A method and system for detecting abnormal installation of pneumatic gun bolts

Abstract

The present invention relates to the technical field of pneumatic gun detection, and discloses a method and system for detecting abnormal bolt installation of a pneumatic gun. The method obtains the actual tightening force value, air pressure fluctuation amplitude and frequency, and surface feature image during the installation process; calculates the dynamic deviation between the actual and preset tightening forces to form a tightening force deviation sequence; performs multi-dimensional discrete analysis on the deviation sequence based on an adaptive clustering algorithm to obtain the number of abnormal discrete point clusters; if the number ratio exceeds a preset threshold, it is marked as a bolt installation abnormality; then, the air pressure fluctuation amplitude and frequency of the pneumatic gun are compared to see if they exceed a preset range to determine whether the air pressure is stable; if not, the surface image is input into a trained image recognition model to output the probability of leakage traces; if the probability exceeds the preset threshold, three-dimensional flow field data of the leakage point is collected to calculate the gas leakage amount. The method can achieve precise detection and quantitative analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of pneumatic gun detection, and in particular to a method and system for detecting abnormal installation of a pneumatic gun bolt. Background Art

[0002] In the industrial field, bolts are an important component in connecting and fixing key components. Their installation quality directly affects the overall stability and safety of the equipment. Especially in pneumatic systems and high-precision equipment, monitoring the tightening force and installation status of bolts is particularly important.

[0003] At present, the detection of abnormal installation of pneumatic gun bolts mainly adopts technologies such as tightening force monitoring, air pressure fluctuation analysis and image recognition. Tightening force monitoring relies on torque sensors, and the installation quality is judged by measuring the tightening force deviation of the bolts, but it is difficult to reflect the trend of the overall installation process. Air pressure fluctuation analysis uses air pressure sensors to detect pressure changes, but air pressure fluctuations are greatly affected by factors such as the external environment and the operating status of the equipment, resulting in unstable detection results. Image recognition technology is used to analyze the airflow characteristics on the surface of the pneumatic gun. Although it can assist in judging abnormal areas, it is difficult to accurately identify leak points due to imaging conditions. At the same time, in terms of gas leak detection, ultrasonic leak detection technology judges the leakage situation by analyzing high-frequency sound waves, but due to background noise interference and sound wave attenuation, its detection results can often only identify the existence of leaks, but cannot quantitatively calculate the amount of leakage.

[0004] Existing technologies often only provide qualitative assessments, failing to quantify specific gas leak data. Many traditional detection methods can only identify airflow anomalies and the presence of leaks, but fail to provide accurate leakage data. This lacks a clear basis for repair and treatment decisions. In summary, existing technologies suffer from low accuracy in detecting bolt installation anomalies in pneumatic guns. Summary of the Invention

[0005] The present invention provides a method and system for detecting abnormal installation of a pneumatic gun bolt, so as to achieve accurate detection and quantitative analysis.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a method for detecting abnormal installation of a pneumatic gun bolt, comprising:

[0007] Obtain the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process;

[0008] Performing dynamic deviation calculation on the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence;

[0009] Based on the adaptive clustering algorithm, a multi-dimensional discrete analysis is performed on the tightening force deviation sequence to obtain the number of abnormal discrete point clusters;

[0010] If the proportion of the abnormal discrete point clusters exceeds a preset threshold, it is marked as a bolt installation abnormality;

[0011] Comparing the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with a preset amplitude threshold and frequency range, respectively, to determine whether the air pressure is stable;

[0012] Comparing the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with a preset amplitude threshold and frequency range, respectively, to determine whether the air pressure is stable;

[0013] If the air pressure is unstable, the surface feature image of the air pressure gun with unstable air pressure is input into the trained image recognition model, and the probability of leakage traces is output;

[0014] Extracting acoustic wave features from the image recognition results based on ultrasonic leak detection technology to obtain acoustic wave frequency features, and determining whether there is a gas leak based on the acoustic wave frequency features;

[0015] If the leakage trace probability is greater than a preset threshold, the three-dimensional flow field data of the leakage point is obtained, and the leakage amount is calculated based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount.

[0016] Preferably, the dynamic deviation calculation of the actual tightening force value and the preset tightening force value to obtain the tightening force deviation sequence includes:

[0017] The tightening force deviation sequence is calculated using the following formula:

[0018]

[0019] in, is the tightening force deviation sequence, is the actual tightening force value, The preset tightening force value.

[0020] Preferably, the method of performing multi-dimensional discrete analysis on the tightening force deviation sequence based on the adaptive clustering algorithm to obtain the number of abnormal discrete point clusters includes:

[0021] Based on a preset deviation direction judgment rule, determining the deviation direction of the tightening force deviation sequence, and obtaining a deviation direction distribution map;

[0022] Based on the adaptive clustering algorithm, anomaly identification is performed on the deviation direction distribution map to obtain the number of abnormal discrete point clusters.

[0023] Preferably, the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun according to the abnormal bolt installation are compared with a preset amplitude threshold and frequency range respectively to determine whether the air pressure is stable, including:

[0024] When the air pressure fluctuation amplitude does not exceed the preset amplitude threshold, it is determined that the air pressure is in a stable state;

[0025] When the air pressure fluctuation amplitude exceeds the preset amplitude threshold, the air pressure fluctuation frequency is further compared with the preset frequency range. If the air pressure fluctuation frequency is not within the preset frequency range, it is determined that the air pressure is in an unstable state.

[0026] Preferably, if the air pressure is unstable, the surface feature image of the air pressure gun with unstable air pressure is input into a pre-trained image recognition model to output the probability of leakage traces, including:

[0027] The image recognition model is trained by the ResNet-18 neural network model;

[0028] The data format of the surface feature image of the air pressure gun is converted through the input layer of the image recognition model to obtain an image in the form of a three-dimensional tensor;

[0029] Performing a convolution operation on the three-dimensional tensor image through the convolution layer of the image recognition model to extract local texture features of the image to obtain a first feature map;

[0030] Performing a nonlinear feature transformation on the first feature map through an activation function layer of the image recognition model to obtain an enhanced second feature map;

[0031] Downsampling the second feature map through a pooling layer of the image recognition model to obtain a feature map with a reduced size;

[0032] Performing feature integration and classification processing on the feature map after size reduction through the fully connected layer of the image recognition model, and outputting classification feature results related to the leakage state;

[0033] The output layer of the image recognition model performs probability calculation on the classification feature results and outputs the leakage trace probability.

[0034] Preferably, if the air pressure is unstable, the surface feature image of the air pressure gun with unstable air pressure is input into a trained image recognition model to output the probability of leakage traces, including:

[0035] The training process of the image recognition model includes:

[0036] Based on the supervised learning algorithm, the pre-prepared training set is optimized to obtain the optimized training set;

[0037] Inputting the optimized training set into the image recognition model and outputting the predicted leakage trace probability;

[0038] Calculate the difference between the predicted leakage trace probability and the actual leakage trace probability to obtain an output error value;

[0039] Based on the output error value, executing a back propagation algorithm to update load demand model parameters and optimize the image recognition model;

[0040] Repeat the above process and continuously optimize the load demand model until the output accuracy of the load demand on the training set reaches the preset output accuracy requirement, and then stop training.

[0041] Preferably, if the probability of the leakage trace is greater than a preset threshold, obtaining three-dimensional flow field data of the leakage point, and calculating the leakage amount based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount, including:

[0042] The three-dimensional flow field data includes: velocity vector, density, leakage start time, leakage end time and leakage cross-sectional area;

[0043] Calculated by the following formula:

[0044]

[0045] Where Q is the gas leakage volume, t1 is the leakage start time, t2 is the leakage end time, V is the velocity vector, p is the density, and A is the leakage cross-sectional area.

