Vibration detection method for constructing edge calculation model based on Simulink physical modeling

Abstract

The invention discloses a vibration detection method for constructing an edge calculation model based on Simulink physical modeling, and the method comprises the steps: constructing a multi-degree-of-freedom vibration model through Simulink, dynamically adjusting the quality, rigidity, a damping matrix, temperature and load parameters, simulating the vibration characteristics of a complex working condition, and fusing simulation and actual measurement data to generate a training set; the model is converted into an edge available format and compressed to be within 100 KB, and edge hardware is adapted; carrying out joint training and online updating on edge nodes by utilizing transfer learning and the like; and the vibration signals are classified in real time through a lightweight model. According to the method, the problems of fixed model parameters, insufficient data fusion and poor edge adaptability in the traditional technology are solved, accurate simulation of the vibration characteristics of the industrial equipment, efficient data utilization and low-delay detection of the edge end are realized, the vibration detection precision and real-time performance are improved, and the method is suitable for fault diagnosis and predictive maintenance of the industrial equipment.

Description

Technical Field

[0001] The present invention relates to the field of industrial equipment vibration detection, and in particular to a vibration detection method for building an edge computing model based on Simulink physical modeling. Background Art

[0002] With the acceleration of industrial intelligence, the stable operation of industrial equipment is crucial to production efficiency and safety. As a key technology for determining equipment operating status, vibration detection, with its accuracy and real-time performance, has become a focus of industry attention. Traditional vibration detection relies heavily on centralized data processing and general algorithm models, making it difficult to meet the demands of complex industrial scenarios and the real-time analysis of massive amounts of data. Meanwhile, the rise of edge computing technology has enabled local processing of device data, but how to effectively apply it to vibration detection and achieve high-precision, low-latency fault diagnosis remains a pressing challenge.

[0003] Existing technologies have obvious defects in vibration detection. First, in terms of model construction and data processing, traditional methods often use general vibration models with fixed parameters, which cannot accurately simulate the complex vibration characteristics of industrial equipment under different working conditions such as load changes, equipment wear, and temperature drift. In addition, its data processing is mostly based on a single data source, lacking the effective integration of simulation data and measured data, resulting in insufficient model training samples, difficulty in dealing with various failure modes, and limited detection accuracy. Second, in terms of equipment adaptation, the models generated by traditional detection methods are mostly general-purpose, and do not fully consider the limited resource characteristics of edge computing devices. There are problems with the large number of model parameters and high computational complexity, which makes it difficult to deploy directly on edge nodes, resulting in poor real-time detection and an inability to meet the needs of rapid diagnosis of industrial equipment failures. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a vibration detection method based on Simulink physical modeling to construct an edge computing model.

[0005] The technical solution adopted by the present invention is to construct a vibration detection method of an edge computing model based on Simulink physical modeling, including the following steps: Step S1: In the Simulink environment, the Newton-Euler equation and the Lagrange equation are used to construct four-degree-of-freedom and six-degree-of-freedom vibration models for industrial equipment. The mass matrix, stiffness matrix, and damping matrix are dynamically adjusted through MATLAB scripts and Simulink parameter management tools to simulate the complex vibration characteristics of industrial equipment under different working conditions such as load changes, equipment wear, and temperature drift. Step S2: Design a lightweight model export process in Simulink, convert the constructed vibration simulation model into a format suitable for edge computing, and use Simulink Coder or MATLAB Coder to generate C / C++ code that is compatible with STM32 and Jetson edge hardware. Step S3: deploy a three-axis acceleration sensor to collect vibration signals of industrial equipment. The collected signals include acceleration, displacement, and angular velocity. DMA technology is used to accelerate the data transmission process and reduce the CPU load. Step S4: Construct a data fusion module in Simulink to mix the simulation data including different fault modes generated in step S1 with the measured data collected in step S3 according to a preset ratio to generate a training set; Step S5: Use the ONNX format to unify the model interface in Simulink, integrate the model compression tool, and generate a model with less than 100KB parameters and suitable for edge deployment through quantization-aware training and pruning operations. This model supports TensorFlowLite, PyTorchMobile, and NanoEdgeAI frameworks and is compatible with STM32 and JetsonNano hardware. Step S6: Design a hybrid training process in Simulink, input the training set generated in step S4 into the edge model obtained in step S5, conduct joint training using transfer learning or federated learning strategies, implement online learning on edge nodes, and update model parameters through incremental learning; Step S7: Deploy a lightweight classification model on the edge node, use Simulink's code generation tool to convert the reasoning logic into C / C++ code, and perform fault classification on the processed vibration signal.

[0006] Furthermore, in step S1, when constructing the vibration model, the vibration frequency parameter is introduced , quality parameters , stiffness parameter , damping parameter , establish the following vibration model formula: ,in, Indicates that industrial equipment is at The acceleration of Indicates that industrial equipment is at speed, Indicates that industrial equipment is at The displacement, Indicates the effect on industrial equipment at time The external excitation force is determined by the quality parameters , stiffness parameters , damping parameters and external incentives Dynamic adjustment to simulate the vibration state of industrial equipment under different working conditions.

[0007] Furthermore, in step S2, in the process of converting the simulation model into a format that can be used for edge computing, the model compression ratio parameter , memory usage parameters , establish the following conversion optimization formula:

[0008]

[0009] in, Indicates the size of the model after conversion suitable for edge computing. Indicates the size of the original simulation model in Simulink, Memor Indicates the memory occupied by the model in the edge computing device. Indicates that the original simulation model occupies memory in the Simulink environment, and the model compression ratio parameter The model size is adjusted to meet the storage and operation requirements of the edge computing device based on the memory usage parameters. Evaluate the memory resource usage of the edge computing device by the converted model.

