarm will be triggered and a warning message will be provided; if it is "normal", no alarm will be triggered.

[0064] S92, Local Output and Feedback: Edge devices transmit fault levels to the host computer via audible and visual alarms and information push notifications, or record fault data locally, providing real-time feedback and enabling lightweight fault diagnosis of rotating machinery. Alarm triggering does not require uploading raw data to the cloud, avoiding bandwidth and latency burdens and ensuring the real-time nature of the diagnostic process.

[0065] Step S9 implements edge-end early warning.

[0066] This invention significantly reduces data transmission latency by directly deploying fault diagnosis tasks on edge computing devices, avoiding the delays caused by data uploading to remote servers in traditional centralized computing methods. This enables fault diagnosis to be performed in a real-time environment, ensuring immediate fault warnings and responses for rotating machinery under high load and complex operating conditions, and greatly improving the safety and stability of the production process. For shutdown protection scenarios with extremely high real-time requirements in rotating machinery fault diagnosis, the adaptive residual enhancement module, multi-group coordinate attention module, and lightweight classifier constructed in this invention feature targeted lightweight designs, eliminating the network latency uncertainty caused by data uploading to the cloud, ensuring inference and feedback are completed within milliseconds, and meeting the real-time warning needs of rotating machinery under sudden faults.

[0067] This application embodiment also provides a lightweight fault diagnosis system for rotating machinery that can be deployed in real time on edge devices, used to apply the aforementioned lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices, including: The data acquisition and edge preprocessing module is used to acquire the raw time-series signal of vibration acceleration during the operation of rotating machinery, cache it to the edge, perform sliding window segmentation, normalize and preprocess it to construct a vibration segment dataset. The lightweight fault diagnosis model module is used to build and train a lightweight fault diagnosis model, and to use the trained lightweight fault diagnosis model for inference to obtain diagnostic results. The lightweight fault diagnosis model includes an adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier. The edge warning module is used to determine the fault level based on the diagnostic results, enabling edge devices to transmit the fault level to the host computer through audible and visual alarms and information push, or to record fault data locally, providing real-time feedback and realizing lightweight fault diagnosis of rotating machinery. If the edge warning module diagnoses a serious fault, the system will trigger a high-priority alarm; if the diagnosis result is an early fault, it will trigger a medium-priority alarm and provide a warning; if the diagnosis result is normal, no alarm will be triggered.

[0068] This application also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the described method for lightweight fault diagnosis of rotating machinery that can be deployed in real time on an edge device.

[0069] The fault diagnosis method of this invention has broad application prospects in multiple industrial scenarios, especially in high-end equipment manufacturing, aerospace, energy and other fields, and will provide a more accurate, efficient and low-cost solution for equipment health management.

[0070] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0071] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0072] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices, characterized in that, include: Vibration sensors are installed on key parts of the rotating machinery to be monitored to collect the raw time-series signals of vibration acceleration during operation and store them at the edge. The original time-series vibration acceleration signal is segmented and preprocessed at the edge to obtain the vibration segment tensor, and then combined with the corresponding real fault category label to construct the vibration segment dataset. Sliding window segments are divided according to fixed window length and step size to obtain multiple vibration segment signals. All vibration segment signals are organized into tensors according to channels and amplitude normalization and noise reduction are performed at the edge. A lightweight fault diagnosis model is built at the edge; the lightweight fault diagnosis model includes an adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier; The vibration segment dataset is input into the adaptive residual enhancement module to extract the temporal fault features. The temporal fault features are then fed into the multi-group coordinate attention module to obtain weighted temporal fault features. The adaptive residual enhancement module includes parallel depthwise separable convolutional branches and inverse residual branches, as well as a gated fusion unit; The weighted temporal fault features are fed into a lightweight classifier to determine the fault category and obtain the classification prediction probability. A lightweight fault diagnosis model is trained by calculating the total loss based on the classification prediction probability; the total loss includes classification loss and reconstruction loss. Vibration signals are collected in real time at the edge, and a trained lightweight fault diagnosis model is used to perform fault diagnosis and obtain the diagnosis results; the fault level is determined based on the diagnosis results. Edge devices transmit fault levels to the host computer via audible and visual alarms and information push notifications, or record fault data locally to provide real-time feedback, enabling lightweight fault diagnosis of rotating machinery.