[0046] In a second aspect, the present invention provides a system for detecting abnormal installation of a pneumatic gun bolt, comprising:

[0047] A data acquisition module is used to obtain the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process;

[0048] a tightening force deviation module, configured to calculate a deviation between the actual tightening force value and a preset tightening force value to obtain a tightening force deviation value;

[0049] A discrete point distribution module is used to determine the deviation direction according to the tightening force deviation value and obtain the discrete point distribution of the bolt tightening force;

[0050] A discrete point quantity module is used to perform abnormality judgment on the discrete points of the bolt tightening force discrete point distribution based on a preset threshold value of the bolt tightening force discrete point distribution interval, and obtain the number of abnormal discrete points;

[0051] an abnormality judgment module, configured to mark the bolt installation as abnormal if the number of the abnormal discrete points is greater than or equal to a preset abnormal discrete point number threshold;

[0052] an air pressure stability judgment module, configured to compare the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with a preset amplitude threshold and frequency range, respectively, to determine whether the air pressure is stable;

[0053] An image recognition module is used to input a surface feature image of the air pressure gun with unstable air pressure into a pre-trained image recognition model to obtain an image recognition result if the air pressure is unstable;

[0054] A gas leakage judgment module is used to extract sound wave features from the image recognition results based on ultrasonic leak detection technology to obtain sound wave frequency features, and judge whether there is a gas leak based on the sound wave frequency features;

[0055] The gas leakage module is configured to obtain gas flow data at the leak point if a gas leak occurs in the pneumatic gun, and calculate the leakage amount based on the gas flow data at the leak point to obtain the gas leakage amount. In a third aspect, the present invention also provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements any of the aforementioned methods for detecting abnormal bolt installation in a pneumatic gun.

[0056] In a fourth aspect, the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned methods for detecting abnormal installation of a pneumatic gun bolt.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] (1) The present invention can accurately identify abnormal bolt installation conditions and improve the intelligent level of installation quality control by dynamically calculating the actual tightening force of the air pressure gun during the bolt installation process and performing discrete analysis on the deviation sequence in combination with an adaptive clustering algorithm.

[0059] (2) The present invention constructs a stability judgment mechanism with the air pressure fluctuation amplitude as the main criterion and the fluctuation frequency as the auxiliary criterion. It adopts a conditional cascade judgment method to effectively reduce the probability of misjudgment and improve the system's recognition reliability of abnormal air pressure fluctuations.

[0060] (3) The present invention uses an image recognition model trained based on the ResNet-18 neural network model to perform deep feature extraction and classification recognition on the surface image of the pneumatic gun, which can effectively identify potential leakage traces and achieve high-precision image-level leakage detection.

[0061] (4) After identifying that the leakage probability exceeds the standard, the present invention obtains the three-dimensional flow field data of the leakage point, combines the density and velocity vectors for time-space integration, and accurately calculates the leakage amount. It has the ability to quantitatively evaluate the degree of leakage, which is helpful for systematic diagnosis and risk warning.

[0062] In summary, the present invention provides a method for detecting abnormal installation of pneumatic gun bolts by integrating multi-source signal analysis and deep learning recognition mechanism, thereby realizing accurate detection and quantitative analysis of abnormal installation of pneumatic gun bolts. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 1 is a flow chart of a method for detecting abnormal bolt installation of a pneumatic gun provided by a first embodiment of the present invention;

[0064] Figure 2 2 is a schematic diagram of a detection system for abnormal bolt installation of a pneumatic gun provided by a second embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0066] Reference Figure 1 A first embodiment of the present invention provides a method for detecting abnormal installation of a pneumatic gun bolt, comprising the following steps:

[0067] S11, obtaining the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process;

[0068] S12, performing dynamic deviation calculation on the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence;

[0069] S13, performing a multi-dimensional discrete analysis on the tightening force deviation sequence based on an adaptive clustering algorithm to obtain the number of abnormal discrete point clusters;

[0070] S14, if the proportion of the abnormal discrete point clusters exceeds a preset threshold, it is marked as a bolt installation abnormality;

[0071] S15, comparing the air pressure fluctuation amplitude and the air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with the preset amplitude threshold and frequency range respectively to determine whether the air pressure is stable;

[0072] S16, if the air pressure is unstable, input the surface feature image of the air pressure gun with unstable air pressure into the trained image recognition model and output the probability of leakage trace;

[0073] S17, if the leakage trace probability is greater than a preset threshold, obtain the three-dimensional flow field data of the leakage point, and calculate the leakage amount based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount.

[0074] In step S11, the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process are obtained;

[0075] It is worth mentioning that first, several key parameters of the air pressure gun during the bolt installation process need to be obtained: actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image.

[0076] To obtain actual tightening force values, torque sensors are first used to monitor the tightening force during bolt installation in real time. These sensors are installed on the working components of the air pressure gun and record the changes in the bolt's tightening force in real time through their output signals. This data reflects the mechanical state of the bolt installation and helps determine whether the bolt is correctly installed within the specified range. The sensors convert the tightening force values ​​into electrical signals, which are transmitted in real time to the data acquisition system for subsequent analysis.

[0077] To obtain the amplitude and frequency of air pressure fluctuations, the air pressure sensor is responsible for monitoring air pressure changes during the operation of the air pressure gun. The air pressure sensor is installed at a key location on the air pressure gun. By collecting air pressure change data, it analyzes air pressure fluctuations over different time periods. The amplitude of air pressure fluctuations refers to the maximum value of air pressure fluctuations over a period of time, while the frequency of air pressure fluctuations refers to the number of air pressure fluctuations per unit time. Through data collection and analysis, the specific value of air pressure fluctuations can be obtained, providing a basis for subsequent judgment of air pressure stability.

[0078] Surface feature images are acquired using an image acquisition device that combines a thermal imaging camera and a high-resolution visible light camera. The thermal imaging camera detects temperature differences caused by airflow fluctuations and captures areas of temperature fluctuation, while the visible light camera records the surface morphology of the air gun. The simultaneous capture of these two cameras yields detailed surface image data, including areas of airflow fluctuations. These images are analyzed using image processing techniques for subsequent leak detection and anomaly identification. Features captured in the images serve as surface feature images, providing detailed information on the location and morphology of surface anomalies.

[0079] Acquiring this critical data requires high-precision sensors and image acquisition equipment, which are then transmitted and processed in real time through a data acquisition system to ensure the real-time and accuracy of the entire inspection process. The integration of this data provides the basis for subsequent bolt installation anomaly detection, air pressure stability assessment, and gas leak analysis.

[0080] In step S12, a dynamic deviation calculation is performed between the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence, including:

[0081] The tightening force deviation sequence is calculated using the following formula:

[0082]

[0083] in, is the tightening force deviation sequence, is the actual tightening force value, The preset tightening force value.

[0084] It's worth noting that in step S12, to identify any abnormalities during the bolt tightening process, the system first needs to obtain the actual tightening force value of the pneumatic gun during the bolt installation process. This tightening force value is collected in real time by a torque sensor installed at the output of the pneumatic gun and uploaded to the control system. The sampling period for the actual tightening force is typically every 10 milliseconds, depending on the device settings, which meets the requirement for continuous monitoring of the entire bolt tightening process. During each sampling period, the system obtains the instantaneous tightening force value at that moment, forming a raw force data sequence for the tightening process.

