Lightweight method, system and instrument for fault monitoring and diagnosis of mechanical equipment component
By designing a lightweight convolutional neural network and a portable fault monitoring and diagnosis system, the problem of high cost and low efficiency of industrial equipment fault diagnosis is solved, and fast and accurate fault diagnosis is achieved, reducing memory usage and running time.
Patent Information
- Application Number
- CN202411439939.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-15
- Publication Date
- 2025-03-04
AI Technical Summary
In the prior art, industrial equipment fault diagnosis is high, low efficiency, complex operation, and deep learning models have large calculations and large parameters, making it difficult to deploy on mobile devices and embedded devices.
A lightweight convolutional neural network is designed, combining signal acquisition module, model training module, intelligent diagnosis module, data and model storage module and human-computer interaction module to build a portable fault monitoring and diagnosis system and instrument, real-time acquisition and rapid diagnosis of vibration signals of mechanical equipment.
It achieves the shortening of running time, memory usage, and the reduction of fault diagnosis costs while ensuring diagnostic accuracy, improving diagnostic efficiency and operation ease.
Smart Images
Figure CN119557758A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mechanical equipment component fault diagnosis, and relates to a method and system for equipment fault monitoring and diagnosis, specifically to a portable fault monitoring and diagnosis instrument, and particularly to a fault diagnosis method based on a lightweight convolutional neural network. Background Art
[0002] With the continuous improvement of industrial automation, a large number of mechanical equipment and systems are widely used in industrial production. Once the core components fail, it will seriously affect production efficiency and safety. Traditional fault diagnosis methods usually rely on the experience and knowledge of professionals, and have problems such as low diagnostic efficiency and high cost. With the acceleration of industrial modernization and informatization and the rapid development of technologies such as machine learning, deep learning, as the hottest branch of machine learning, is widely used in the field of fault diagnosis due to its powerful data processing and feature learning capabilities and architectural innovation. As the model structure of deep learning continues to deepen, the accuracy of fault diagnosis has increased accordingly. However, the number of parameters, computing power, and resources required for training have also begun to increase significantly, making it difficult to deploy deep learning models on mobile devices and embedded devices, resulting in their inability to be directly applied to fault diagnosis of actual industrial equipment.
[0003] Therefore, it is necessary to design a lightweight model with low computational complexity, light model structure, and strong portability, so that it can quickly and effectively identify equipment faults with high accuracy; at the same time, in view of the high cost, low efficiency, and complex operation of industrial equipment fault diagnosis, it is necessary to develop a portable fault monitoring and diagnosis instrument for simple and fast fault diagnosis at the production site. To this end, we propose a lightweight method and system for monitoring and diagnosing mechanical equipment components, which provides guarantee for industrial production and promotes the intelligent and automated development of industrial production, which has important practical significance. Summary of the invention
[0004] The purpose of the present invention is to provide a lightweight method, system and instrument for monitoring and diagnosing mechanical equipment component faults. This method addresses the problems of high cost, low efficiency and complex operation of industrial equipment fault diagnosis mentioned in the above background technology, and realizes a mechanical equipment component fault monitoring and diagnosis system and instrument with model training functions and real-time collection and real-time diagnosis of mechanical equipment vibration signals. At the same time, a lightweight convolutional neural network with low computational complexity, lightweight model structure and strong portability is proposed. The network can achieve faster diagnosis speed and smaller memory usage while ensuring diagnostic accuracy.
[0005] The system is composed of a signal acquisition module, a model training module, an intelligent diagnosis module, a data and model storage module, and a human-computer interaction module. The vibration signal collected by the sensor is input into the constructed lightweight convolutional neural network after data preprocessing. The trained model can quickly diagnose it, thereby outputting the current health status of the component. This series of actions are all realized by the human-computer interaction module. The user operates the instrument and the instrument provides feedback to the user. The present invention is intended to be applied in actual industrial production, and can diagnose the health status of mechanical equipment components in real time, and assist equipment maintenance personnel in fault monitoring.