2. The method for lightweight fault diagnosis of rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The original time-series signal of vibration acceleration is a sequence of acceleration amplitude values ​​collected at the sampling frequency; The original vibration acceleration timing signal is transmitted in real time to the local cache of the edge diagnostic unit and stored therein, without being uploaded to a remote server. The edge diagnostic unit is a resource-constrained terminal that can be deployed in an industrial field and has a million-level floating-point computing power, which can support the execution of subsequent processing steps of the present invention.

3. The lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The process of performing sliding window segmentation and preprocessing on the original time-series vibration acceleration signal at the edge to obtain the vibration segment tensor includes: The sliding window is segmented according to a fixed window length and step size to obtain multiple vibration segment signals; All vibration segment signals are organized into tensors according to channels, and amplitude normalization and noise reduction are performed at the edges to obtain vibration segment tensors.

4. The lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The step of inputting the vibration segment dataset into the adaptive residual enhancement module to extract time-series fault features includes: The vibration segment tensors in the vibration segment dataset are input into the depthwise separable convolution branch and the inverse residual branch, respectively. The depthwise separable convolution branch performs depthwise separable convolution on the vibration segment signal to generate depthwise separable convolution features. The inverse residual branch is an inverse residual structure that extracts inverse residual features from the vibration segment signal. The depthwise separable convolutional features and the inverse residual features are fed into the gated fusion unit. After global average pooling, they are concatenated at the vector level, and then subjected to learnable linear weighting and sigmoid activation to obtain the gated weight vector. Then, based on the gated weight vector, the depthwise separable convolutional features and the inverse residual features are fused channel by channel to obtain the temporal fault features.

5. The lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The step of feeding the temporal fault features into the multi-group coordinate attention module to obtain weighted temporal fault features includes: The time-series fault features are divided into multiple sub-group features according to the channel dimension. The features of each sub-group are averaged and fused along the channel dimension at each time step to obtain the time response vector of each sub-group feature. Lightweight convolutions are applied to the temporal response vectors of each subgroup feature to generate temporal attention weights for each subgroup feature; The features of each subgroup are weighted at each time step according to the temporal attention weight to obtain the weighted subgroup features; By concatenating all weighted subgroup features along the channel dimension, weighted temporal fault features are obtained.

6. The method for lightweight fault diagnosis of rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The step of inputting weighted temporal fault features into a lightweight classifier for fault category determination and obtaining classification prediction probabilities includes: The weighted temporal fault features are averaged and compressed over the time dimension, and then passed through a dropout layer, a linear classification layer, and a softmax activation layer to obtain the classification prediction probability.

7. The lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The classification loss is the cross-entropy loss calculated between the classification prediction probability output by the lightweight fault diagnosis model and the true fault category label.

8. A lightweight fault diagnosis method for rotating machinery that can be deployed in real time on edge devices according to claim 1, characterized in that: The edge end acquires vibration signals in real time, and uses a trained lightweight fault diagnosis model to perform fault diagnosis, obtaining diagnostic results, including: The collected vibration signals are normalized, denoised, and preprocessed. The preprocessed vibration signal is fed into the trained lightweight fault diagnosis model for inference to obtain the classification prediction probability of each vibration signal. Based on the classification prediction probability, the category with the highest probability is selected as the final diagnostic result for the vibration signal.

9. A lightweight fault diagnosis system for rotating machinery that can be deployed in real time on an edge device, used to apply the lightweight fault diagnosis method for rotating machinery that can be deployed in real time on an edge device as described in any one of claims 1-8, characterized in that, include: The data acquisition and edge preprocessing module is used to acquire the raw time-series signal of vibration acceleration during the operation of rotating machinery, cache it to the edge, perform sliding window segmentation, normalize and preprocess it to construct a vibration segment dataset. The lightweight fault diagnosis model module is used to build and train a lightweight fault diagnosis model, and to use the trained lightweight fault diagnosis model for inference to obtain diagnostic results. The lightweight fault diagnosis model includes an adaptive residual enhancement module, a multi-group coordinate attention module, and a lightweight classifier. The edge-end early warning module is used to determine the fault level based on the diagnostic results, enabling edge devices to transmit the fault level to the host computer through audible and visual alarms and information push, or to record fault data locally, providing real-time feedback and realizing lightweight fault diagnosis of rotating machinery.

10. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the rotating machinery lightweight fault diagnosis method that can be deployed in real time on an edge device according to any one of claims 1-8.

Images