Bolt pretightening force monitoring system and method based on ultrasonic waves

Abstract

The invention discloses a bolt pre-tightening force monitoring system and method based on ultrasonic waves. The system comprises an ultrasonic probe, a data acquisition and processing analyzer and the like, a probe array is uniformly mounted on the surface of a bolt, an improved depth residual network is arranged in the data acquisition and processing analyzer and can process a flight time difference signal, and an audible and visual alarm can trigger a three-level alarm according to pre-tightening force deviation. The data remote transmission device provides data for the digital twin system, the display unit displays a pre-tightening force distribution atlas in real time, and a digital twin system platform generates a three-dimensional visual dynamic model and the like. According to the method, bolt pre-tightening force monitoring is achieved through the system. The system is high in measurement precision, convenient to install, capable of achieving real-time early warning and intelligent operation and maintenance, and suitable for various industrial fields.

Description

Technical Field

[0001] This invention relates to the field of bolt preload monitoring technology, and in particular to an ultrasonic-based bolt preload monitoring system and method. Background Technology

[0002] In modern industrial production and infrastructure construction, bolted connections are a common mechanical connection method, widely used in various equipment and structures. The magnitude of the bolt preload directly affects the reliability and safety of the connection. Insufficient preload may lead to loosening of the connection, increased vibration, or even equipment failure or structural damage; while excessive preload may overload the bolt, causing fatigue cracks and reducing its service life.

[0003] Traditional methods for monitoring bolt preload mainly include the torque method and the strain gauge method. The torque method indirectly estimates the preload by measuring the tightening torque, but this method is greatly affected by factors such as the friction coefficient and thread precision, resulting in a large measurement error. Although the strain gauge method has high measurement accuracy, it requires attaching strain gauges to the bolt, which is complex to install, and the strain gauges are easily affected by factors such as ambient temperature and humidity, resulting in lower reliability.

[0004] With the development of industrial automation and intelligence, higher requirements have been placed on the real-time monitoring of bolt preload. For example... Figure 4 As shown, ultrasonic technology, as a non-contact measurement method, has advantages such as high measurement accuracy, fast response speed, and real-time monitoring, and has received increasing attention in the field of bolt preload monitoring. For example, Chinese invention patent application number 2023112118291 discloses a wireless passive ultrasonic bolt preload monitoring device and method. The wireless passive ultrasonic bolt preload monitoring device includes a wireless passive preload sensing tag based on ultrasonic time-of-flight and a matching reader system. The wireless passive preload sensing tag includes a piezoelectric ultrasonic transducer and a tag coupling coil; the reader system includes a reader coupling coil, an ultrasonic excitation and echo receiving analog front end, and an ultrasonic time-of-flight processing module. It can monitor the bolt status and preload changes in real time, promptly grasp the bolt's working condition, and has the significant advantages of convenient and non-destructive installation.

[0005] However, existing ultrasonic-based bolt preload monitoring systems still have some problems in practical applications. For example, the aforementioned patent only filters and amplifies the collected ultrasonic signals before comparing them with a threshold. However, in actual use, industrial environments (oil, vibration, temperature) can cause ultrasonic echo signal distortion and low signal-to-noise ratio. Traditional threshold comparison methods are prone to misjudging propagation time, leading to significant errors. Furthermore, the propagation of ultrasound in bolts is affected by stress (preload), resulting in slight nonlinear changes in propagation speed. Traditional linear fitting also has a correspondingly large error. This leads to insufficient overall signal processing accuracy, affecting the monitoring effect of bolt preload. Additionally, existing technologies can only monitor preload unidirectionally, failing to achieve three-dimensional visualization monitoring, making observation less intuitive and lacking in intelligence. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an ultrasonic-based bolt preload monitoring system and method, which solves the problems of large signal processing errors, insufficient detection accuracy, and inability to achieve three-dimensional visualization monitoring in existing technologies.

[0007] In a first aspect, the present invention proposes an ultrasonic-based bolt preload monitoring system, comprising:

[0008] An ultrasonic testing probe is uniformly fitted and installed on the surface of the bolt body to detect the time-of-flight difference signal of ultrasonic waves transmitted inside the bolt.

[0009] The data processing module receives the time-of-flight signal collected by the ultrasonic testing probe, processes the signal through a neural network model, and finally calculates the bolt prestress.