[0085] In order to determine whether the current tightening behavior meets the operating specifications, it is not possible to directly judge by relying solely on the original actual tightening force value, because different types of bolts are designed to correspond to different tightening force requirements. To this end, the system needs to compare and refer to a standardized tightening force reference value. This reference value is the preset tightening force value. The system has pre-established a mapping table between bolt specifications and target tightening force values. This mapping table can be obtained through laboratory calibration, standard parameters provided by equipment manufacturers, or process specifications. In actual application, the control system automatically matches the preset tightening force value corresponding to the bolt by identifying the bolt specification information used in the current work station. The bolt specification can be identified by a QR code scanner integrated in the mechanical equipment to read the bolt label information, or by the operator selecting the bolt type of the current batch through the operation terminal during the operation preparation stage. The system completes the configuration of the tightening force threshold accordingly.

[0086] Taking the M10 steel structure bolt as an example, the expected tightening force set at the construction site is 35 Nm. When the system recognizes that the bolt used in the current operation is of M10 specification, it will automatically call 35 Nm as the preset tightening force value for this task. At the same time, the actual tightening force data collected by the pneumatic gun are uploaded in time series, such as 32, 34, 36, 37, 35, 34 and other real-time torque values ​​in units of Nm. At this time, the control system calculates each actual value with 35 Nm, and obtains the deviation value at each moment in turn. This calculation process is a pure subtraction operation, which is completed in a single time and does not require the introduction of complex calculation modules such as neural networks and optimization algorithms. Therefore, the calculation speed is fast and the error is controllable.

[0087] The set of difference data obtained through the above method is called the tightening force deviation sequence. This sequence is essentially a temporal arrangement of the mechanical deviations at each sampling moment during the actual tightening process. Because bolt tightening is a dynamic process, the actual torque values ​​may exhibit different trends before, during, and after tightening. Therefore, the deviation sequence is inherently time-correlated. The deviation sequence recorded by the system not only reflects the single-point deviation at a specific moment but, more importantly, can also indicate whether there are structural anomalies in the entire tightening process, such as persistently low values ​​or drastic fluctuations over a period of time.

[0088] For example, during a batch installation operation, if the deviation sequence during bolt tightening at a certain workstation is consistently negative, the system might record it as -4, -5, -6, -5, -7, and so on. This persistent deviation from the preset value indicates a systematic deficiency in the tightening process. Conversely, if the deviation fluctuates within a reasonable range, such as repeatedly jumping between -1 and +2 Nm, the process can be considered to meet the allowable deviation specification. The system then evaluates the entire deviation sequence based on the set tolerance interval, providing criteria for triggering the anomaly clustering algorithm.

[0089] In addition, the dynamic adaptability of the preset tightening force is another key feature of this step. Since the pneumatic gun in the industrial field may need to adapt to bolts of different types, grades, and lengths, and these parameters directly affect the final target torque requirements, the preset value must have the ability to be dynamically adjusted. If the system only uses a fixed value as a comparison benchmark, it will cause some bolts to be misjudged as abnormal, while others will be missed. Therefore, this system realizes the automatic scheduling of the target torque value in the parameter library by binding with the material information, thereby ensuring that the calculation of the tightening force deviation sequence has an accurate benchmark reference.

[0090] In step S13, based on the adaptive clustering algorithm, a multi-dimensional discrete analysis is performed on the tightening force deviation sequence to obtain the number of abnormal discrete point clusters, including:

[0091] Based on a preset deviation direction judgment rule, determining the deviation direction of the tightening force deviation sequence, and obtaining a deviation direction distribution map;

[0092] Based on the adaptive clustering algorithm, anomaly identification is performed on the deviation direction distribution map to obtain the number of abnormal discrete point clusters.

[0093] It's worth noting that in step S13, to further analyze abnormal behavior during bolt tightening, the system incorporates a multidimensional discrete analysis method based on an adaptive clustering algorithm. This method is specifically designed to identify possible abnormal point clusters within the tightening force deviation sequence. This step already constructs the tightening force deviation sequence in the previous stage, identifying the difference between the actual tightening force and the preset tightening force at each time point. This sequence reflects the deviations throughout the tightening process. To further analyze the distribution trend, directionality, and degree of clustering of the deviations, the system employs a deviation direction determination rule in conjunction with the adaptive clustering algorithm to identify discrete point clusters.

[0094] First, the system directional-classifies each data point in the tightening force deviation sequence based on a set of clear deviation direction judgment rules. This judgment rule is implemented based on numerical symbol judgment. When the deviation value at a certain time point is greater than zero, it means that the actual tightening force at that moment is higher than the preset value, and the system determines that the point is in an "overtight" state; conversely, when the deviation value is less than zero, it means that the actual tightening force is lower than the preset value, and the system determines that it is in an "undertight" state. If the deviation value is equal to zero, it is considered to be consistent with the preset value. According to this judgment rule, the system traverses the deviation sequence in sequence, judges the directionality of each deviation value point by point, and draws a deviation direction distribution diagram during the tightening process in the coordinate system. In this distribution diagram, the horizontal axis is the time series and the vertical axis is the deviation value. The positive value area represents overtightening and the negative value area represents undertightening. The changing trend of the deviation on the time axis can be intuitively observed through graphics.

[0095] After obtaining the deviation direction distribution map, the system further performs clustering and identification operations on the deviation points that show a central trend in the map. The goal of clustering is to classify deviation points with similar deviation directions and similar time positions into a cluster, so as to determine whether there is a systematic abnormal trend. In order to improve the stability and adaptability of the clustering effect, this system uses an adaptive clustering algorithm. This algorithm is an improved clustering method based on the traditional density clustering principle. Its core lies in the ability to automatically adjust clustering parameters, such as neighborhood radius and minimum number of samples, according to the distribution characteristics of the input data. Unlike the fixed-parameter K-means or DBSCAN algorithms, the adaptive clustering algorithm does not require manual specification of the number of clusters before clustering. Instead, it dynamically determines the cluster boundaries by automatically scanning and adjusting the data density, so as to better fit the distribution structure of the actual data.

[0096] Before performing the clustering operation, the system first normalizes the deviation direction distribution map, and maps the deviation value and the time axis to a standardized numerical range respectively, in order to eliminate the influence of different bolt specifications and different tightening cycles. Subsequently, the adaptive clustering algorithm traverses the data space and determines whether each data point belongs to an existing cluster or should be used as a new cluster center based on the local density distribution. During the clustering process, the algorithm iterates continuously to identify one or more groups of local high-density point clusters, which represent periods of continuous anomalies during the tightening process. Finally, the system counts the number of all identified abnormal point clusters and compares it with the abnormality ratio threshold set by the system to determine whether the tightening process needs to be marked as a bolt installation anomaly.

[0097] For example, during a bolt tightening process, the tightening force deviation sequence contains multiple consecutive negative points. After determining the deviation direction, the system determines that the tightening force is "undertightened." If these points are concentrated on the time axis, the system plots them in a deviation direction distribution diagram and identifies a distinct cluster of low-value points through clustering. Assuming that the number of outliers in this cluster accounts for 25% of the total number of sampling points in the entire tightening cycle, and the system's default threshold is 20%, the system determines that the tightening process has a clear "undertightening" trend and outputs an abnormality prompt.

[0098] In step S14, if the proportion of the abnormal discrete point clusters exceeds a preset threshold, it is marked as a bolt installation abnormality.

