A bolt tightening abnormality diagnosis method based on historical data and deep learning

Abstract

This invention discloses a bolt tightening anomaly diagnosis method based on historical data and deep learning, belonging to the field of intelligent manufacturing technology. It solves the problems of excessive reliance on human expert experience, inconsistent subjective diagnostic standards, lagging and low coverage of anomaly identification, and lack of actionable diagnostic results in existing technologies. This invention constructs a dynamically updated statistical baseline model, establishes an objective deviation quantification scoring system for multi-dimensional tightening parameters, and introduces a hybrid model of autoencoder-long short-term memory network to learn the deep temporal features of normal tightening patterns. Finally, it employs an information entropy weighted fusion strategy to achieve collaborative decision-making between the two independent models, significantly improving the accuracy and robustness of anomaly detection. Furthermore, this invention enables the direct conversion of intelligent diagnostic results into production execution, ultimately constructing an autonomous, continuously evolving bolt tightening quality assurance system with zero human intervention.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a method for diagnosing bolt tightening anomalies based on historical data and deep learning. Background Technology

[0002] Threaded connections, as the most widely used fastening method in mechanical assembly, directly affect the safety, reliability, and service life of the entire machine product through their tightening quality. On modern large-scale production lines, hundreds of thousands of bolt tightening data points are generated daily. The tightening process of a single bolt involves multiple dimensions of parameters, including torque, angle, time, speed, assembly tool number, workpiece identification, and operator information. The data characteristics exhibit complex features such as high dimensionality, strong temporal sequence, parameter coupling, and nonlinear fluctuations. Traditional tightening quality control mainly relies on the following two models: The first type is the manual sampling inspection mode. Quality control personnel periodically sample finished products from the production line and re-inspect them using a manual torque wrench, or rely on experienced torque experts to periodically review the tightening curve data exported from the controller. This mode has significant lag (usually several hours to several days), making real-time interception impossible; and the inspection coverage is extremely low (usually less than 1%), allowing a large number of potentially defective bolts to flow into subsequent processes or to the customer; more importantly, it relies too heavily on personnel experience, and the subjective differences in the judgment standards of different experts lead to poor consistency and traceability in quality judgment. With the continuous rise in labor costs and the increasing complexity of products, this mode can no longer meet the requirements of high-quality manufacturing.

[0003] The second type is the threshold-based judgment mode. Some advanced production lines preset fixed thresholds such as upper and lower torque limits and angle thresholds in the tightening controller, triggering an alarm when real-time data exceeds these limits. However, the quality of bolt tightening is not determined solely by a single point value, but rather by the comprehensive characteristics of the entire tightening curve (such as yield point, slope change, and fluctuation pattern). The fixed threshold mode cannot identify hidden risks such as "within the threshold but with abnormal curve shape" (e.g., jamming during tightening or early signs of thread stripping). Furthermore, the threshold setting itself still relies on engineer experience and lacks a dynamic update mechanism, making it difficult to adapt to systemic disturbances such as material batch fluctuations, tool wear, and changes in ambient temperature, resulting in persistently high false alarm and false negative rates.

[0004] In recent years, with the development of Industrial Internet of Things (IIoT) and big data technologies, some research has attempted to introduce machine learning algorithms for bolt tightening anomaly detection. For example, supervised learning classification models are used to train labeled qualified / unqualified samples, or traditional statistical process control (SPC) methods are employed to monitor the mean torque. However, these solutions still have significant drawbacks: supervised learning relies on a large amount of manually labeled data, while bolt tightening anomaly samples are scarce and labeling is extremely costly, limiting the model's generalization ability; traditional SPC methods only monitor a single statistic, ignoring the rich information in high-dimensional time-series data, and are insufficient for identifying complex fault modes. More importantly, existing technical solutions generally lack interpretability and operability of diagnostic results. When the system alarms, on-site operators cannot understand the root cause of the anomaly, let alone obtain specific corrective action guidance, leading to a disconnect between the intelligent system and production execution, significantly reducing its practical application value.