[0010] Furthermore, in step S3, when using a three-axis acceleration sensor to collect vibration signals of industrial equipment, the sensor sampling frequency parameter , sensor sensitivity parameters , establish the following signal acquisition model: ,in, Indicates the voltage value of the vibration signal collected by the sensor. Indicates the actual vibration acceleration of industrial equipment by adjusting the sensor sensitivity parameters and sampling frequency parameters , optimize the sensor's collection effect on industrial equipment vibration signals.

[0011] Furthermore, in step S4, when constructing the data fusion module to mix the simulated data with the measured data, the data mixing ratio parameter is introduced. , establish the following data fusion formula: ,in, Represents the generated training set data, represents the simulation data, Represents the measured data, by adjusting the data mixing ratio parameters , control the proportion of simulation data and measured data in the training set.

[0012] Furthermore, in step S5, in the process of generating a model suitable for edge deployment, the number of model parameters is combined , model computational complexity parameter , establish the following model optimization formula: ,in, Indicates the first parameters, Indicates the The time complexity of the parameters in the calculation process is calculated by the number of model parameters. The control of model computational complexity parameters Optimize the edge computing equipment to generate a lightweight model that meets the computing power and operating efficiency requirements.

[0013] Furthermore, in step S6, when performing joint training, the training iteration number parameter is introduced , the learning rate parameter , establish the following training optimization formula: ,in, Indicates in The model parameters at the training iteration, Indicates in The model parameters at the training iteration, Indicates in The loss function for the training iteration Model parameters The gradient of the training iteration number parameter is adjusted by and the learning rate parameter , optimize the joint training process.

[0014] Furthermore, in step S7, when fault classification is performed on the edge node, the vibration amplitude fluctuation coefficient is introduced. , vibration frequency deviation , and combined with the classification decision threshold parameter , establish the following fault classification formula:

[0015] in, Indicates the fault classification results of industrial equipment; It is the classification probability value of the vibration amplitude of industrial equipment by the lightweight classification model, reflecting the possibility that the vibration amplitude is within the normal range; It is the classification probability value of the vibration frequency of industrial equipment by the lightweight classification model, reflecting the possibility that the vibration frequency conforms to the normal state; the vibration amplitude fluctuation coefficient Used to measure the weight of vibration amplitude in fault judgment, vibration frequency deviation Used to measure the weight of vibration frequency in fault judgment; by adjusting as well as , fault classification of industrial equipment is performed based on the comprehensive characteristics of vibration amplitude and frequency.

[0016] Furthermore, in the entire method process, the operating environment temperature parameters of the industrial equipment are combined , the temperature parameter As an influencing factor, it is introduced into the vibration model of step S1, and the expanded vibration model formula is established:

[0017] in, Indicates the temperature The quality parameters of the following industrial equipment, Indicates the temperature The stiffness parameters of the following industrial equipment, Indicates the temperature The damping parameters of the industrial equipment are as follows: Indicates the temperature and time The external excitation force acting on industrial equipment, Indicates the temperature and time Acceleration of industrial equipment, Indicates the temperature and time Lower the speed of industrial equipment, Indicates the temperature and time The displacement of industrial equipment under different temperature environments is simulated through this extended model.

[0018] Furthermore, in the whole process, the load variation parameters of industrial equipment are combined , the load change parameter As an influencing factor, it is introduced into the vibration model of step S1 and the data fusion process of step S4, and the formula is established in the vibration model:

[0019] in, Indicates that the load The quality parameters of industrial equipment are as follows, Indicates that the load The stiffness parameters of the following industrial equipment, Indicates that the load The damping parameters of the industrial equipment are as follows: Indicates that the load and time The external excitation force acting on industrial equipment, Indicates that the load and time Acceleration of industrial equipment, Indicates that the load and time Lower the speed of industrial equipment, Indicates that the load and time The displacement of industrial equipment under load; in the process of data fusion, according to the load change parameters Adjust the data mixing ratio parameters , establish the formula:

[0020] in, and and load The relevant weight coefficient is obtained by combining the load change parameters , optimize the vibration model and data fusion process.