[0006] In order to achieve the above object, the present invention provides the following technical solutions:
[0007] In a first aspect, the present invention provides a lightweight method for monitoring and diagnosing mechanical equipment component faults, comprising:
[0008] Obtain equipment vibration signals under different health conditions of mechanical equipment components;
[0009] The vibration signal is imported into the processor for preprocessing as the input of the model: In order to eliminate the order of magnitude difference between the data of each dimension and avoid the network prediction error caused by the large order of magnitude difference between the input and output data, the acquired raw data is normalized according to formula (1) to obtain data x′ in the range of [0,1].
[0010]
[0011] Where x is the original data, x′ is the normalized data, min(x) and max(x) are the minimum and maximum values of the original data respectively; overlapping sampling is adopted, 4096 sampling points are intercepted as a group as a single sample sequence each time, and each one-dimensional original signal is converted into a two-dimensional grayscale image sample of size 64×64. After traversing the entire signal data, samples are generated as input for model training;
[0012] Constructing a lightweight convolutional neural network: the lightweight convolutional neural network includes a feature extraction module, a feature enhancement module and a classification module; the feature extraction module is composed of a standard convolution and a depth-separable convolution, the feature enhancement module is composed of two inverted residual structures and a point convolution, and the classification module is composed of a fully connected layer and a softmax classifier;
[0013] Update model parameters and conduct model training: Divide the prepared samples into training sets and test sets, input them into the network model for training, and perform back propagation according to the cross entropy loss value obtained by formula (2) after each round of training, and continuously adjust the training parameters. When the iteration reaches the set final number of iterations, the model training is completed.
[0014]
[0015] Among them, C is the number of categories, β is the input data, and β i represents the i-th element of the true label, γ i It represents the probability that the model predicts that β belongs to the i-th category.
[0016] Furthermore, the construction of the lightweight convolutional neural network specifically includes:
[0017] The feature extraction module first uses a standard convolutional layer with a convolution kernel size of 3×3, a padding of 1, and a stride of 2 to perform an initial convolution operation on the input data to extract low-level feature information; after the convolution operation, the output data is input into a batch normalization layer to normalize each feature channel so that its mean is 0 and its variance is 1; after the batch normalization layer, the ReLU activation function is applied to set all negative values to zero and keep all positive values unchanged. The process is shown in formula (3):
[0018] y1=max{H(f 3 (x)),0} (3)
[0019] Among them, x represents the input data, f 3 () represents a 3×3 standard convolution operation, H() represents a batch normalization operation, the ReLU activation function formula is max{x,0}, and y1 represents the output feature after the standard convolution. The extracted feature information is input into a deep convolution layer consisting of a convolution kernel size of 3×3, a padding of 1, a stride of 1, and a group number of 32, and a point convolution layer with a 1×1 convolution kernel, no padding, and a stride of 1 for further feature extraction. The calculation process is shown in formula (4). This feature extraction module significantly reduces the number of parameters and computational costs while effectively extracting features.
[0020]
[0021] Among them, y1 represents the input feature, represents a 3×3 depthwise convolution operation, represents a 1×1 point convolution operation, and y2 represents the output features of the feature extraction module.
[0022] The feature enhancement module is designed to maintain computational efficiency while learning richer feature representations for further feature information extraction. The output features of the feature extraction module first pass through the first inverted residual structure, in which the features first pass through a lightweight dilated convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1 to increase the receptive field without adding additional parameters; then, each input channel is independently convolved through a convolution kernel size of 3×3, padding of 1, stride of 2, and grouping number of 48; finally, a point convolution layer with a 1×1 convolution kernel, no padding, and stride of 1 is used to mix channel information and reduce the number of channels. The calculation process is shown in formula (5):
[0023]
[0024] Among them, y2 represents the input feature, Represents a 1×1 lightweight dilated convolution operation, and y3 represents the output feature of the first inverted residual structure of the feature enhancement module. Similarly, the second inverted residual structure is composed of a 1×1 convolution kernel, no padding, and a lightweight dilated convolution layer with a stride of 1, a deep convolution layer with a convolution kernel size of 5×5, a padding of 2, a stride of 2, and a grouping number of 72, and a point convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1. This is to learn deep features while further optimizing the computational efficiency. The calculation process is shown in formula (6):
[0025]
[0026] Among them, y3 represents the input feature, represents a 5×5 depth convolution operation, and y4 represents the output feature of the second inverted residual structure of the feature enhancement module. The output features of the two inverted residual structures are input into a point convolution layer with a convolution kernel size of 1×1 for dimensionality increase. The calculation process is shown in formula (7).