[0010] The alarm device triggers multi-level alarms based on the deviation between the bolt prestress output by the data processing module and the set threshold.

[0011] The digital twin platform establishes a digital twin system based on the bolt prestress and the bolt's three-dimensional model, and simultaneously generates a three-dimensional visualized dynamic model of the prestress.

[0012] Furthermore, the ultrasonic probe uses a wideband piezoelectric sensor and is installed using an adjustable magnetic base and a threaded dual-mode fixing assembly. The surface of the ultrasonic probe is coated with a waveguide coupling agent.

[0013] Furthermore, the alarm device includes an audible and visual alarm with at least three different alarm modes.

[0014] Furthermore, it also includes a display module, which is simultaneously connected to the data processing module and the digital twin platform to display bolt preload parameters or a three-dimensional visualized dynamic model of the bolt preload.

[0016] Secondly, the present invention also proposes a method for monitoring bolt preload, comprising the following steps:

[0017] S1. Install an ultrasonic testing probe on the surface of the bolt, and then collect the time-of-flight difference signal of ultrasonic waves transmitted inside the bolt through the ultrasonic testing probe.

[0018] S2. The time-of-flight difference signal is processed through a neural network model to remove noise and improve signal accuracy. The neural network model is an improved deep residual network model, comprising:

[0019] The single signal input is replaced with two parallel input branches in the time and frequency domains. The first convolutional layer is replaced with a time-frequency dual-channel feature extraction structure. Then, an adaptive feature fusion layer is added after the first convolutional layer to replace the traditional direct connection method.

[0020] An attention module is inserted between the output of each residual block and the input of the next residual block to dynamically weight data from different sensor channels.

[0021] The final features are modeled temporally, and time-series memory units are used to track the trajectory of bolt preload changes.

[0022] Finally, the bolt prestress is calculated based on the time-of-flight difference signal obtained by the improved deep residual network model.

[0023] S3. Based on the deviation between the calculated bolt prestress and the standard value, multi-level early warning is implemented;

[0024] S4. Upload the calculated bolt prestress to the digital twin platform, simultaneously generate a three-dimensional visualized dynamic model of preload, integrate finite element simulation data to generate a stress distribution cloud map of the bolt group, establish a knowledge graph mapping of acoustic characteristics, preload, and fatigue damage, and automatically generate a maintenance work order containing suggested torque values ​​and push it to the inspection terminal.

[0025] Furthermore, in the improved deep residual network model of step S2, the method of replacing the input layer and the first convolutional layer of the traditional deep residual network with a time-frequency dual-stream feature extraction layer is as follows:

[0026] The original acoustic signal is processed in the time domain branch to extract time-domain features such as peak time and rise slope; a short-time Fourier transform (STFT) is performed on the original acoustic signal in the frequency domain branch.

[0027]

[0028] Where x(τ) is the original signal, ω(τ-t) is the window function, t is the time variable, and f is the frequency variable;

[0029] After feature extraction via temporal and frequency domain branches, global average pooling is performed on the temporal and frequency domain features, followed by full-connected layer expansion. Finally, the sigmoid function is used for activation to obtain the fusion weights, ultimately yielding the input feature map F.

[0030]

[0031] Where α is the fusion weight, F_t and F_f are the time-domain and frequency-domain features, respectively, GAP is global flat pooling, FC is the flattening operation, and sigmoid is the activation function.

[0032] Furthermore, in the improved deep residual network model of step S2, an attention module is inserted, which simultaneously acts on the output features of the time-frequency dual-stream feature extraction layer and the features of the subsequent residual layer. The channel attention mapping is as follows:

[0033]

[0034] Where F is the input feature map, GAP is global average pooling, W1 and W2 are the weights of the fully connected layer, and σ is the activation function.

[0035] Furthermore, in the improved deep residual network model of step S2, after all residual blocks and attention modules have been processed, when performing temporal modeling on the final features, a gated recurrent unit (GRU) is used to track the trajectory of bolt preload changes:

[0036]

[0037] Among them, z t To update the gate, r t To reset the door, h t To be in a hidden state, W r Reset the weight matrix of the gate, W z It is the weight matrix of the updated gate, h t-1 h is the hidden state from the previous moment. t This is the candidate hidden state.