[0021] Beneficial effects: The present invention proposes a vibration detection method for constructing an edge computing model based on Simulink physical modeling. At the model construction and data processing level, the method uses Simulink to construct a multi-degree-of-freedom vibration model. By dynamically adjusting the mass, stiffness, and damping matrix and introducing environmental parameters such as temperature and load, the vibration characteristics of industrial equipment under complex working conditions are accurately simulated, breaking through the limitations of traditional fixed parameter models. At the same time, the simulation data and measured data covering various fault modes are mixed in proportion through the data fusion module to generate a training set, which solves the problem of insufficient samples from a single data source and improves the model's detection accuracy for various fault modes. In terms of device adaptation, the method uses the Simulink lightweight model export process to convert the simulation model into a format that can be used for edge computing and generate code that is adapted to edge hardware such as STM32 and Jetson. The model compression tool is combined to realize the deployment of lightweight models with a parameter amount of less than 100KB. At the same time, a hybrid training process is designed, and transfer learning, federated learning, and online incremental learning strategies are used to realize dynamic updates of model parameters at edge nodes, avoiding the problem of insufficient real-time performance of traditional general models due to excessive computational complexity. In addition, the lightweight classification model deployed on the edge node achieves rapid classification of equipment failures through a comprehensive decision-making mechanism that integrates multi-dimensional parameters such as vibration amplitude and frequency. While ensuring detection accuracy, the overall solution significantly reduces the resource usage of edge computing devices, improves the real-time performance and adaptability of industrial equipment vibration detection to working conditions, and provides efficient technical support for intelligent industrial operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0023] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 As shown in the figure, the vibration detection method of the edge computing model based on Simulink physical modeling includes the following steps: Step S1: In the Simulink environment, use the Newton-Euler equations and Lagrange equations to construct four-degree-of-freedom and six-degree-of-freedom vibration models for industrial equipment. The four-degree-of-freedom model covers translation and rotation, while the six-degree-of-freedom model includes three-dimensional translation and three-dimensional rotation. Dynamically adjust the mass matrix, stiffness matrix, and damping matrix using MATLAB scripts and Simulink parameter management tools to simulate the complex vibration characteristics of industrial equipment under different operating conditions such as load changes, equipment wear, and temperature drift. Specifically, within the Simulink environment, Newton-Euler and Lagrange equations are used to construct four-degree-of-freedom (translation + rotation) and six-degree-of-freedom (3D translation + 3D rotation) vibration models for industrial equipment. The four-degree-of-freedom model simulates complex in-plane motion of the equipment, such as the vertical vibration and pitching motion of a vehicle suspension, while the six-degree-of-freedom model further covers three-dimensional motion, such as the axial play and radial oscillation of an industrial pump. Using MATLAB scripts or the Simulink parameter management tool, the values ​​of the mass, stiffness, and damping matrices can be dynamically adjusted. For example, the mass matrix can be adjusted from 50 kg to 200 kg to simulate varying equipment loads, and the damping matrix from 10 N·s / m to 50 N·s / m to reflect bearing wear. These parameter adjustments accurately replicate the vibration characteristics of the equipment under conditions such as sudden load changes, component wear, and ambient temperature fluctuations (e.g., -20°C to 80°C), providing a high-precision model foundation for subsequent simulation data generation. The core significance of this step is to break the limitations of the traditional single-degree-of-freedom model through multi-physics field modeling, so that the simulation data is closer to the actual equipment operation status and reduces dependence on measured data.

[0025] First, an appropriate degree-of-freedom model is selected based on the structural characteristics and motion of the industrial equipment. A four-degree-of-freedom model is used for planar motion equipment, and a six-degree-of-freedom model is used for spatial motion equipment. Initial mass, stiffness, and damping matrix parameters are then calibrated using finite element analysis or experimental data. Finally, using Simulink's parameter sweep function, model parameters are batch adjusted under different operating conditions to generate a simulation data set encompassing multiple failure modes. This step establishes the physical model foundation for the entire vibration detection method, ensuring that subsequent analysis accurately reflects the equipment's true vibration state.

[0026] Step S2: Design a lightweight model export process in Simulink to convert the constructed vibration simulation model into a format suitable for edge computing, such as ONNX or TensorFlow Lite. Use Simulink Coder or MATLAB Coder to generate C / C++ code that is compatible with edge hardware such as STM32 and Jetson. Specifically, a dedicated lightweight model export process is designed within Simulink to convert the constructed vibration simulation model into a format compatible with edge computing devices, such as ONNX or TensorFlow Lite. During this conversion process, optimized C / C++ code is generated using the Simulink Coder or MATLAB Coder tools. The code is optimized for the hardware architectures of the STM32 microcontroller (200MHz) and the Jetson edge computing platform (equipped with a 128-core GPU). For example, the STM32's DMA direct memory access technology is leveraged to reduce CPU data transfer load, and matrix operation logic is optimized for the Jetson's parallel computing capabilities. The model conversion process requires controlling the output file size, ensuring that the converted model file is less than 1MB, while maintaining an extraction accuracy of above 95% for key vibration characteristic parameters (such as acceleration amplitude and frequency distribution).

[0027] Simulink's model analysis tools are used to assess model complexity and identify layers and parameters that require compression. Quantization-aware training techniques are then applied to convert 32-bit floating-point parameters to 8-bit integers while maintaining model accuracy. Pruning is then performed to remove connections and neurons that have little impact on the output, reducing the number of model parameters. Finally, an appropriate inference framework and optimization strategy are selected based on the hardware characteristics of the target edge device to generate a ready-to-deploy code package. This step addresses the difficulty of traditional simulation models running on resource-constrained edge devices, achieving a precise match between the model's computing power and hardware resources.

[0028] Step S3: deploy a three-axis acceleration sensor to collect vibration signals of industrial equipment. The collected signals include acceleration, displacement, and angular velocity. DMA technology is used to accelerate the data transmission process and reduce the CPU load. Specifically, triaxial accelerometers are deployed to collect vibration signals from industrial equipment, with a sampling frequency of 10kHz-100kHz to cover the equipment's primary vibration frequency range (e.g., the 10Hz-10kHz characteristic frequency of rotating machinery). The sensors utilize industrial-grade MEMS devices with a measurement range of ±200g and a sensitivity of 0.001g, capable of detecting micron-level vibration displacements. The collected data, including multi-dimensional information such as acceleration, displacement, and angular velocity, is transmitted to edge computing nodes via SPI or I2C interfaces. DMA technology is used to accelerate data transmission, achieving a data transfer rate of up to 1MB / s, reducing CPU load to less than 30%, and ensuring real-time collection and transmission of high-frequency vibration data.

[0029] The sensor installation location is selected based on the equipment structure and vibration characteristics to ensure that vibration signals from key areas are captured. The sensor's sampling frequency and range are then configured to match the equipment's vibration frequency range. Hardware filtering circuits and software digital filtering algorithms are then used to remove noise interference and improve signal quality. Finally, DMA technology is used to enable direct data transfer to memory, reducing CPU intervention and improving data acquisition efficiency. This step ensures high-quality acquisition and efficient transmission of vibration signals, providing a reliable data source for subsequent data analysis.