[0027]
[0028] Among them, y4 represents the input feature, and y5 represents the output feature of the feature enhancement module.
[0029] The classification module classifies the fault type by a fully connected layer and a classifier, and outputs a classification result.
[0030] In a second aspect, the present invention provides a mechanical equipment component fault monitoring and diagnosis system and instrument, comprising:
[0031] The signal acquisition module is used to collect vibration signals of mechanical equipment in real time and consists of a sensor, an amplifier and an analog-to-digital converter;
[0032] The model training module is used to adjust the hyperparameters of the lightweight model and train the model. The lightweight model includes a depthwise separable module, an inverted residual module and a classification module; the depthwise separable module is composed of a standard convolution and a depthwise separable convolution, the inverted residual module is composed of two inverted residual structures and a point convolution, and the classification module is composed of a fully connected layer and a softmax classifier;
[0033] Intelligent diagnostic module, when inputting the vibration signal of mechanical equipment components, can quickly determine the health status of the current components;
[0034] The data and model storage module is used to collect and save all historical and real-time vibration signal data acquired by the signal acquisition module, and store untrained and trained lightweight models respectively;
[0035] The human-computer interaction module is used to issue instructions to control the signal acquisition module, model training module and intelligent diagnosis module to perform corresponding functions and receive feedback.
[0036] Furthermore, the signal acquisition module uses a sensor to collect vibration signals in real time, converts the vibration signals into voltage signals in the millivolt range, and then inputs the signals into analog-to-digital converters, which convert the signals into digital signals for subsequent processing.
[0037] Furthermore, the sensor model is a piezoelectric acceleration sensor SA12SC with high sensitivity, high resolution and no need for external power supply, so as to realize the collection of vibration signals.
[0038] Furthermore, the analog-to-digital converter is a Raspberry Pi high-precision AD / DA expansion board, which has an onboard 24-bit resolution analog-to-digital conversion chip ADS1256 to achieve conversion of analog signals to digital signals.
[0039] Furthermore, the model training module and the intelligent diagnosis module select Raspberry Pi 4B as the core processor. The Raspberry Pi 4B has a built-in Broadcom BCM2711 chip and is equipped with a high-performance 64-bit quad-core CortexA72 processor.
[0040] Furthermore, the hyperparameter adjustment includes: the number of training iterations, the batch size and the learning rate.
[0041] Furthermore, the human-computer interaction module uses a 5-inch, 800×480 resolution capacitive touch TFT screen connected to a processor to develop a portable fault monitoring and diagnostic instrument with an integrated graphical user interface, allowing users to issue instructions through the touch screen to control the work of the signal acquisition module, model training module and intelligent diagnosis module, and at the same time receive feedback from each module and display it on the screen.
[0042] Furthermore, the human-computer interaction module and the processor are connected via an HDMI interface.
[0043] Furthermore, the housing of the portable fault monitoring and diagnostic instrument is modeled using SolidWorks software, the design is converted into a physical object using an FDM printer, and PrusaSlicer is selected as the 3D printing slicing software.
[0044] Compared with the prior art, the present invention provides a lightweight method, system and instrument for monitoring and diagnosing mechanical equipment component faults, which has the following advantages and technical effects: The present invention can realize the real-time collection and real-time diagnosis of mechanical equipment vibrations through the mechanical equipment component fault monitoring and diagnosis system, making the mechanical equipment component fault diagnosis cost lower, more efficient and easier to operate; wherein, the lightweight convolutional neural network embedded in the system has low computational complexity, light model structure and strong portability, and the network can shorten the running time and reduce memory usage while ensuring the diagnostic accuracy. The method and system can assist equipment maintenance personnel in timely discovering mechanical equipment component faults, and provide a technical basis for the development of small and flexible fault diagnosis instruments for industrial production site equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0046] Figure 1 A flowchart of a mechanical equipment component fault monitoring and diagnosis system provided by the present invention;
[0047] Figure 2 Diagnostic process and model structure diagram of the lightweight method for monitoring and diagnosing mechanical equipment component faults provided by the present invention;
[0048] Figure 3 A hardware connection diagram of the portable fault monitoring and diagnostic instrument provided by the present invention;
[0049] Figure 4 The confusion matrix of the constructed lightweight convolutional neural network under bearing data;
[0050] Figure 5 The classification accuracy and loss curve of the constructed lightweight convolutional neural network. DETAILED DESCRIPTION
[0051] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0052] Example 1
[0053] The present invention provides a lightweight method, system and instrument for monitoring and diagnosing mechanical equipment component faults. Figure 1 As shown in the figure, the mechanical equipment component fault monitoring and diagnosis system is mainly composed of a signal acquisition module, a model training module, an intelligent diagnosis module, a data and model storage module, and a human-computer interaction module. The vibration signal collected by the sensor is input into the constructed lightweight convolutional neural network after data preprocessing. The trained model can quickly diagnose it and output the current health status of the component. This series of actions are realized through the human-computer interaction module to realize the user's operation of the instrument and the instrument's feedback to the user.