[0038] Furthermore, in the improved deep residual network model of step S2, after the temporal modeling is completed, a pre-training fine-tuning strategy is adopted. Pre-training is performed on a pre-collected large-scale standard bolt preload dataset, and then fine-tuning is performed on the target dataset. The pre-training loss function is:

[0039]

[0040] Where N is the number of samples, C is the number of categories, and y i,c For real labels, p i,c To predict probabilities.

[0041] Furthermore, in step S3, a level three audible and visual alarm is triggered based on the deviation between the calculated bolt prestress and the standard value.

[0042] Level 1 warning: When the preload deviation is within ±3%, a green light will be emitted as a warning.

[0043] Level 2 warning: When the preload deviation is between ±3% and ±5%, a yellow light will be emitted and an intermittent buzzer will sound;

[0044] Level 3 warning: When the preload deviation exceeds ±5%, a red light will be emitted and a continuous beeping sound will be heard.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention optimizes the traditional signal input by improving the deep residual network model, replacing the single time-domain signal input with two parallel input branches in the time and frequency domains, thus obtaining two-dimensional time-frequency features. Considering that industrial noise is mostly single-frequency interference, the two-dimensional time-frequency features can automatically separate noise from the effective signal, resulting in more accurate feature extraction.

[0047] 2. The present invention also introduces an attention module corresponding to the residual block, which simultaneously acts on the output features of the time-frequency dual-stream feature extraction layer and the features of the subsequent residual layer. It adopts a combination of channel attention and spatial attention to dynamically weight data from different sensor channels, enhance the model's sensitivity to preload, and can dynamically reflect the changing trend of preload, thus better judging the working condition of the bolt.

[0048] 3. This invention also introduces a digital twin system to generate a three-dimensional visualized dynamic model of preload force and generate the latest stress distribution cloud map, which has a more intuitive visual effect. At the same time, the stress distribution cloud map is used to analyze and predict the trend of preload force change and determine the fatigue damage state of the bolt, which has a higher level of intelligence. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the monitoring system according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the improved deep residual network structure in an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the digital twin platform architecture in an embodiment of the present invention.

[0052] Figure 4 This is a schematic diagram of the principle of ultrasonic monitoring of bolt preload. Detailed Implementation

[0053] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 As shown in the figure, this invention proposes an ultrasonic-based bolt preload monitoring system, comprising: an ultrasonic detection probe, which is uniformly fitted and installed on the surface of the bolt body to detect the time-of-flight signal of ultrasonic waves transmitted inside the bolt; a data processing module, which receives the time-of-flight signal collected by the ultrasonic detection probe, processes the signal through a neural network model, and finally calculates the bolt preload; an alarm device, which triggers multi-level alarms based on the deviation between the bolt preload output by the data processing module and a set threshold; and a digital twin platform, which establishes a digital twin system based on the bolt preload and a three-dimensional model of the bolt, and synchronously generates a three-dimensional visualized dynamic model of the preload.

[0055] The ultrasonic probe, serving as the signal acquisition front end of the system, employs a wideband piezoelectric sensor. Its wideband characteristics cover the vibration frequency range of bolts under different preload conditions, ensuring comprehensive signal acquisition. Regarding the installation structure, an adjustable magnetic base and threaded fixing components are used to adapt to installation requirements under various working conditions. For scenarios requiring frequent disassembly or temporary monitoring, the adjustable magnetic base uses a strong magnet to adhere to the bolt surface; while for long-term fixed monitoring, the threaded fixing components engage with pre-drilled threads on the bolt body to securely install the probe. To improve ultrasonic coupling efficiency, the probe surface is coated with a high-temperature waveguide coupling agent. This agent maintains good waveguide performance even at high temperatures, effectively reducing ultrasonic wave reflection loss at the probe-bolt contact surface.

[0056] In this embodiment, each probe can be assigned a unique ID, which is bound to information such as the probe's installation location and model parameters and stored in the system's database. This design facilitates the system's management and identification of each probe, ensuring accurate mapping to specific bolt positions during data acquisition and processing.

[0057] The data processing module is the core of the system's data processing. It employs a high-performance GPU and connects to each ultrasonic probe via cables to acquire the electrical signals output by the probes in real time. The data processing module incorporates an improved neural network model to process the time-of-flight signal to calculate the bolt preload.