[0030] Step S4: Construct a data fusion module in Simulink to mix the simulation data including different fault modes generated in step S1 with the measured data collected in step S3 according to a preset ratio to generate a training set; Specifically, a data fusion module was constructed in Simulink. The simulation data generated in step S1, covering various fault modes (such as bearing pitting, gear cracks, and rotor imbalance), was mixed with the measured data collected in step S3 in a ratio of 3:7 to 7:3 to generate a training set. The simulation data included vibration responses for over 1,000 fault conditions, while the measured data was collected from the operating records of over 50 units of the same type, ensuring the diversity and representativeness of the training set. During the data fusion process, the simulation and measured data were time-synchronized and feature-matched to eliminate data bias, ensuring that the fused training set accurately reflects the actual operating status of the equipment.

[0031] This step involves preprocessing the simulated and measured data to unify the data format and sampling rate. Timestamp matching or feature alignment algorithms are then used to synchronize the data. Feature extraction and dimensionality reduction techniques are then used to extract key features from the raw data that characterize the equipment's operating status. Finally, the simulated and measured data are mixed according to a preset ratio and data augmentation is performed to expand the number of training samples. This multi-source data fusion step addresses the issue of insufficient samples from a single data source, improving the quality of the training set and the model's generalization capabilities.

[0032] Step S5: Use the ONNX format to unify the model interface in Simulink, integrate the model compression tool, and generate a model with less than 100KB parameters and suitable for edge deployment through quantization-aware training, pruning, and other operations. This model supports frameworks such as TensorFlowLite, PyTorchMobile, and NanoEdgeAI, and is compatible with hardware such as STM32 and JetsonNano. Specifically, the ONNX format is used in Simulink to unify the model interface, integrate model compression tools, and generate a model with less than 100KB of parameters suitable for edge deployment through quantization-aware training, pruning, and other operations. During the model compression process, key neuron connections sensitive to vibration characteristics are retained, and the pruning rate is controlled at 70%-90%. At the same time, 8-bit integer quantization is used instead of 32-bit floating-point operations to reduce computational complexity. The optimized model supports multiple inference frameworks such as TensorFlowLite, PyTorchMobile, and NanoEdgeAI. The optimal deployment solution can be selected based on the hardware configuration of the edge device, such as selecting the NanoEdgeAI framework on the ARMCortex-M4F processor and the TensorFlowLite framework on the NVIDIAJetson platform.

[0033] The trained model is converted to the ONNX format to achieve compatibility across different frameworks. Model compression tools are then used to prune and quantize the model, reducing the number of model parameters and computational complexity. Based on the hardware characteristics of the target edge device, an appropriate inference framework is selected for model deployment. Finally, performance testing and accuracy evaluation are conducted to verify the effectiveness of the optimized model on the edge device. This step, through model optimization and framework adaptation, enables complex deep learning models to run efficiently on resource-constrained edge devices, achieving a balance between high accuracy and low resource consumption.

[0034] Step S6: Design a hybrid training process in Simulink, input the training set generated in step S4 into the edge model obtained in step S5, conduct joint training using transfer learning or federated learning strategies, implement online learning on edge nodes, and update model parameters through incremental learning; Specifically, a hybrid training process is designed within Simulink, where the training set generated in step S4 is input into the edge model obtained in step S5, and joint training is performed using transfer learning or federated learning strategies. During transfer learning, the edge model is initialized using pre-trained general vibration model parameters to reduce the need for large-scale training data. During federated learning, multiple edge nodes collaboratively update model parameters without sharing original data to protect data privacy. At the same time, online learning is implemented at the edge nodes, and the model parameters are updated through incremental learning. The learning rate is set to 0.001-0.01 to ensure that the model can quickly adapt to changes in the operating status of the equipment. During online learning, a model update is triggered every time 1,000 sets of new data are collected, and the update time is controlled within 100ms to ensure real-time performance.

[0035] Based on the computing power and data privacy requirements of the edge device, an appropriate training strategy (transfer learning or federated learning) is selected. The edge model is then initialized using pre-trained model parameters and fine-tuned on a local training set. The trigger conditions and update frequency for online learning are then set, and the model parameters are regularly updated with newly acquired data. Finally, a model evaluation mechanism is used to verify the performance and stability of the updated model. This step enables continuous model learning and optimization, enabling the detection system to adapt to challenges brought about by factors such as equipment aging and changing operating conditions, maintaining long-term detection accuracy.

[0036] Step S7: Deploy a lightweight classification model on the edge node, use Simulink's code generation tool to convert the reasoning logic into C / C++ code, and perform fault classification on the processed vibration signal.

[0037] Specifically, a lightweight classification model is deployed at the edge node. Simulink's code generation tools are used to convert the inference logic into C / C++ code to perform fault classification on the processed vibration signals. The classification model utilizes a multilayer perceptron or convolutional neural network architecture. The input layer receives the vibration signal's time-domain features (such as RMS value and crest factor) and frequency-domain features (such as spectral amplitude and energy distribution). The output layer provides classification results for 5-10 fault types (such as normal, imbalance, misalignment, bearing fault, and gear fault). The model's inference time is kept within 10ms, with a classification accuracy exceeding 98% and a false alarm rate below 1%.

[0038] The trained classification model is converted into a code format that can be run on edge devices. The model is then deployed to edge nodes, and input data preprocessing is configured to extract the time and frequency domain features of the vibration signal. Model inference is then executed to classify the device's operating status in real time. Finally, fault alarm information is generated based on the classification results, and the diagnosis results are uploaded to the cloud server. This step enables real-time fault diagnosis through edge computing, reducing data transmission delays and cloud computing pressure, and improving the system's response speed and reliability.