[0005] Therefore, a method for diagnosing bolt tightening anomalies based on historical data and deep learning is needed. Summary of the Invention

[0006] To address the problems of excessive reliance on human expert experience, inconsistent subjective diagnostic standards, lagging and low coverage of anomaly identification, and lack of actionable diagnostic results in existing technologies, this invention provides a bolt tightening anomaly diagnosis method based on historical data and deep learning. It establishes an objective deviation quantification scoring system for multi-dimensional tightening parameters by constructing a dynamically updated statistical baseline model. Furthermore, it introduces a hybrid model of autoencoder-long short-term memory network to learn the deep temporal characteristics of normal tightening patterns. Finally, it employs an information entropy weighted fusion strategy to achieve collaborative decision-making between the two independent models, significantly improving the accuracy and robustness of anomaly detection. Simultaneously, by mapping diagnostic results to a structured knowledge graph, this invention can automatically generate natural language solutions for non-technical personnel, including temporary measures, troubleshooting sequences, and parameter adjustment suggestions. This enables the direct conversion of intelligent diagnostic results into production execution, ultimately constructing an autonomous, continuously evolving bolt tightening quality assurance system with zero human intervention. The specific technical solution is as follows: A method for diagnosing bolt tightening anomalies based on historical data and deep learning includes the following steps: Multi-source heterogeneous time-series data of bolt tightening process is captured in real time. The raw data is cleaned, outliers are removed, and dimensions are standardized to obtain a standardized tightening dataset. ; The dataset D is automatically clustered into multiple monitoring subsets according to potential failure modes. A sliding time window mechanism is used to calculate the statistical feature matrix F for each subset in real time. The statistical features include mean, standard deviation, skewness and kurtosis. A purely data-driven statistical baseline model is constructed based on historical compliance data to calculate the overall mean of each tightening parameter. with standard deviation Establish dynamic confidence intervals And construct the objective offset scoring function: in, These are real-time measured values; For the first j Objective deviation score for each parameter; A hybrid model of autoencoder-long short-term memory network was constructed using TensorFlow. The statistical feature matrix F was input, and the model was trained unsupervised using historical normal data. The deep learning anomaly score was calculated using the reconstruction error. in An adaptive threshold is calculated based on the median error and absolute median difference of the reconstruction based on historical data. To reconstruct the root mean square error; Calculate the statistical baseline score S o b and deep learning score The information entropy is used to dynamically weight and fuse the entropy values ​​to obtain the final diagnostic score. in For dynamic fusion weights, when When the value is below the threshold τ, an anomaly is identified and a risk point matrix R is generated.

[0007] Preferably, the multi-source heterogeneous timing data includes torque value, rotation angle, rotation speed, timestamp, tool number, workpiece serial number, and controller error information.

[0008] Preferably, the monitoring subset includes a qualified steady-state set, a set with high torque, a set with low torque, a set with out-of-tolerance angle, and a set with time-series fluctuations, and each column of the statistical feature matrix F is a statistical feature vector of a monitoring subset.

[0009] Preferably, the dynamic confidence interval The objective offset score Take each parameter The minimum value is used as the overall score.

[0010] Preferably, the adaptive threshold value is as follows: in, This represents the absolute deviation of the median.

[0011] Preferably, the values ​​for the dynamic fusion weights are as follows: in, For the entropy of the statistical model, Entropy for deep learning models.

[0012] Preferably, the following steps are also included: The risk point matrix R is mapped to a failure mode-response knowledge graph, and structured solution suggestions are automatically generated and pushed to the production site terminal through the PyQT front-end interface.

[0013] A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the bolt tightening anomaly diagnosis method based on historical data and deep learning as described above.