[0099] It's worth noting that in step S14, the system further comprehensively assesses the bolt's installation status based on the number of abnormal discrete point clusters identified in the previous stage. The core of this step is to establish a set of criteria for identifying "abnormal bolt installation" by setting a threshold for the proportion of abnormal point clusters in the entire tightening process. When the number of abnormal discrete point clusters counted by the system exceeds this preset threshold, the system deems the current bolt installation process to have a significant risk of abnormality, requiring a flag or alarm.

[0100] In specific implementations, the system first calculates the total number of clusters identified as abnormal discrete point clusters based on the deviation direction distribution map and the output of the adaptive clustering algorithm. Each discrete point cluster contains a certain number of continuous or approximately adjacent deviation data points. These data points are consistent in deviation amplitude and deviation direction, reflecting the long-term behavioral trend of the tightening force output by the air pressure gun deviating from the target value within a certain period of time. The system then obtains the total number of data points involved in the sampling during the entire tightening cycle and calculates the proportion of abnormal discrete point clusters. The proportion value is calculated by dividing the sum of the number of deviation points contained in all abnormal point clusters by the total number of sampling points in the entire tightening process.

[0101] The system's default threshold is 20 percent. That is, when the number of data points in an abnormal discrete point cluster accounts for 20 percent or more of the total number of samples, the system will mark the bolt as "installed abnormally." The setting of this threshold is based on engineering verification results and quality control standards. In actual industrial assembly processes, pneumatic tools and sensors may fluctuate within a certain range in a short period of time due to factors such as air pressure fluctuations, mechanical wear, or external interference. Therefore, a certain tolerance space is retained in engineering judgment. Statistical analysis of a large amount of experimental data found that when the proportion of abnormal points exceeds 20 percent, the error trend often no longer falls into the category of occasional disturbances, but has the characteristics of continuous and systematic deviations. Such deviations are often associated with installation defects such as under-tightening, over-tightening, or missing bolts.

[0102] For example, in a specific operational scenario, if the total number of sampling points within a bolt tightening cycle is 200, adaptive cluster analysis reveals three abnormal discrete point clusters, containing 12, 15, and 17 abnormal points, respectively, for a total of 44 abnormal points. At this point, the abnormal point ratio is 22%, exceeding the set threshold of 20%. The system then determines that the bolt has a significant trend of concentrated tightening force deviation and immediately marks the bolt as an installation anomaly. This is graphically indicated in the system monitoring interface, prompting the operator to recheck or tighten the bolt.

[0103] To enhance the system's engineering adaptability, the system supports configurable adjustments to the threshold. In mass production assembly lines or high-consistency processes, the threshold can be set to 15% or even lower to enhance quality control sensitivity. In field maintenance or non-standard parts assembly, the threshold can be set to 25% to relax the judgment criteria to avoid false alarms.

[0104] In step S15, the air pressure fluctuation amplitude and the air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation are compared with the preset amplitude threshold and frequency range respectively to determine whether the air pressure is stable, including:

[0105] When the air pressure fluctuation amplitude does not exceed the preset amplitude threshold, it is determined that the air pressure is in a stable state;

[0106] When the air pressure fluctuation amplitude exceeds the preset amplitude threshold, the air pressure fluctuation frequency is further compared with the preset frequency range. If the air pressure fluctuation frequency is not within the preset frequency range, it is determined that the air pressure is in an unstable state.

[0107] It's worth noting that in this step, after the system identifies a bolt tightening anomaly, it further determines whether the air pressure conditions causing the anomaly are stable. This judgment logic is based on a dual-feature analysis of the air pressure fluctuation signal: the amplitude and frequency of the pressure fluctuation. This two-tiered judgment mechanism improves the robustness of the system's recognition and reduces the risk of misjudgment due to errors in a single feature.

[0108] First, the system performs envelope extraction on the collected air pressure signal, calculates the difference between its maximum and minimum values ​​within the unit time window, and obtains the air pressure fluctuation amplitude for that time period. The amplitude reflects the intensity of the air pressure within a certain period. The larger the value, the more intense the air pressure fluctuation. The system sets the threshold value of the air pressure fluctuation amplitude to 0.15 MPa. Under standard stable air supply conditions, the air pressure fluctuation amplitude is mostly between 0.08 MPa and 0.12 MPa. In the event of an abnormality in the air pressure system, especially due to a delayed response of the pressure regulating valve or discontinuous air supply, the fluctuation amplitude will increase to above 0.15 MPa. Therefore, it is reasonable to set this value as a standard for judging whether the air pressure fluctuation is too large.

[0109] If the pressure fluctuation amplitude does not exceed 0.15 MPa within a given detection cycle, the system can directly determine that the current pressure state is stable, without further processing, and proceed to the next detection process. If the fluctuation amplitude exceeds this threshold, it indicates that the current air pressure supply is experiencing abnormal fluctuations, and the system will further analyze the frequency characteristics of the pressure signal.

[0110] The air pressure fluctuation frequency is obtained by performing fast Fourier transform processing on the air pressure time series signal to obtain its main frequency component. This main frequency value reflects the oscillation frequency of the air pressure signal and is closely related to the pressure regulation frequency and air supply rhythm in the air source system. The air pressure fluctuation frequency range set by the system is 1 to 5 Hz. This range also comes from the statistical results of the signal spectrum under a large number of normal air pressure working conditions. Under standard operating conditions, the oscillation frequency of the regulating valve of the air supply system generally fluctuates between 2 and 3.5 Hz, and there are very few rapid fluctuations exceeding 5 Hz or slow responses below 1 Hz. Therefore, setting 1 to 5 Hz as a reasonable frequency stability range can cover the fluctuation rhythm under normal working conditions and effectively eliminate abnormal signal interference.

[0111] In practice, if the air pressure fluctuation exceeds 0.15 MPa within a given detection cycle, the system performs frequency analysis on that pressure signal to extract its dominant frequency component. If the extracted dominant frequency falls outside the 1 to 5 Hz range, the system determines that the air pressure during that cycle is unstable, triggering image recognition to further determine if there is a leak.

[0112] To illustrate, consider a specific example: during a bolt tightening operation using a pneumatic gun, the air pressure signal collected fluctuated by 0.19 MPa and had a dominant frequency of 6.2 Hz. The system initially determined that the amplitude exceeded 0.15 MPa, precluding the direct determination of stability. Therefore, it extracted the dominant frequency. However, since the dominant frequency exceeded 5 Hz, exceeding the system's upper limit for stable frequency, the system ultimately determined the current air pressure state to be unstable.

[0113] This dual-layered judgment structure effectively prevents false triggering of the leak detection module due to brief anomalies. If the air pressure fluctuates significantly but the frequency remains within the normal range, the system will not immediately identify it as unstable, thereby improving the overall system's accuracy and fault tolerance. This judgment logic, based on actual data patterns from engineering sites, has a solid theoretical basis and is feasible, making it directly applicable to the stability assessment phase of intelligent detection systems.

[0114] In step S16, if the air pressure is unstable, the surface feature image of the air pressure gun with unstable air pressure is input into a pre-trained image recognition model, and the leakage trace probability is output, including:

[0115] The image recognition model is trained by the ResNet-18 neural network model;

[0116] The data format of the surface feature image of the air pressure gun is converted through the input layer of the image recognition model to obtain an image in the form of a three-dimensional tensor;

[0117] Performing a convolution operation on the three-dimensional tensor image through the convolution layer of the image recognition model to extract local texture features of the image to obtain a first feature map;

[0118] Performing a nonlinear feature transformation on the first feature map through an activation function layer of the image recognition model to obtain an enhanced second feature map;

[0119] Downsampling the second feature map through a pooling layer of the image recognition model to obtain a feature map with a reduced size;

[0120] Performing feature integration and classification processing on the feature map after size reduction through the fully connected layer of the image recognition model, and outputting classification feature results related to the leakage state;

[0121] The output layer of the image recognition model performs probability calculation on the classification feature results and outputs the leakage trace probability.