[0039] Preferably, in step S1, when constructing the vibration model, the vibration frequency parameter is introduced , quality parameters , stiffness parameter , damping parameter , establish the following vibration model formula: ,in, Indicates that industrial equipment is at The acceleration of Indicates that industrial equipment is at speed, Indicates that industrial equipment is at The displacement, Indicates the effect on industrial equipment at time The external excitation force is determined by the quality parameters , stiffness parameters , damping parameters and external incentives Dynamic adjustment to simulate the vibration state of industrial equipment under different working conditions.

[0040] Specifically, a parametric vibration model is constructed by introducing vibration frequency parameters, mass parameters, stiffness parameters, and damping parameters. The vibration frequency parameter range is 10Hz-10kHz, covering the common vibration frequencies of industrial equipment; the mass parameter can be dynamically adjusted between 50kg-500kg according to the equipment load; the stiffness parameter reflects the elastic characteristics of the structure and has a value range of 10 4 N / m-10 7 N / m; the damping parameter reflects energy dissipation capacity and ranges from 10N・s / m to 1000N・s / m. During implementation, initial parameters are calibrated through finite element analysis or experimental data. Simulink parameter management tools are then used to adjust parameter values ​​in real time to simulate vibration responses under conditions such as load changes and component wear. This approach is significant in that the dynamic coupling between parameters allows for accurate reproduction of the equipment's true vibration characteristics, providing more physically meaningful input variables for simulation data generation and edge model training.

[0041] Preferably, in step S2, in the process of converting the simulation model into a format available for edge computing, the model compression ratio parameter , memory usage parameters , establish the following conversion optimization formula:

[0042]

[0043] in, Indicates the size of the model after conversion suitable for edge computing. Indicates the size of the original simulation model in Simulink, Memor Indicates the memory occupied by the model in the edge computing device. Indicates that the original simulation model occupies memory in the Simulink environment, and the model compression ratio parameter The model size can be adjusted to meet the storage and operation requirements of the edge computing device according to the memory usage parameters. Evaluate the memory resource usage of the edge computing device by the converted model.

[0044] Specifically, the model conversion process introduces model compression ratio parameters (ranging from 0.1 to 0.5) and memory usage parameters (in MB), achieving model lightweighting through quantization-aware training and architectural optimization. During implementation, the 32-bit floating-point parameters of the Simulink simulation model are first compressed into 8-bit integers, with a compression ratio controlled between 0.2 and 0.3, reducing the model size from 5MB to under 1MB. Memory mapping technology is then used to optimize edge device storage allocation, ensuring a 60%-80% reduction in memory usage compared to the original model. This addresses the issue of traditional simulation models being bulky and prohibitive for deployment on edge nodes, enabling the model to maintain feature extraction accuracy exceeding 95% on low-resource devices such as the STM32 (64KB to 256KB of memory).

[0045] Preferably, in step S3, when using a three-axis acceleration sensor to collect vibration signals of industrial equipment, the sensor sampling frequency parameter , sensor sensitivity parameters , establish the following signal acquisition model: ,in, Indicates the voltage value of the vibration signal collected by the sensor. Indicates the actual vibration acceleration of industrial equipment by adjusting the sensor sensitivity parameters and sampling frequency parameters , optimize the sensor's collection effect on industrial equipment vibration signals, and ensure that the collected signals can accurately reflect the vibration status of industrial equipment.

[0046] Specifically, sampling frequency parameters (5kHz-50kHz) and sensitivity parameters (0.001g-0.1g) are introduced into the sensor acquisition process to enable dynamic adjustment of data collection strategies. During implementation, edge node preprocessing logic determines the device's operating status: during the startup phase (vibration frequency <100Hz), a 5kHz sampling frequency and 0.1g sensitivity are used to reduce data redundancy; during stable operation (vibration frequency >1kHz), the sampling frequency is increased to 50kHz and 0.001g sensitivity to capture high-frequency fault characteristics. Combined with DMA technology (transmission rate 1MB / s), the CPU load is reduced to below 30%, ensuring real-time data transmission at a 100kHz sampling rate and providing high-fidelity raw signals for subsequent data fusion.

[0047] Preferably, in step S4, when constructing the data fusion module to mix the simulation data with the measured data, the data mixing ratio parameter is introduced. , establish the following data fusion formula: ,in, Represents the generated training set data, represents the simulation data, Represents the measured data, by adjusting the data mixing ratio parameters , control the proportion of simulation data and measured data in the training set to achieve the best training effect.

[0048] Specifically, a data mixing ratio parameter (ranging from 0.3 to 0.7) is introduced into the data fusion module to dynamically adjust the mixing ratio of simulated and measured data based on the equipment's operating cycle. During implementation, the ratio parameter is set to 0.7 during the initial equipment commissioning period (failure probability <5%), expanding the sample space with 70% simulated data. During mid-operation (failure probability 5%-20%), the ratio is adjusted to 0.5 to balance data diversity and authenticity. During the aging phase (failure probability >20%), the ratio is set to 0.3, reinforcing the learning of real-world failure modes with 70% measured data. Through the weight configuration functionality of the Simulink data fusion module, adaptive generation of training sets at different stages is achieved, improving the detection accuracy of the model throughout its lifecycle.

[0049] Preferably, in step S5, in the process of generating a model suitable for edge deployment, the number of model parameters is combined with the number of model parameters. , model computational complexity parameter , establish the following model optimization formula: ,in, Indicates the first parameters, Indicates the The time complexity of the parameters in the calculation process is calculated by the number of model parameters. The control of model computational complexity parameters Optimize the edge computing equipment to generate a lightweight model that meets the computing power and operating efficiency requirements.