[0058] Based on the calculated preload value, the system compares it with a preset threshold to determine the preload status. When the preload deviation exceeds the corresponding threshold, the three-level early warning mechanism of the audible and visual alarm in the alarm device is triggered, and the early warning information is uploaded to the digital twin system platform.

[0059] In the specific implementation plan, a touch screen display is also included as a display module. After receiving the bolt preload information transmitted by the data processing module, it displays the bolt preload distribution map, historical trend, and other information in real time using charts and curves. This facilitates the observation, operation, and adjustment of the monitoring system by the staff.

[0060] like Figure 3 As shown, this embodiment also includes a data remote transmission device that uploads real-time collected preload and temperature data to a remote server, which then transmits the data to the digital twin platform. Upon receiving the data, the digital twin platform synchronously updates the three-dimensional visualized preload dynamic model, generating the latest stress distribution cloud map. Simultaneously, based on the preload status assessment results, fatigue damage prediction, and maintenance strategy, the digital twin system platform automatically generates a maintenance work order containing suggested torque values ​​and pushes it to the inspection terminal. Inspection personnel perform on-site maintenance according to the work order information. After maintenance is completed, the maintenance results are fed back to the system, which updates the relevant data, completing a full monitoring and maintenance cycle.

[0061] Throughout the monitoring process, the touchscreen display shows the real-time distribution of preload force on all bolts of the generator set, allowing operators to check the status of each bolt at any time. The data transmission device uploads the monitoring data to a remote server, enabling management personnel to monitor the bolt preload force of the generator set remotely, ensuring safe operation of the equipment. This method achieves real-time and accurate monitoring of bolt preload force. Combined with digital twin technology, it improves the intelligence level of equipment maintenance, reduces the risk of equipment failure, and enhances the operational reliability and service life of the equipment.

[0062] Taking bolt group monitoring of large generator sets as an example:

[0063] An array of ultrasonic probes is installed on bolts at key locations in the generator set, such as cylinder mating surfaces and bearing housings. Each probe is securely mounted using an adjustable magnetic base or threaded fixing assembly, and its surface is coated with a high-temperature waveguide coupling agent to adapt to the high-temperature operating environment of the generator set.

[0064] The data acquisition, processing, and analysis unit collects signals from each probe in real time and calculates the bolt preload using a built-in deep residual neural network model. When the preload deviation of a bolt exceeds the secondary warning threshold, an orange alarm is triggered by both sound and light. Simultaneously, the digital twin system platform generates a stress distribution cloud map of the bolt, displaying areas of stress concentration. The system automatically generates a maintenance work order, recommending torque adjustment for the bolt, and pushes the work order to the handheld terminal of the inspection personnel.

[0065] The inspection personnel arrived at the site according to the work order instructions and adjusted the bolts using a torque wrench according to the recommended torque value. After the adjustment was completed, the maintenance results were fed back to the system via a handheld terminal. The system updated the preload data in the digital twin model, the stress distribution cloud map returned to normal, and the warning was lifted.

[0067] The bolt preload monitoring method of this invention is referenced in the embodiments. Figure 2 As shown, it includes the following steps:

[0068] (1) Install an ultrasonic testing probe on the surface of the bolt, and then collect the time-of-flight signal of ultrasonic waves transmitted inside the bolt through the ultrasonic testing probe.

[0069] (2) The time difference of flight signal is processed by a neural network model to remove noise and improve signal accuracy. The neural network model is an improved deep residual network model, specifically:

[0070] (2.1) Replace the single signal input with two parallel input branches in the time and frequency domains, replace the first convolutional layer with a time-frequency dual-channel feature extraction structure, and then add an adaptive feature fusion layer after the first convolutional layer to replace the traditional direct connection method:

[0071] The original acoustic signal is processed in the time domain branch to extract time-domain features such as peak time and rise slope; a short-time Fourier transform (STFT) is performed on the original acoustic signal in the frequency domain branch.

[0072]

[0073] Where x(τ) is the original signal, ω(τ-t) is the window function, t is the time variable, and f is the frequency variable;

[0074] After feature extraction via temporal and frequency domain branches, global average pooling is performed on the temporal and frequency domain features, followed by full-connected layer expansion. Finally, the sigmoid function is used for activation to obtain the fusion weights, ultimately yielding the input feature map F.