[0122] It's worth noting that in this step, once the system determines that the pneumatic gun is experiencing unstable pressure based on the previous stage's determination process, it enters the image recognition-based leak detection phase. The core objective of this phase is to extract visual features and perform pattern recognition on the surface image of the currently operating pneumatic gun to determine whether there are any visible signs of a gas leak. To achieve this, the system pre-deploys an image recognition model trained on the ResNet-18 architecture. This model features a complete deep convolutional neural network architecture, extracting multi-layer texture features from the input image and performing pattern classification, ultimately outputting a probability value for the presence of a leak.

[0123] Before performing the image recognition task, the system first retrieves the surface image corresponding to the air pressure gun device currently experiencing unstable air pressure. This image is captured by an industrial camera mounted on the workstation, with a resolution of 1280×720 and three RGB channels. Before being input into the neural network, the original image undergoes data format conversion at the input layer. The input layer of the image recognition model does not accept image files directly; instead, it requires the image data to be converted into a three-dimensional tensor format—a numerical matrix structured by the number of channels, height, and width. This conversion process includes image scaling, normalization, and channel transformation. Specifically, the system scales the original image to a size of 224×224 and normalizes the pixel values ​​of the three RGB channels to a value between 0 and 1. This results in a three-dimensional floating-point tensor of size 3×224×224, which serves as the input to the neural network.

[0124] This tensor first passes through the model's first convolutional layer. The initial convolutional layer of ResNet-18 uses a 7×7 kernel with a stride of 2 to quickly capture low-level texture information such as edges and contours in the image. This layer outputs a 64-channel feature map. Subsequent batch normalization and ReLU activation further enhance the initial responsiveness of edge features. The primary function of the convolutional layer is to automatically extract spatial local features at different scales from the raw pixels, enabling the model to perform more complex pattern inference based on these features in subsequent layers.

[0125] After the initial convolutional layer, the image data enters the backbone of the model. ResNet-18 consists of four residual module stages, each of which consists of several basic residual units, and each residual unit contains two 3×3 convolutional layers. The introduction of these residual modules solves the problems of gradient vanishing and feature degradation that are prone to occur in traditional deep networks when the number of layers is deepened, enabling the model to extract deeper image semantic features while maintaining training stability. In each residual unit, after the feature map undergoes convolution operation, it will be element-by-element added with the input feature map of the unit to form a residual connection. This design allows the model to learn the "difference from the input" rather than completely reconstructing the output, which helps to focus on subtle feature differences in different areas of the image.

[0126] This multi-layer, cross-scale convolutional and residual connection structure is crucial for identifying leak traces. Leak traces appear as small, localized spots, blurred edges, or image textures such as aerosol deposits, with subtle morphology. Shallow convolutions can capture texture boundaries, while deep convolutions can incorporate broader contextual information to determine whether an area exhibits leak characteristics. ResNet-18 extracts spatial information from images at multiple depth levels, enabling the model to accurately distinguish leaking from non-leaking areas in image environments with complex backgrounds and diverse textures.

[0127] After each set of convolution operations, the model introduces a nonlinear activation function called ReLU. This function suppresses negative values ​​in the convolution output to zero, retaining only positive values. This operation not only improves the model's expressiveness but also increases the network's responsiveness to high-intensity features, enhancing the texture response in areas with potential leakage, thereby providing more discriminative feature maps for the subsequent discriminative layer.

[0128] To avoid wasting computational resources due to excessive dimensionality of feature maps during convolutional stacking, the model uses a maximum pooling operation between residual modules. The pooling layer downsamples the feature map output by the convolution using a fixed window, retaining the maximum response value in each local region. This operation helps compress the feature map size while enhancing feature invariance, making the model more robust to small geometric perturbations such as image shifts and rotations. This is crucial in leak detection scenarios, as the positional information of leak traces may shift slightly when the same air pressure gun is captured from different angles. The introduction of the pooling layer ensures that the model can accurately identify leak features under different shooting conditions.

[0129] After all convolutional and pooling layers complete, the model feeds the resulting low-dimensional feature map into a fully connected layer. This fully connected layer expands the previously extracted spatial feature vectors into a one-dimensional vector and weights all features to form a comprehensive representation for classification. In this step, the fully connected layer converts the feature map into a probability vector representing leaky and non-leakage states. Each element in the vector corresponds to a class, with larger values ​​indicating a higher probability for that class.

[0130] Finally, the model inputs this probability vector into the output layer for normalization. The output layer uses the Softmax function to map the output of the fully connected layer to a probability value between 0 and 1, where the maximum value represents the category most likely determined by the model. In this system, only two output categories are set: "leakage traces present" and "no leakage traces." The leakage trace probability output by the model is the percentage probability of the presence of leakage traces in the image. For example, an output of 0.87 indicates an 87% probability of the presence of leakage traces in the image. The system then compares this probability value with the set threshold. If it exceeds 0.75, the subsequent leakage amount calculation process is entered; otherwise, it is considered that no valid leakage signs are found in the image.

[0131] For example, during a tightening process, the air pressure at a certain workstation fluctuated significantly. The system extracted an image of the air pressure gun surface at that workstation and fed it into the image recognition model. The model processed the image and output a leak probability of 0.91, significantly exceeding the system's threshold of 0.75. Based on this, the system determined that the image contained a high probability of leaks and proceeded to the next stage: 3D flow field acquisition and leakage calculation.

[0132] In step S16, the surface feature image of the air pressure gun with unstable air pressure is input into the trained image recognition model, and the leakage trace probability is output, including:

[0133] The training process of the image recognition model includes:

[0134] Based on the supervised learning algorithm, the pre-prepared training set is optimized to obtain the optimized training set;

[0135] Inputting the optimized training set into the image recognition model and outputting the predicted leakage trace probability;

[0136] Calculate the difference between the predicted leakage trace probability and the actual leakage trace probability to obtain an output error value;

[0137] Based on the output error value, executing a back propagation algorithm to update load demand model parameters and optimize the image recognition model;

[0138] Repeat the above process and continuously optimize the load demand model until the output accuracy of the load demand on the training set reaches the preset output accuracy requirement, and then stop training.

[0139] In step S16, the image recognition model undergoes supervised training to ensure it can identify leaks on the pneumatic gun surface. The entire training process focuses on the degree of match between the input image and the ground-truth label. By continuously adjusting model parameters to reduce prediction error, the model achieves high recognition accuracy. This training process follows standard deep neural network training mechanisms, relying on a data-driven approach to gradually converge the model's recognition capabilities. This training was completed before system deployment.

[0140] The training phase primarily relies on a high-quality image training set. This training set's image data is derived from images of pneumatic gun surfaces captured at real-world work sites and in simulated laboratory environments. These images are captured using fixed industrial cameras or cameras mounted on the end of a robot. Multi-angle, multi-lighting, and multi-distance sampling strategies are employed during the capture process to ensure that the training images cover the differences in working conditions under different environments. For example, in a wind turbine flange bolt tightening experiment, researchers created an initial dataset of 3,000 images by creating abnormal air pressure fluctuations and intentionally creating tiny leaks. These images included 2,000 images showing no leaks and 1,000 images showing signs of leaks. All images were annotated by professionals based on actual test results, with labels categorized as either "leakage present" or "no leak."