[0050] Specifically, the number of parameters (<100KB) and computational complexity (operations per second <10 6 The model size is controlled by pruning the network at a rate of 70%-90% and the number of quantization bits (8 bits). For microcontrollers like the STM32, the number of convolutional neural network layers is limited to 5, the pruning rate is set to 90%, the parameter size is compressed to less than 50KB, and the computational complexity is controlled at 100,000 operations per second. For edge computing platforms like Jetson, a 10-layer network is allowed, a pruning rate of 70%, a parameter size of 100KB, and computational complexity is increased to 1 million operations per second. This is crucial for optimizing models based on hardware computing power, ensuring real-time inference latency of less than 10ms across all edge nodes.

[0051] Preferably, in step S6, when performing joint training, the training iteration number parameter is introduced , the learning rate parameter , establish the following training optimization formula: ,in, Indicates in The model parameters at the training iteration, Indicates in The model parameters at the training iteration, Indicates in The loss function for the training iteration Model parameters The gradient of the training iteration number parameter is adjusted by and the learning rate parameter , optimize the joint training process and improve the training effect and convergence speed of the model.

[0052] Specifically, parameters for the number of training iterations (10-100) and the learning rate (0.001-0.01) are introduced into the hybrid training process, dynamically adjusting the training strategy based on the equipment failure risk level. During implementation, when the RMS value of the equipment vibration signal is <0.5g (normal state), the number of iterations is set to 10 and the learning rate is set to 0.001 to maintain model stability. When an abnormal peak (RMS value >1.5g) is detected, emergency training is automatically triggered, increasing the number of iterations to 100 and the learning rate to 0.01 to rapidly update the fault signature parameters. Combined with a federated learning mechanism, multiple edge nodes can collaboratively optimize model parameters, improving global detection accuracy while protecting data privacy.

[0053] Preferably, in step S7, when fault classification is performed at the edge node, the vibration amplitude fluctuation coefficient is introduced , vibration frequency deviation , and combined with the classification decision threshold parameter , establish the following fault classification formula:

[0054] in, Indicates the fault classification results of industrial equipment; It is the classification probability value of the vibration amplitude of industrial equipment by the lightweight classification model, reflecting the possibility that the vibration amplitude is within the normal range; It is the classification probability value of the vibration frequency of industrial equipment by the lightweight classification model, reflecting the possibility that the vibration frequency conforms to the normal state; the vibration amplitude fluctuation coefficient Used to measure the weight of vibration amplitude in fault judgment, vibration frequency deviation Used to measure the weight of vibration frequency in fault judgment; by adjusting as well as , realizing fault classification of industrial equipment based on the comprehensive characteristics of vibration amplitude and frequency.

[0055] Specifically, the fault classification process incorporates vibration amplitude fluctuation coefficient (threshold 0.1-0.5), vibration frequency deviation (threshold 5%-20%), and classification decision threshold parameters (0.5-0.9), establishing a multi-parameter decision logic. During implementation, the amplitude fluctuation coefficient (the deviation between the current amplitude and the historical mean) and frequency deviation (the difference between the current main frequency and the device's natural frequency) of the real-time vibration signal are calculated. If both exceed the threshold and the combined value is greater than the classification decision threshold (e.g., 0.7), the fault is considered abnormal. If only a single parameter exceeds the limit, continuous monitoring is initiated within a 10-second window to avoid false alarms. This method upgrades traditional single-probability judgment to a multi-dimensional feature fusion decision, reducing the false alarm rate from 5% to below 1%.

[0056] Preferably, in the entire method process, the operating environment temperature parameters of the industrial equipment are combined , the temperature parameter As an influencing factor, it is introduced into the vibration model of step S1, and the expanded vibration model formula is established:

[0057] in, Indicates the temperature The quality parameters of industrial equipment are as follows, Indicates the temperature The stiffness parameters of the following industrial equipment, Indicates the temperature The damping parameters of the industrial equipment are as follows: Indicates the temperature and time The external excitation force acting on industrial equipment, Indicates the temperature and time Acceleration of industrial equipment, Indicates the temperature and time Lower the speed of industrial equipment, Indicates the temperature and time The displacement of industrial equipment under different temperature environments is simulated through this extended model.

[0058] Specifically, a temperature parameter (-20°C to 80°C) is introduced into the vibration model to establish a mapping relationship between temperature and mass, stiffness, and damping parameters. During implementation, the surface temperature of the equipment is collected in real time through thermocouples, and the thermal expansion coefficient of the material (such as 1.2×10⁻ for steel) is used to calculate the temperature. 5 The extended model adjusts the mass matrix based on the temperature coefficient of the elastic modulus (e.g., -0.002 / °C for steel) (mass increases by 0.1%-0.5% for every 10°C increase in temperature). The stiffness matrix is ​​adjusted based on the temperature coefficient of the elastic modulus (e.g., -0.002 / °C for steel) (stiffness decreases by 1%-3% for every 10°C increase in temperature). Furthermore, the damping matrix is ​​modified using viscous damping (damping increases by 5%-10% for every 10°C increase in temperature). This extended model enables simulation data to accurately reflect the vibration characteristics of equipment in both high and low temperature environments, reducing the prediction error caused by temperature drift from 15% to less than 3%.