[0075]

[0076] Where α is the fusion weight, F_t and F_f are the time-domain and frequency-domain features, respectively, GAP is global flat pooling, FC is the flattening operation, and sigmoid is the activation function.

[0077] (2.2) An attention module is inserted between the output of each residual block and the input of the next residual block to dynamically weight the data from different sensor channels. This attention module simultaneously affects the output features of the time-frequency dual-stream feature extraction layer and the features of the subsequent residual layers. The channel attention mapping is as follows:

[0078]

[0079] Where F is the input feature map, GAP is global average pooling, W1 and W2 are the weights of the fully connected layer, and σ is the activation function.

[0080] (2.3) Perform time-series modeling on the final features, and use time-series memory units and gated loop units (GRUs) to track the trajectory of bolt preload changes:

[0081]

[0082] Among them, z t To update the gate, r t To reset the door, h t To be in a hidden state, W r Reset the weight matrix of the gate, W z It is the weight matrix of the updated gate, h t-1 h is the hidden state from the previous moment. t This is the candidate hidden state.

[0083] (2.4) After the time series modeling is completed, a pre-training and fine-tuning strategy is adopted. Pre-training is performed on a large-scale standard bolt preload dataset collected in advance, and then fine-tuning is performed on the target dataset. The pre-training loss function is:

[0084]

[0085] Where N is the number of samples, C is the number of categories, and y i,c For real labels, p i,c To predict probabilities.

[0086] (2.5) After processing by the improved deep residual network model described above, the processed time-of-flight difference signal Δt is obtained. Finally, based on the obtained time-of-flight difference signal, the bolt prestress F0 is calculated. The specific formula is based on Hooke's Law:

[0087]

[0088] Where E is the elastic modulus of the bolt material, S is the effective cross-sectional area of ​​the bolt, and v is the ultrasonic propagation velocity. L0 represents the time difference of flight of the ultrasonic wave, and L0 represents the bolt clamping length.

[0089] (3) Trigger a level three audible and visual alarm based on the deviation between the calculated bolt prestress and the standard value:

[0090] Level 1 warning: When the preload deviation is within ±3%, a green light will be emitted as a warning.

[0091] Level 2 warning: When the preload deviation is between ±3% and ±5%, a yellow light will be emitted and an intermittent buzzer will sound;

[0092] Level 3 warning: When the preload deviation exceeds ±5%, a red light will be emitted and a continuous beeping sound will be heard.

[0093] (4) Upload the calculated bolt prestress to the digital twin platform, generate a three-dimensional visualized preload dynamic model, integrate finite element simulation data to generate a bolt group stress distribution cloud map, establish a mapping knowledge graph of acoustic characteristics-preload-fatigue damage, and automatically generate a maintenance work order containing suggested torque values ​​and push it to the inspection terminal.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An ultrasonic-based bolt preload monitoring system, characterized in that, include: An ultrasonic testing probe is uniformly fitted and installed on the surface of the bolt body to detect the time-of-flight difference signal of ultrasonic waves transmitted inside the bolt. The data processing module receives the time-of-flight signal collected by the ultrasonic testing probe, processes the signal through a neural network model, and finally calculates the bolt prestress. The alarm device triggers multi-level alarms based on the deviation between the bolt prestress output by the data processing module and the set threshold. The digital twin platform establishes a digital twin system based on the bolt prestress and the bolt's three-dimensional model, and simultaneously generates a three-dimensional visualized dynamic model of the prestress.

2. The ultrasonic-based bolt preload monitoring system as described in claim 1, characterized in that: The ultrasonic probe uses a wideband piezoelectric sensor and is installed using an adjustable magnetic base and a threaded dual-mode fixing assembly. The surface of the ultrasonic probe is coated with a waveguide coupling agent.

3. The ultrasonic-based bolt preload monitoring system as described in claim 1, characterized in that: The alarm device includes an audible and visual alarm and has no fewer than three different alarm modes.

4. The ultrasonic-based bolt preload monitoring system as described in claim 1, characterized in that: It also includes a display module, which is simultaneously connected to the data processing module and the digital twin platform to display bolt preload parameters or a three-dimensional visual dynamic model of the bolt preload.