[0141] After image annotation, the system optimizes the image training set. This optimization process includes unifying the image format, resizing it to 224×224 pixels, normalizing the pixels, and performing data augmentation. Data augmentation, including random image flipping, color perturbation, and cropping, is used to improve the model's generalization and avoid overfitting due to a single sample distribution. The optimized training set is divided into a training set and a validation set, typically in a ratio of 8 to 2. The training set is used to fit the model parameters, while the validation set is used to periodically evaluate the model's performance to prevent overfitting during training.

[0142] The training process uses supervised learning. The model receives input from a training set of images and outputs a probability value for the presence of a leak in each image. This probability value is a floating-point number between 0 and 1, expressing the model's confidence that the image belongs to the "leak" category. For example, if an image known to have a leak is input, the model's initial output value may be 0.42, indicating that the model is not currently correctly identifying the leak. The system compares this predicted value with the image's true label and calculates the error. This error is calculated based on the cross-entropy loss function. This loss function measures the degree of deviation between the model's prediction and the true label. Larger values ​​indicate less accurate predictions, while smaller values ​​indicate a closer match between the model's recognition results and the true situation.

[0143] The calculated error value is used to guide the adjustment of model parameters. The system uses the backpropagation algorithm to update the parameters of each layer of the model. The backpropagation algorithm uses the chain rule to calculate the contribution of each parameter in a neural network to the total error. In the current structure, the system starts from the output layer and calculates the partial derivative of the loss function with respect to the output of each layer. The gradient is propagated to the previous layer, and so on until it reaches the input of the first layer. The weight parameters in each layer are updated based on this partial derivative. The specific update method is to subtract the learning rate from the original parameter value and multiply it by the gradient value. The learning rate is a preset constant that controls the step size of each parameter update. A common value is 0.001.

[0144] In this neural network architecture, convolutional layers and fully connected layers all contain parameters that need to be updated. In convolutional layers, the weight matrices within the convolution kernels are updated, while in fully connected layers, the weights between connected nodes are updated. For example, in the first convolutional layer of the model, if a convolution kernel does not respond strongly to a specific leaky texture feature in a training sample, the system calculates its gradient through backpropagation and adjusts the kernel's weights to better respond to that feature in the next training run. This parameter iteration method based on error feedback continuously optimizes the model's ability to identify leaky images.

[0145] Each training cycle is called a training epoch. During a epoch, the system traverses all training set samples and performs a complete forward prediction and backpropagation process. The training process continues for multiple epochs until the system evaluates that the recognition accuracy of the model on the validation set meets the set standard, or the loss function changes by less than a set threshold over multiple consecutive epochs. Taking this system as an example, during a model training process, the training termination conditions are set as a validation set accuracy of at least 94% and a loss value that converges to less than 0.05. Ultimately, the model meets these conditions after the 28th epoch, the training process automatically stops, and the current parameters are saved as the final model.

[0146] For example, in one training experiment, the system's initial model's prediction values ​​for a set of leaking images were generally below 0.5, indicating inaccurate detection of leaks. As training rounds progressed, the error gradually decreased. By the 10th round, the prediction value for this set of images had risen to 0.63, and by the 20th round, it had risen to 0.81, ultimately reaching a final prediction value of 0.91, indicating that the model could detect the presence of leaks with a high degree of confidence.

[0147] In step S17, if the leakage trace probability is greater than a preset threshold, the three-dimensional flow field data of the leakage point is obtained, and the leakage amount is calculated based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount, including:

[0148] The three-dimensional flow field data includes: velocity vector, density, leakage start time, leakage end time and leakage cross-sectional area;

[0149] Calculated by the following formula:

[0150]

[0151] Where Q is the gas leakage volume, t1 is the leakage start time, t2 is the leakage end time, V is the velocity vector, p is the density, and A is the leakage cross-sectional area.

[0152] It's worth noting that in this step, after the system outputs the probability of leak traces through the image recognition model, if the probability value is greater than the preset threshold set by the system, the current air pressure gun is considered to have a high risk of leakage. To further verify the leakage situation and quantitatively assess the degree of leakage, the system enters the gas leakage calculation process. This process uses the three-dimensional flow field data of the leakage point and calculates the total amount of gas leakage within a specified time period based on the mass flux integral calculation method. This calculation step not only has recognition functions but also quantitative diagnostic capabilities, which can provide a core basis for on-site maintenance operations, equipment performance analysis, and energy consumption estimation.

[0153] The setting of the preset threshold is the trigger condition of this step. In the system, the leakage trace probability threshold is set to 0.75. This value comes from the experimental data statistics in the model verification phase. After the image recognition model training is completed, the model recognition accuracy is evaluated using an independent validation set. The measured results show that when the predicted probability is greater than 0.75, the model's recognition accuracy for leakage images reaches 96.4%, and the misjudgment rate is controlled within 3%. Therefore, in engineering practice, using 0.75 as the trigger threshold for judging leakage can take into account both the accuracy of the model and the stability of the system response. If the value is set too low, the system will frequently trigger the flow field acquisition and calculation modules, increasing the computational burden; if it is set too high, some cases where there are real leaks but the image features are not obvious will be missed. Therefore, selecting 0.75 as the balance point is a reasonable range determined after experimental verification and performance evaluation.

[0154] When the probability of a leak exceeds a set threshold, the system first calls the 3D flow field data module to extract gas flow data for the leak area corresponding to the current pneumatic gun within a specified time period. This data primarily includes five parameters: velocity vector, gas density, leak start time, leak end time, and leak cross-sectional area. Based on these five parameters, the system constructs a complete flow field integral formula input to accurately calculate the amount of gas leakage.

[0155] Velocity vector data is used to describe the direction and velocity of gas flow at each spatial location in the leakage area and is a vector value in three-dimensional space. The method of obtaining the velocity vector relies on a sensor array or a CFD simulation model built based on the principles of computational fluid dynamics. In industrial deployment, a flow velocity sensor array is often used to perform real-time sampling of the three-dimensional space near the leakage point. The arrangement of the sensors is customized according to the geometric shape of the air pressure gun. For example, nine-point gas flow velocity sensors are arranged in three dimensions on the surface of the gun body. Each sensor outputs flow velocity values ​​in three directions, which are combined to form a local velocity vector field. The system uses an interpolation algorithm to expand the data of limited measurement points to the entire leakage area to form complete velocity field data.

[0156] The density parameter is used to represent the mass of gas per unit volume at each location and is a key factor in calculating mass flow rate. The density acquisition method is based on the pressure-temperature combined measurement method. The system deploys temperature sensors and pressure sensors in the leakage area and uses the gas state equation to calculate the density value in real time. For example, the density of air at normal pressure and 25 degrees Celsius is approximately 1.18 kilograms per cubic meter. When the pressure is detected to drop to 0.9 atmospheres and the temperature rises to 30 degrees Celsius, the system calculates the density to be approximately 1.09 kilograms per cubic meter using the gas state formula. This density value changes dynamically with time and local airflow state. The system inputs it as a function parameter into the flow field integration process to achieve dynamic coupling of density.

[0157] The leak start time and leak end time represent the start and end moments of the leak event, respectively, and are used to limit the time boundaries of the integration. These two time points are automatically determined by the system. The start time is taken as the timestamp when the image recognition model first outputs a leak trace probability exceeding the set threshold, and the end time is taken as the time point after the system shuts down the current leak assessment process or the leak image recognition result recovers to below the threshold. In the system architecture, detailed timestamp information is recorded for each operation cycle, so the boundary of the leak period can be accurately locked through the log system, and the error is controlled within the sampling period. This design ensures that the integration time period is highly consistent with the actual time of the leak event, thereby improving the accuracy of the calculation.