[0054] like Figure 3 As shown in the figure, a portable fault monitoring and diagnostic instrument mainly includes a sensor, an analog-to-digital converter, a processor, a display and a shell. The piezoelectric accelerometer SA12SC is used to collect vibration signals in real time, and the vibration signals are converted into voltage signals of the millivolt level and then input into the analog-to-digital converter Raspberry Pi high-precision AD / DA expansion board, which has a 24-bit resolution analog-to-digital conversion chip ADS1256 on board to realize the conversion of analog signals to digital signals; the Raspberry Pi 4B is used as the core processor for subsequent processing, which has a built-in Broadcom BCM2711 chip and a high-performance 64-bit quad-core CortexA72 processor; a 5-inch, 800×480 resolution capacitive touch TFT screen connected to the processor is selected as the display, and the display and processor are connected through an HDMI interface, and a graphical user interface is integrated at the same time, so that users can issue instructions to control the work of the signal acquisition module, model training module and intelligent diagnosis module by touching the screen, and receive feedback from each module and display it on the screen; finally, the shell is modeled using SolidWorks software, and the design is converted into a physical object through an FDM printer, and PrusaSlicer is selected as the 3D printing slice software.
[0055] The detection object in this embodiment is a bearing, and the specific use method of the system and the instrument is as follows:
[0056] Step 1: In the signal acquisition module, obtain the vibration signal of the bearing fault equipment and preprocess it to obtain a grayscale image sample.
[0057] Specifically, there are four states of the bearing components, namely healthy, inner race fault, rolling element fault and outer race fault. First, in order to eliminate the order of magnitude difference between the data of each dimension and avoid the network prediction error caused by the large order of magnitude difference between the input and output data, the acquired raw data is normalized according to formula (1) to obtain data x′ in the range of [0,1].
[0058]
[0059] Among them, x is the original data, x′ is the normalized data, min(x) and max(x) are the minimum and maximum values of the original data respectively; secondly, the first 120,000 sampling points of the vibration signal in each state are selected, the overlap is set to 2926, 4096 sampling points are intercepted as a group as a single sample sequence each time, and each one-dimensional original signal is converted into a two-dimensional grayscale image sample with a size of 64×64. After traversing the entire signal data, 100 samples are generated for each state, and a total of 400 samples for the four types of bearing states. Samples of the same type are stored in the same path, and "0, 1, 2, 3" are used to represent the fault type as the label name as the input for model training.
[0060] Step 2: In the model training module, adjust the lightweight model hyperparameters and train the model to a satisfactory accuracy.