5. A monitoring method for the bolt preload monitoring system as described in claim 1, characterized in that, Includes the following steps: S1. Install an ultrasonic testing probe on the surface of the bolt, and then collect the time-of-flight difference signal of ultrasonic waves transmitted inside the bolt through the ultrasonic testing probe. S2. The time-of-flight difference signal is processed through a neural network model to remove noise and improve signal accuracy. The neural network model is an improved deep residual network model, comprising: The single signal input is replaced with two parallel input branches in the time and frequency domains. The first convolutional layer is replaced with a time-frequency dual-channel feature extraction structure. Then, an adaptive feature fusion layer is added after the first convolutional layer to replace the traditional direct connection method. An attention module is inserted between the output of each residual block and the input of the next residual block to dynamically weight the data from different sensor channels. The final features are modeled temporally, and time-series memory units are used to track the trajectory of bolt preload changes. Finally, the bolt prestress is calculated based on the time-of-flight difference signal obtained by the improved deep residual network model. S3. Based on the deviation between the calculated bolt prestress and the standard value, multi-level early warning is implemented; S4. Upload the calculated bolt prestress to the digital twin platform, simultaneously generate a three-dimensional visualized dynamic model of preload, integrate finite element simulation data to generate a stress distribution cloud map of the bolt group, establish a knowledge graph mapping of acoustic characteristics, preload, and fatigue damage, and automatically generate a maintenance work order containing suggested torque values ​​and push it to the inspection terminal.

6. The monitoring method as described in claim 5, characterized in that: In the improved deep residual network model of step S2, the method of replacing the input layer and the first convolutional layer of the traditional deep residual network with a time-frequency dual-stream feature extraction layer is as follows: The original acoustic signal is processed in the time domain branch to extract time-domain features such as peak time and rise slope; a short-time Fourier transform (STFT) is performed on the original acoustic signal in the frequency domain branch. Where x(τ) is the original signal, ω(τ-t) is the window function, t is the time variable, and f is the frequency variable; After feature extraction via temporal and frequency domain branches, global average pooling is performed on the temporal and frequency domain features, followed by full-connected layer expansion. Finally, the sigmoid function is used for activation to obtain the fusion weights, ultimately yielding the input feature map F. Where α is the fusion weight, F_t and F_f are the time-domain and frequency-domain features, respectively, GAP is global flat pooling, FC is the flattening operation, and sigmoid is the activation function.

7. The monitoring method as described in claim 5, characterized in that: In the improved deep residual network model of step S2, an attention module is inserted, which simultaneously acts on the output features of the time-frequency dual-stream feature extraction layer and the features of the subsequent residual layer. The channel attention mapping is as follows: Where F is the input feature map, GAP is global average pooling, W1 and W2 are the weights of the fully connected layer, and σ is the activation function.

8. The monitoring method as described in claim 5, characterized in that: In the improved deep residual network model of step S2, after all residual blocks and attention modules have been processed, when performing temporal modeling on the final features, a gated recurrent unit (GRU) is used to track the trajectory of bolt preload changes. Among them, z t To update the gate, r t To reset the door, h t To be in a hidden state, W r Reset the weight matrix of the gate, W z It is the weight matrix of the updated gate, h t-1 h is the hidden state from the previous moment. t This is the candidate hidden state.

9. The monitoring method as described in claim 5, characterized in that: In the improved deep residual network model of step S2, after the temporal modeling is completed, a pre-training and fine-tuning strategy is adopted. Pre-training is performed on a pre-collected large-scale standard bolt preload dataset, and then fine-tuning is performed on the target dataset. The pre-training loss function is: Where N is the number of samples, C is the number of categories, and y i,c For real labels, p i,c To predict probabilities.

10. The monitoring method as described in claim 5, characterized in that, In step S3, a three-level audible and visual alarm is triggered based on the deviation between the calculated bolt prestress and the standard value: Level 1 warning: When the prestress deviation is within ±3%, a green light is emitted as a warning; Level 2 warning: When the prestress deviation is between ±3% and ±5%, a yellow light is emitted and an intermittent buzzer is sounded; Level 3 warning: When the prestress deviation exceeds ±5%, a red light is emitted and a continuous buzzer is sounded.

Images