[0158] The leakage cross-sectional area refers to the geometric area where the gas passes through the leakage site and is the core boundary of the spatial integration. This cross-sectional area is obtained through pneumatic gun structural modeling and image recognition-assisted calibration. Before the model is deployed, engineers establish a standard geometric model based on the configuration of the pneumatic gun tip and preset the structural gap area where leakage may occur. When the image recognition model determines that the probability of leakage is high, the system will further analyze the specific spatial position of the leakage feature area in the image and map the area back to the corresponding actual position in the geometric model. The system uses this position as the leakage cross-sectional area to input the integration process. Its area value unit is square meters, and the shape is set as a regular or irregular surface according to the model.

[0159] After obtaining the above five parameters, the system enters them into the mass flow integral formula to calculate the leakage. This formula consists of two integration steps. First, the instantaneous mass flux on the leakage cross section is spatially integrated to calculate the total mass outflow at each moment. Second, this result is integrated over the entire time period to obtain the cumulative leakage. The integration process is completed by the numerical integration module. The time step is consistent with the sensor sampling period, and the spatial step is set according to the velocity vector field interpolation density.

[0160] The calculation result is a scalar value, which represents the total mass of gas escaping from the leakage cross-sectional area within a specified time period. The unit of the result is kilograms, which reflects the actual physical quantity of the gas leakage. For example, in a leakage event, the system records the start time as 10:15:04 on May 12, 2024, the end time as 10:15:14, and the duration as 10 seconds. During this time period, the average velocity vector output by the flow sensor was 3.2 meters per second, the density value was 1.1 kilograms per cubic meter, and the area of ​​the leakage cross-sectional area was 5 square centimeters. The system substitutes the above data and calculates the leakage volume to be 0.0176 kilograms. This result is further used to generate a maintenance report, prompting the operator to perform a sealing inspection on the air pressure gun.

[0161] The following describes the working process of the present invention using a common scenario as an example. Figure 2 , which is Figure 1 Schematic diagram of the working scenario of the method.

[0162] The following combination Figure 1 The method flow shown is similar to Figure 2 The system structure shown here describes the actual working process of the present invention in detail, based on a typical pneumatic gun usage scenario during wind turbine tower flange connection operations. This scenario is representative: wind farm construction sites often feature a mix of bolt specifications, fluctuating air supply conditions, and high tightening precision requirements. Therefore, intelligent, closed-loop detection of pneumatic gun bolt installation anomalies is particularly important.

[0163] In step S11, the data acquisition module uses real-time sensors connected to the pneumatic gun to obtain the actual tightening force, air pressure fluctuation amplitude, air pressure fluctuation frequency, and an image of the pneumatic gun surface characteristics at that moment. The tightening force is output in real time by a torque sensor with a sampling frequency of 100Hz; air pressure parameters are acquired by a combination of a pressure sensor and an acoustic sensor; and image data is captured by an industrial camera mounted on the workbench. All data is timestamped and uploaded to the processing module for subsequent dynamic analysis.

[0164] In step S12, the tightening force deviation module compares the actual tightening force values ​​obtained with the preset tightening force values ​​set by the process, calculating a time-varying tightening force deviation sequence. In this example, the bolt currently being used is M16, corresponding to a preset tightening force of 60 Nm. The sensor continuously records actual torque values ​​of 58, 59, 55, and 61, which are processed to form a deviation sequence such as -2, -1, -5, and +1. This process is a one-dimensional time series calculation operation and forms the basis for subsequent discrete identification.

[0165] In step S13, the adaptive clustering module uses built-in deviation direction judgment rules based on the aforementioned deviation sequence to determine whether each data point is overtightened or undertightened, and constructs a deviation direction distribution map. This map reflects the deviation distribution trend along the time axis. The system then performs adaptive clustering on this distribution map, identifying dense, directional deviation point clusters and extracting the number of abnormal point clusters. For example, in this operation, the system detected three relatively dense undertightening point clusters, each containing more than eight consecutive undertightening data points, initially indicating a potential abnormal tightening trend.

[0166] In step S14, the installation anomaly module determines the proportion of these abnormal discrete point clusters. During the current operation sampling cycle, a total of 100 deviation points were obtained, of which 24% were abnormal point clusters. The system's set abnormality threshold is 20%, which has been exceeded. Therefore, the system marks this bolting operation as "bolt installation abnormality" and notifies the operator in real time on the operation interface.

[0167] Entering step S15, the air pressure stabilization module further assesses the air pressure status of the air pressure gun at the abnormal station. The system first detects that the pressure fluctuation is 0.19 MPa, exceeding the system's upper limit of 0.15 MPa. The system then extracts the dominant frequency of this pressure signal, which is 6.1 Hz, exceeding the system's allowable range of 1 to 5 Hz. Based on this, the system determines that the current pressure is unstable and requires further inspection to determine if a leak is the cause.

[0168] In step S16, the leak probability module starts the image recognition process. The system calls the previously collected image of the pneumatic gun surface and inputs it into the trained ResNet-18 image recognition model. The model performs format conversion through the input layer, extracts local features through the convolution layer, enhances texture response through the activation layer, and reduces the dimension through the pooling layer. After entering the fully connected layer for feature integration, the output layer finally calculates the leak probability. In this example, the model outputs a probability value of 0.84, which exceeds the system-set threshold of 0.75. Therefore, the system believes that there is a leak in the image.

[0169] Finally, in step S17, the gas leakage module initiates the three-dimensional flow field calculation process based on the image recognition results. The system uses the velocity array sensor network deployed at the device location to obtain the velocity vector field of the current leak area. Simultaneously, the gas density is inferred using the combined pressure and temperature sensors. The system records the leak start time as 10:15:06 and the current time as the end time. Based on the cross-sectional area of ​​the pneumatic gun (preset to 4.5 square centimeters through modeling), the system substitutes these parameters into the integral calculation module, ultimately outputting a leakage value of 0.023 kg, which is recorded in real time in the job log.

[0170] In summary, through the above steps, the system completes the closed-loop control of the entire process from data acquisition, deviation analysis, abnormality judgment, air pressure status confirmation, image recognition to leakage calculation. Figure 2 The functional modules shown respond in sequence to complete Figure 1 Each processing node from S11 to S17 forms a detection system with clear structure, rigorous logic and information exchange, which significantly improves the efficiency and accuracy of leakage anomaly identification during the installation operation of pneumatic gun bolts.

[0171] Reference Figure 2 A second embodiment of the present invention provides a system for detecting abnormal bolt installation of a pneumatic gun, comprising:

[0172] A data acquisition module is used to obtain the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process;

[0173] a tightening force deviation module, configured to perform dynamic deviation calculation on the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence;

[0174] An adaptive clustering module is used to perform multi-dimensional discrete analysis on the tightening force deviation sequence based on an adaptive clustering algorithm to obtain the number of abnormal discrete point clusters;

[0175] An installation anomaly module is used to mark the bolt installation as abnormal if the proportion of the abnormal discrete point clusters exceeds a preset threshold;

[0176] An air pressure stabilization module is used to compare the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with a preset amplitude threshold and frequency range respectively to determine whether the air pressure is stable;

[0177] The leakage trace probability module is used to input the surface feature image of the air pressure gun with unstable air pressure into the trained image recognition model and output the leakage trace probability if the air pressure is unstable;

[0178] An image recognition module is used to input a surface feature image of the air pressure gun with unstable air pressure into a pre-trained image recognition model to obtain an image recognition result if the air pressure is unstable;

[0179] The gas leakage module is used to obtain the three-dimensional flow field data of the leakage point if the probability of the leakage trace is greater than a preset threshold, and calculate the leakage amount based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount.