[0061] The lightweight convolutional neural network in this embodiment includes a feature extraction module, a feature enhancement module and a classification module. The feature extraction module is composed of a standard convolution and a depth-separable convolution, the feature enhancement module is composed of two inverted residual structures and a point convolution, and the classification module is composed of a fully connected layer and a softmax classifier. The network diagnosis process and the lightweight network model structure are shown in Figure 2 Specifically, the training of the network includes the following steps:
[0062] (1) Divide the samples prepared in step 1 into training set and test set in a ratio of 7:3, and then input them into the network for training;
[0063] (2) First, a standard convolutional layer with a kernel size of 3×3, padding of 1, and stride of 2 is used to perform an initial convolution operation on the input data to extract low-level feature information. After the convolution operation, the output data is input into a batch normalization layer to normalize each feature channel so that its mean is 0 and its variance is 1. After the batch normalization layer, the ReLU activation function is applied to set all negative values to zero and keep all positive values unchanged. The process is shown in formula (3):
[0064] y1=max{H(f 3 (x)),0} (3)
[0065] Among them, x represents the input data, f 3 () represents a 3×3 standard convolution operation, H() represents a batch normalization operation, the ReLU activation function formula is max{x,0}, and y1 represents the output feature after standard convolution;
[0066] The extracted feature information is input into a depthwise separable convolution layer consisting of a convolution kernel size of 3×3, padding of 1, stride of 1, and grouping number of 32, and a point convolution layer with a 1×1 convolution kernel, no padding, and stride of 1 to further extract features. The calculation process is shown in formula (4). This feature extraction module significantly reduces the number of parameters and computational costs while effectively extracting features.
[0067]
[0068] Among them, y1 represents the input feature, represents a 3×3 depthwise convolution operation, represents a 1×1 point convolution operation, and y2 represents the output feature of the feature extraction module;
[0069] (3) In order to learn richer feature representations while maintaining computational efficiency, the feature enhancement module further extracts feature information. The output features of the feature extraction module first pass through the first inverted residual structure. In this structure, the features first pass through a lightweight dilated convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1 to increase the receptive field without adding additional parameters; then, each input channel is independently convolved through a deep convolution layer with a convolution kernel size of 3×3, padding of 1, stride of 2, and grouping number of 48; finally, the features pass through a point convolution layer with a 1×1 convolution kernel, no padding, and stride of 1 to mix channel information and reduce the number of channels. The calculation process is shown in formula (5):
[0070]
[0071] Among them, y2 represents the input feature, represents a 1×1 lightweight dilated convolution operation, and y3 represents the output feature of the first inverted residual structure of the feature enhancement module;
[0072] Similarly, the second inverted residual structure is composed of a lightweight dilated convolution layer with a 1×1 convolution kernel, no padding, and a stride of 1, a deep convolution layer with a convolution kernel size of , a padding of 2, a stride of 2, and a group number of 72, and a point convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1, which is stacked in sequence to learn deep features while further optimizing computational efficiency. The calculation process is shown in formula (6):
[0073]
[0074] Among them, y3 represents the input feature, represents a 5×5 depth convolution operation, and y4 represents the output feature of the second inverted residual structure of the feature enhancement module;
[0075] The output features of the two inverted residual structures are input into a point convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1 for dimensionality increase. The calculation process is shown in formula (7).
[0076]
[0077] Among them, y4 represents the input feature, y5 represents the output feature of the feature enhancement module, and then it is input into the fully connected layer and classifier of the classification module to classify the fault type and output the classification result.
[0078] (4) Finally, by touching the display of the portable fault monitoring and diagnostic instrument, give instructions to set the model hyperparameters for training. In this embodiment, the number of iterations is set to 50, the batch size is set to 120, and the learning rate is set to 0.001. After each round of training, back propagation is performed according to the cross entropy loss value obtained by formula (2), and the parameters of the network model are updated using the Adam optimizer. When the iteration reaches the set final number of iterations, the model training is completed, and a trained lightweight convolutional neural network is obtained.
[0079]
[0080] Wherein, C is the number of categories, which is 4 in this embodiment; β is the vector representation of the category to which the input data belongs, β i Determine whether the input data belongs to the i-th category, if so, the value is 1, otherwise it is 0; γ i It represents the probability that the model predicts that the input data belongs to the i-th category, which is obtained by the softmax classifier after the fully connected layer.
[0081] The classification effect of the lightweight convolutional neural network proposed in the present invention is represented by a confusion matrix, such as Figure 4 As shown in the figure, the vertical axis and horizontal axis represent the actual label and predicted label of the sample respectively. It can be seen that the proposed lightweight method has achieved better results in classification.