[0059] Preferably, in the entire method process, the load change parameters of the industrial equipment are combined , the load change parameter As an influencing factor, it is introduced into the vibration model of step S1 and the data fusion process of step S4, and the formula is established in the vibration model:

[0060] in, Indicates that the load The quality parameters of industrial equipment are as follows, Indicates that the load The stiffness parameters of the following industrial equipment, Indicates that the load The damping parameters of the industrial equipment are as follows: Indicates that the load and time The external excitation force acting on industrial equipment, Indicates that the load and time Acceleration of industrial equipment, Indicates that the load and time Lower the speed of industrial equipment, Indicates that the load and time The displacement of industrial equipment under load; in the process of data fusion, according to the load change parameters Adjust the data mixing ratio parameters , establish the formula:

[0061] in, and and load The relevant weight coefficient is obtained by combining the load change parameters , optimize the vibration model and data fusion process, and improve the adaptability of the vibration detection method to different load conditions.

[0062] Specifically, a load variation parameter (0%-120% of rated load) is introduced in steps S1 and S4 to achieve dual adaptive adjustment of the vibration model and data fusion. During model construction, the moment of inertia (for every 10% increase in load, the moment of inertia increases by 2%-5%) and contact stiffness (for every 10% increase in load, the stiffness increases by 3%-8%) are dynamically adjusted based on the load parameters to simulate the impact of load changes on vibration characteristics. During data fusion, the ratio of simulated to measured data is adjusted using load-related weighting coefficients (ω1=0.6-0.3, ω2=0.4-0.7). Under high-load conditions (>80% of rated load), the weight of measured data is increased to 70%, while under low-load conditions (<30% of rated load), 60% of simulated data is used to supplement rare load state samples. This technology improves the system's detection accuracy by 12%-15% across load conditions, significantly enhancing generalization capabilities.

[0063] The present invention addresses the deficiencies of traditional technologies in model construction and data processing, and achieves precise simulation and sample expansion through multi-dimensional modeling and data fusion technology. This method constructs a four-degree-of-freedom or six-degree-of-freedom vibration model of industrial equipment based on the Simulink environment. By dynamically adjusting the mass matrix, stiffness matrix, and damping matrix, combined with real-time feedback of environmental parameters such as temperature and load, it can accurately characterize the vibration characteristics of the equipment under different working conditions, breaking through the limitations of traditional fixed parameter models that are insufficient to simulate complex working conditions. At the same time, through the data fusion module, simulation data covering various fault modes such as bearing wear and gear failure are mixed with measured data in a specific proportion to generate a training set, which solves the problem of insufficient sample diversity of a single data source, enables the model to learn more comprehensive fault characteristics, and significantly improves the detection accuracy of early faults and rare faults.

[0064] In response to the challenge of poor compatibility of traditional technology equipment, this method achieves a balance between model size and computational efficiency through lightweight model conversion and edge deployment optimization technology. Simulink Coder generates C / C++ code adapted to edge hardware such as STM32 and Jetson, and utilizes model compression technologies such as quantization-aware training and pruning to control the number of model parameters to less than 100KB, significantly reducing the storage and computing pressure on edge devices. At the same time, a hybrid training process is designed, employing transfer learning and federated learning strategies to enable edge nodes to use local real-time data for online incremental learning, avoiding the latency issues associated with traditional centralized training's reliance on the cloud. This ensures that the model continuously adapts to changes in operating conditions during equipment operation, improving real-time detection. The lightweight classification model deployed on the edge node can directly analyze the vibration signals collected by the three-axis accelerometer in real time, reducing data transmission delays while achieving millisecond-level response for fault classification.

[0065] In addition, this method forms an integrated solution for industrial vibration detection through full-process tool chain integration and multi-parameter decision-making mechanism. Based on the unified modeling environment of Simulink, it realizes seamless connection of the entire process from physical model construction, data fusion, model compression to edge deployment, avoiding the data format conversion and compatibility issues brought about by traditional multi-tool chain collaboration. In the fault classification link, by integrating the decision logic of multi-dimensional parameters such as vibration amplitude and frequency, combined with the real-time computing power of edge nodes, it can quickly distinguish the normal operating status of the equipment from the abnormal vibration mode, and avoid missed detection or false alarm caused by misjudgment of a single parameter. While ensuring detection accuracy, the overall solution reduces model deployment delay by more than 80% and computing resource usage by more than 60%, significantly improving the intelligence level of industrial equipment vibration detection, and providing efficient and reliable technical support for predictive maintenance in intelligent manufacturing.

[0066] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A vibration detection method based on Simulink physical modeling to build an edge computing model, characterized by: The steps include: Step S1: In the Simulink environment, the Newton-Euler equation and the Lagrange equation are used to construct four-degree-of-freedom and six-degree-of-freedom vibration models for industrial equipment. The mass matrix, stiffness matrix, and damping matrix are dynamically adjusted through MATLAB scripts and Simulink parameter management tools to simulate the complex vibration characteristics of industrial equipment under different working conditions such as load changes, equipment wear, and temperature drift. Step S2: Design a lightweight model export process in Simulink, convert the constructed vibration simulation model into a format suitable for edge computing, and use Simulink Coder or MATLAB Coder to generate C / C++ code that is compatible with STM32 and Jetson edge hardware. Step S3: deploy a three-axis acceleration sensor to collect vibration signals of industrial equipment. The collected signals include acceleration, displacement, and angular velocity. DMA technology is used to accelerate the data transmission process and reduce the CPU load. Step S4: Construct a data fusion module in Simulink to mix the simulation data including different fault modes generated in step S1 with the measured data collected in step S3 according to a preset ratio to generate a training set; Step S5: Use the ONNX format to unify the model interface in Simulink, integrate the model compression tool, and generate a model with less than 100KB parameters and suitable for edge deployment through quantization-aware training and pruning operations. This model supports TensorFlowLite, PyTorchMobile, and NanoEdgeAI frameworks and is compatible with STM32 and JetsonNano hardware. Step S6: Design a hybrid training process in Simulink, input the training set generated in step S4 into the edge model obtained in step S5, conduct joint training using transfer learning or federated learning strategies, implement online learning on edge nodes, and update model parameters through incremental learning; Step S7: Deploy a lightweight classification model on the edge node, use Simulink's code generation tool to convert the reasoning logic into C / C++ code, and perform fault classification on the processed vibration signal.

2. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S1, when constructing the vibration model, the vibration frequency parameter is introduced , quality parameters , stiffness parameter , damping parameter , establish the following vibration model formula: ,in, Indicates that industrial equipment is at The acceleration of Indicates that industrial equipment is at speed, Indicates that industrial equipment is at The displacement, Indicates the effect on industrial equipment at time The external excitation force is determined by the quality parameters , stiffness parameters , damping parameters and external incentives Dynamic adjustment to simulate the vibration state of industrial equipment under different working conditions.

3. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S2, in the process of converting the simulation model into a format available for edge computing, the model compression ratio parameter is combined , memory usage parameters , establish the following conversion optimization formula: in, Indicates the size of the model after conversion suitable for edge computing. Indicates the size of the original simulation model in Simulink, Memor Indicates the memory occupied by the model in the edge computing device. Indicates that the original simulation model occupies memory in the Simulink environment, and the model compression ratio parameter The model size is adjusted to meet the storage and operation requirements of the edge computing device based on the memory usage parameters. Evaluate the memory resource usage of the edge computing device by the converted model.

4. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S3, when using a triaxial acceleration sensor to collect vibration signals of industrial equipment, the sensor sampling frequency parameter , sensor sensitivity parameters , establish the following signal acquisition model: ,in, Indicates the voltage value of the vibration signal collected by the sensor. Indicates the actual vibration acceleration of industrial equipment by adjusting the sensor sensitivity parameters and sampling frequency parameters , optimize the sensor's collection effect on industrial equipment vibration signals.

5. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S4, when constructing the data fusion module to mix the simulation data with the measured data, the data mixing ratio parameter is introduced. , establish the following data fusion formula: ,in, Represents the generated training set data, represents the simulation data, Represents the measured data, by adjusting the data mixing ratio parameters , control the proportion of simulation data and measured data in the training set.

6. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S5, in the process of generating a model suitable for edge deployment, the number of model parameters is combined , model computational complexity parameter , establish the following model optimization formula: ,in, Indicates the first parameters, Indicates the The time complexity of the parameters in the calculation process is calculated by the number of model parameters. The control of model computational complexity parameters Optimize the edge computing equipment to generate a lightweight model that meets the computing power and operating efficiency requirements.

7. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S6, when performing joint training, the training iteration number parameter is introduced. , the learning rate parameter , establish the following training optimization formula: ,in, Indicates in The model parameters at the training iteration, Indicates in The model parameters at the training iteration, Indicates in The loss function for the training iteration Model parameters The gradient of the training iterations parameter is adjusted by and the learning rate parameter , optimize the joint training process.

8. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In step S7, when fault classification is performed at the edge node, the vibration amplitude fluctuation coefficient is introduced. , vibration frequency deviation , and combined with the classification decision threshold parameter , establish the following fault classification formula: in, Indicates the fault classification results of industrial equipment; It is the classification probability value of the vibration amplitude of industrial equipment by the lightweight classification model, reflecting the possibility that the vibration amplitude is within the normal range; It is the classification probability value of the vibration frequency of industrial equipment by the lightweight classification model, reflecting the possibility that the vibration frequency conforms to the normal state; the vibration amplitude fluctuation coefficient Used to measure the weight of vibration amplitude in fault judgment, vibration frequency deviation Used to measure the weight of vibration frequency in fault judgment; by adjusting as well as , fault classification of industrial equipment is performed based on the comprehensive characteristics of vibration amplitude and frequency.

9. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In the whole process, the operating environment temperature parameters of industrial equipment are combined , the temperature parameter As an influencing factor, it is introduced into the vibration model of step S1 to establish the expanded vibration model formula: in, Indicates the temperature The quality parameters of the following industrial equipment, Indicates the temperature The stiffness parameters of the following industrial equipment, Indicates the temperature The damping parameters of industrial equipment are as follows: Indicates the temperature and time The external excitation force acting on industrial equipment, Indicates the temperature and time Acceleration of industrial equipment, Indicates the temperature and time Lower the speed of industrial equipment, Indicates the temperature and time The displacement of industrial equipment under different temperature environments is simulated through this extended model.

10. The vibration detection method based on Simulink physical modeling to build an edge computing model according to claim 1, characterized in that: In the whole process of the method, the load change parameters of industrial equipment are combined , the load change parameter As an influencing factor, it is introduced into the vibration model of step S1 and the data fusion process of step S4, and the formula is established in the vibration model: in, Indicates that the load The quality parameters of the following industrial equipment, Indicates that the load The stiffness parameters of the following industrial equipment, Indicates that the load The damping parameters of industrial equipment are as follows: Indicates that the load and time The external excitation force acting on industrial equipment, Indicates that the load and time Acceleration of industrial equipment, Indicates that the load and time Lower the speed of industrial equipment, Indicates that the load and time The displacement of industrial equipment under load; in the process of data fusion, according to the load change parameters Adjust data mixing ratio parameters , establish the formula: in, and and load The relevant weight coefficient is obtained by combining the load change parameters , optimize the vibration model and data fusion process.

Images