[0180] It should be noted that the device for detecting abnormal installation of a pneumatic gun bolt provided in an embodiment of the present invention is used to execute all the process steps of the method for detecting abnormal installation of a pneumatic gun bolt in the above embodiment. The working principles and beneficial effects of the two correspond one to one, and therefore will not be repeated here.

[0181] An embodiment of the present invention further provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a program for detecting abnormal installation of a pneumatic gun bolt. When the processor executes the computer program, the steps of the above-mentioned embodiments of the method for detecting abnormal installation of a pneumatic gun bolt are implemented, such as Figure 1 Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned device embodiments are realized, such as the gas leakage module.

[0182] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of implementing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device.

[0183] The electronic device may be a computing device such as a desktop computer, notebook, PDA, or smart tablet. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the aforementioned components are merely examples of electronic devices and do not constitute a limitation of the electronic device. The electronic device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0184] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of the electronic device and connects various parts of the entire electronic device using various interfaces and lines.

[0185] The memory can be used to store the computer programs and / or modules. The processor implements the various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and accessing the data stored in the memory. The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as a sound playback function or an image playback function); the data storage area may store data generated based on the use of the mobile phone (such as audio data, a phone book, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0186] If the module / unit integrated into the electronic device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention can implement all or part of the process steps in the above-mentioned method embodiments by using a computer program to instruct the relevant hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. It should be noted that the content of the computer-readable medium can be appropriately increased or decreased based on the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media does not include electric carrier signals and telecommunication signals.

[0187] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive effort.

[0188] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for detecting abnormal installation of a pneumatic gun bolt, characterized in that: include: Obtain the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process; Performing dynamic deviation calculation on the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence; Based on the adaptive clustering algorithm, a multi-dimensional discrete analysis is performed on the tightening force deviation sequence to obtain the number of abnormal discrete point clusters; If the proportion of the abnormal discrete point clusters exceeds a preset threshold, it is marked as a bolt installation abnormality; Comparing the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with a preset amplitude threshold and frequency range, respectively, to determine whether the air pressure is stable; If the air pressure is unstable, the surface feature image of the air pressure gun with unstable air pressure is input into the trained image recognition model, and the probability of leakage traces is output; If the leakage trace probability is greater than a preset threshold, the three-dimensional flow field data of the leakage point is obtained, and the leakage amount is calculated based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount; If the air pressure is unstable, the surface feature image of the air pressure gun with unstable air pressure is input into the trained image recognition model to output the probability of leakage traces, including: The image recognition model is trained by the ResNet-18 neural network model; Performing data format conversion on the surface feature image of the pneumatic gun through the input layer of the image recognition model to obtain an image in the form of a three-dimensional tensor; Performing a convolution operation on the image in the form of a three-dimensional tensor through a convolution layer of the image recognition model to extract local texture features of the image and obtain a first feature map; Performing a nonlinear feature transformation on the first feature map through an activation function layer of the image recognition model to obtain an enhanced second feature map; Downsampling the second feature map through a pooling layer of the image recognition model to obtain a feature map with a reduced size; Performing feature integration and classification processing on the feature map after size reduction through the fully connected layer of the image recognition model, and outputting classification feature results related to the leakage state; Performing probability calculation on the classification feature results through the output layer of the image recognition model and outputting the leakage trace probability; The training process of the image recognition model includes: Based on the supervised learning algorithm, the pre-prepared training set is optimized to obtain the optimized training set; Inputting the optimized training set into the image recognition model and outputting the predicted leakage trace probability; Calculate the difference between the predicted leakage trace probability and the actual leakage trace probability to obtain an output error value; Based on the output error value, executing a back propagation algorithm to update load demand model parameters and optimize the image recognition model; Repeat the above process and continuously optimize the load demand model until the output accuracy of the load demand on the training set reaches the preset output accuracy requirement, and then stop training.

2. The method for detecting abnormal bolt installation of a pneumatic gun according to claim 1, characterized in that: The step of performing dynamic deviation calculation on the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence includes: The tightening force deviation sequence is calculated using the following formula: ； in, is the tightening force deviation sequence, is the actual tightening force value, The preset tightening force value.

3. The method for detecting abnormal bolt installation of a pneumatic gun according to claim 1, characterized in that: The adaptive clustering algorithm is used to perform multi-dimensional discrete analysis on the tightening force deviation sequence to obtain the number of abnormal discrete point clusters, including: Based on a preset deviation direction judgment rule, determining the deviation direction of the tightening force deviation sequence, and obtaining a deviation direction distribution map; Based on the adaptive clustering algorithm, anomaly identification is performed on the deviation direction distribution map to obtain the number of abnormal discrete point clusters.

4. The method for detecting abnormal bolt installation of a pneumatic gun according to claim 1, characterized in that: The air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun according to the abnormal bolt installation are compared with the preset amplitude threshold and frequency range respectively to determine whether the air pressure is stable, including: When the air pressure fluctuation amplitude does not exceed the preset amplitude threshold, it is determined that the air pressure is in a stable state; When the air pressure fluctuation amplitude exceeds a preset amplitude threshold, the air pressure fluctuation frequency is further compared with a preset frequency range. If the air pressure fluctuation frequency is not within the preset frequency range, it is determined that the air pressure is in an unstable state.

5. The method for detecting abnormal bolt installation of a pneumatic gun according to claim 1, characterized in that: If the probability of the leakage trace is greater than a preset threshold, the three-dimensional flow field data of the leakage point is obtained, and the leakage amount is calculated based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount, including: The three-dimensional flow field data includes: velocity vector, density, leakage start time, leakage end time and leakage cross-sectional area; Calculated by the following formula: ； in, is the gas leakage, is the leakage start time, is the leakage end time, is the velocity vector, is the density, is the leakage cross-sectional area.

6. A system for detecting abnormal installation of a pneumatic gun bolt, for implementing the method for detecting abnormal installation of a pneumatic gun bolt according to any one of claims 1 to 5, characterized in that: include: A data acquisition module is used to obtain the actual tightening force value, air pressure fluctuation amplitude, air pressure fluctuation frequency and surface feature image of the air pressure gun during the bolt installation process; a tightening force deviation module, configured to perform dynamic deviation calculation on the actual tightening force value and the preset tightening force value to obtain a tightening force deviation sequence; An adaptive clustering module is used to perform multi-dimensional discrete analysis on the tightening force deviation sequence based on an adaptive clustering algorithm to obtain the number of abnormal discrete point clusters; An installation anomaly module is used to mark the bolt installation as abnormal if the proportion of the abnormal discrete point clusters exceeds a preset threshold; An air pressure stabilization module is used to compare the air pressure fluctuation amplitude and air pressure fluctuation frequency of the air pressure gun with the abnormal bolt installation with a preset amplitude threshold and frequency range respectively to determine whether the air pressure is stable; The leakage trace probability module is used to input the surface feature image of the air pressure gun with unstable air pressure into the trained image recognition model and output the leakage trace probability if the air pressure is unstable; The gas leakage module is used to obtain the three-dimensional flow field data of the leakage point if the probability of the leakage trace is greater than a preset threshold, and calculate the leakage amount based on the three-dimensional flow field data of the leakage point to obtain the gas leakage amount.

7. An electronic device, characterized in that: The invention comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for detecting abnormal installation of a bolt of a pneumatic gun according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for detecting abnormal installation of a pneumatic gun bolt according to any one of claims 1 to 5.

Images