[0082] like Figure 5 As shown in the figure, it is the accuracy and loss curve of the lightweight convolutional neural network training process proposed by the present invention, and its horizontal axis represents the number of training times. An epoch refers to the complete data set being forward propagated and backward propagated once in the network. The left axis vertical axis represents the accuracy rate. The accuracy rate in the figure shows an upward trend and gradually approaches 1. There is no abnormal fluctuation in the whole training process. The right axis vertical axis represents the loss value. The loss value in the figure shows a downward trend and gradually approaches 0. There is no abnormal fluctuation in the whole training process, which proves the effectiveness of this method.
[0083] In order to illustrate the advantages of the lightweight convolutional neural network proposed in the present invention, it is compared with other classic lightweight methods, namely MobileNet V2 and ShuffleNet V2. The running time in the table refers to the time required for 50 training iterations of each model. It can be seen from Table 1 that the lightweight method proposed in the present invention has lower model calculation amount and parameter amount and shorter running time than the classic lightweight model.
[0084] Table 1. Comparison of lightweight degree of different models
[0085]
[0086] Step 3: In the signal acquisition module, collect the vibration signal of the mechanical equipment to be diagnosed, input the trained lightweight model in the intelligent diagnosis module, output the diagnosis result, and feedback it on the display through the human-computer interaction module. Specifically, the original data processing method can be found in step 1, which will not be repeated here.
[0087] The embodiment of the present invention uses a signal acquisition module, a model training module, an intelligent diagnosis module, a data and model storage module, and a human-computer interaction module to input the vibration signal collected by the sensor into the constructed lightweight convolutional neural network after data preprocessing. The trained model can quickly diagnose it, thereby outputting the current health status of the component. This series of actions are all realized by the human-computer interaction module. The user operates the instrument and the instrument provides feedback to the user. The embodiment of the present invention can assist equipment maintenance personnel to promptly discover mechanical equipment component failures, and provides a technical basis for the development of small and flexible fault diagnosis instruments for industrial production field equipment.
Claims
1. A lightweight system for monitoring and diagnosing mechanical equipment component faults, characterized in that: include: Signal acquisition module, used to collect vibration signals of mechanical equipment in real time; Model training module, used to adjust the hyperparameters of lightweight network models and train models; The lightweight network model includes a feature extraction module, a feature enhancement module and a classification module; the feature extraction module is composed of a standard convolution and a depth-separable convolution, the feature enhancement module is composed of two inverted residual structures and a point convolution, and the classification module is composed of a fully connected layer and a classifier; Intelligent diagnostic module, when inputting the vibration signal of mechanical equipment components, can determine the health status of the current components; The data and model storage module is used to collect and save all historical and real-time vibration signal data acquired by the signal acquisition module, and store untrained and trained lightweight network models respectively; The human-computer interaction module is used to issue instructions to control the signal acquisition module, model training module and intelligent diagnosis module to perform corresponding functions and receive feedback.
2. A method for applying the system as claimed in claim 1, characterized in that: include: S1: Obtain equipment vibration signals of bearing components of mechanical equipment in different health states through the signal acquisition module, including four states: healthy, inner ring fault, rolling element fault and outer ring fault, and save them to the data and model storage module; S2: Import the vibration signal in the data and model storage module into the model training module for preprocessing as the input of the model: normalize the acquired raw data, then select the first 120,000 sampling points of the vibration signal in each state, set the overlap amount to 2926, and use overlapping sampling to intercept 4096 sampling points each time as a group as a single sample sequence, and convert each one-dimensional raw signal into a two-dimensional grayscale image sample of size 64×64, and generate samples after traversing the entire signal data; S3: Building a lightweight convolutional neural network: It includes a feature extraction module, a feature enhancement module and a classification module; the feature extraction module is composed of a standard convolution and a depth-separable convolution, the feature enhancement module is composed of two inverted residual structures and a point convolution, and the classification module is composed of a fully connected layer and a classifier; The feature extraction module first uses a standard convolutional layer with a convolution kernel size of 3×3, a padding of 1, and a stride of 2 to perform an initial convolution operation on the input data to extract low-level feature information; after the convolution operation, the output data is input into a batch normalization layer to normalize each feature channel so that its mean is 0 and its variance is 1; after the batch normalization layer, the ReLU activation function is applied to set all negative values to zero and keep all positive values unchanged. The process is shown in formula (1): y1=max{H(f 3 (x)),0} (1) Among them, x represents the grayscale image sample converted from the vibration signal of the mechanical equipment as input data, f 3 () represents a 3×3 standard convolution operation, H() represents a batch normalization operation, the ReLU activation function formula is max{x,0}, and y1 represents the output feature after the standard convolution; the extracted feature information is input into a deep convolution layer consisting of a convolution kernel size of 3×3, a padding of 1, a stride of 1, and a group number of 32, and a point convolution layer with a 1×1 convolution kernel, no padding, and a stride of 1 for further feature extraction. The calculation process is shown in formula (2). This feature extraction module significantly reduces the number of parameters and computational costs while effectively extracting features; Among them, y1 represents the input feature, represents a 3×3 depthwise convolution operation, represents a 1×1 point convolution operation, and y2 represents the output feature of the feature extraction module; The feature enhancement module is designed to maintain computational efficiency while learning richer feature representations for further feature information extraction. The output features of the feature extraction module first pass through the first inverted residual structure. In this structure, the features first pass through a lightweight dilated convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1 to increase the receptive field without adding additional parameters. Then, each input channel is independently convolved through a deep convolution layer with a convolution kernel size of 3×3, padding of 1, a stride of 2, and a grouping number of 48. Finally, the features pass through a point convolution layer with a 1×1 convolution kernel, no padding, and a stride of 1 to mix channel information and reduce the number of channels. The calculation process is shown in formula (3): Among them, y2 represents the input feature, represents a 1×1 lightweight dilated convolution operation, and y3 represents the output feature of the first inverted residual structure of the feature enhancement module; similarly, the second inverted residual structure is composed of a lightweight dilated convolution layer with a 1×1 convolution kernel, no padding, and a stride of 1, a deep convolution layer with a convolution kernel size of 5×5, a padding of 2, a stride of 2, and a group number of 72, and a point convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1, which is stacked in sequence to learn deep features while further optimizing the computational efficiency; the calculation process is shown in formula (4): Among them, y3 represents the input feature, represents a 5×5 depth convolution operation, and y4 represents the output feature of the second inverted residual structure of the feature enhancement module; the output features of the two inverted residual structures are input into a point convolution layer with a convolution kernel size of 1×1, no padding, and a stride of 1 for dimensionality increase. The calculation process is shown in formula (5); Among them, y4 represents the input feature, and y5 represents the output feature of the feature enhancement module; The classification module classifies the fault type by a fully connected layer and a classifier, and outputs a classification result; S4: Update model parameters and perform model training: Divide the prepared samples into training sets and test sets, input them into the network model for training, perform back propagation according to the obtained cross entropy loss value after each round of training, and continuously adjust the training parameters; set the number of iterations to 50, the batch size to 120, and the learning rate to 0.001; when the iteration reaches the set final number of iterations, the model training is completed and saved to the data and model storage module; S5: The trained lightweight model is loaded through the intelligent diagnosis module, and the trained lightweight network is used to diagnose the vibration signal to be diagnosed that is collected in real time.
3. A mechanical equipment component fault monitoring and diagnostic instrument, used to implement the lightweight method according to claim 2, characterized in that: Includes sensor, analog-to-digital converter, processor, display and housing; The sensor uses a piezoelectric acceleration sensor SA12SC to collect vibration signals; The analog-to-digital converter uses the Raspberry Pi AD / DA expansion board, which has an onboard 24-bit resolution analog-to-digital conversion chip ADS1256 to achieve accurate conversion of voltage signals to digital signals; The processor uses Raspberry Pi 4B as the core processor of the system, which has a built-in Broadcom BCM2711 chip and a 64-bit quad-core CortexA72 processor to realize the functions of the model training module, intelligent diagnosis module, and data and model storage module; The display uses a 5-inch capacitive touch TFT screen with a resolution of 800×480, and an integrated graphical user interface, so that the user can issue instructions to control the work of the signal acquisition module, model training module and intelligent diagnosis module by touching the screen, and receive feedback from each module and display it on the screen; the display and Raspberry Pi 4B are connected through an HDMI interface; The shell was modeled using SolidWorks software, and the design was converted into a physical object using an FDM printer, with PrusaSlicer selected as the 3D printing slicing software.