[0014] A processor for running a program, wherein the program executes the bolt tightening anomaly diagnosis method based on historical data and deep learning as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention first abandons the traditional approach of relying on expert experience to set thresholds. By constructing a dynamic statistical baseline model based on massive amounts of historical qualified data, it uses an objective deviation scoring function to quantify and adaptively update quality standards, completely eliminating the problem of inconsistent judgments caused by subjective factors. Second, this invention innovatively uses a hybrid model of unsupervised autoencoder-long short-term memory network for tight time series analysis. It learns deep patterns using only normal samples, avoiding the high-cost manual annotation required for supervised learning, and accurately identifying hidden risks of "within the threshold but with abnormal morphology." Its output deep learning score achieves objective anomaly measurement through adaptive thresholds. Most importantly, this invention achieves complementary advantages between statistical models and deep learning models through an information entropy weighted fusion mechanism. Dynamic weights ensure that decisions always favor the more reliable model. At the same time, it automatically generates natural language solution suggestions by combining a fault mode-response knowledge graph, upgrading the traditional "information silo" that only alarms without guidance into a complete closed loop from diagnosis to execution. This enables non-technical personnel to respond and handle issues quickly, significantly reducing quality costs and reliance on experts. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] In one embodiment of the present invention, a method for diagnosing bolt tightening anomalies based on historical data and deep learning is provided, such as... Figure 1 As shown, it includes the following steps: Step 1: Capture multi-source heterogeneous time-series data during the bolt tightening process in real time, clean the raw data, remove outliers and standardize the dimensions to obtain a standardized tightening dataset D; This step retrieves raw tightening data from the production line data source. In actual production, the complete assembly process of a bolt is recorded and stored, including but not limited to the following fields: work order number, workpiece serial number, bolt unique identifier, tightening program number, tightening tool (high-precision tightening gun) number, workstation number, operator number, tightening start timestamp, sampling frequency, torque timing array, rotation angle timing array, speed timing array, and the final status code returned by the controller (e.g., qualified, high torque, low torque). This data constitutes the raw multi-source heterogeneous timing data.

[0023] The process of obtaining the standardized tightening dataset D is as follows: S101: Outliers are identified using the Z-score statistical method: The mean μᵀ and standard deviation σᵀ are calculated for the torque time series array T = [t1, t2, ..., tᵤ]. Data points that satisfy |tᵢ - μᵀ|>3σᵀ are marked as outliers, and they are removed only when a point and its two adjacent points both satisfy the Z-score condition. After removal, linear interpolation is used to fill the gaps.

[0024] S102: Convert all physical quantities to SI units (torque: N·m, angle: °). Subsequently, the Z-score normalization method is used to normalize the time series of each parameter, converting all parameter time series into dimensionless sequences with a mean of 0 and a standard deviation of 1.

[0025] After the above processing, each original tightening record is converted into a high-quality standardized dataset. ,in, ,in These are the standardized timing arrays for torque, angle, and speed, respectively. This is a metadata vector containing information such as workstations and tools. All dᵢ are entered into a real-time data cache queue for subsequent steps.

[0026] Step 2: The dataset D is automatically clustered into multiple monitoring subsets according to potential failure modes. A sliding time window mechanism is used to calculate the statistical feature matrix F for each subset in real time. The statistical features include mean, standard deviation, skewness and kurtosis. This step introduces a dynamic classification management mechanism after data preprocessing, which automatically divides the standardized dataset into different monitoring subsets based on its potential failure mode characteristics, thereby achieving modular and refined data management.

[0027] For example, the steps are as follows: S201: Preliminary clustering based on K-means++: For each record, the torque time series T' and angle time series A' are used to construct a two-dimensional feature point set according to their time index correspondence. The central moments of the point set are calculated as the clustering feature vectors, including: first moment (mean vector), second moment (elements of the covariance matrix), third moment (skewness vector), and total curve length (sum of Euclidean distances between points). Each record is ultimately represented as a 12-dimensional feature vector fᵢ. During algorithm initialization, 100 labeled samples each of three categories—known acceptable, high torque, and low torque—are extracted from the historical database (only used for initializing cluster centers; no labeled data is needed for subsequent model operation). The K-means++ strategy is used to select initial cluster centers, and the number of clusters k=5, corresponding to the following five failure modes: acceptable steady-state mode, high torque mode, low torque mode, angle deviation mode, and time-series fluctuation anomaly mode.

[0028] S201: Sliding Time Window Real-Time Classification Monitoring: An independent circular buffer is established for each subset, and a First-In-First-Out (FIFO) mechanism is used to manage data. The sliding window width w is set to 100 records, meaning each subset retains a maximum of the 100 most recent records of the same type. Whenever a new record... Upon arrival, the system first calculates its 12-dimensional feature vector. Quantity, through calculation Assign the cluster to the nearest subset based on its Euclidean distance from each cluster center. Subsequently, real-time statistical analysis was performed on the current data within this subset to extract the following key features: Mean vector μ m The average torque level of this pattern is obtained by averaging all torque values ​​in the subset. Standard deviation vector σ m : Calculate the degree of fluctuation in torque and angle timing to quantify the stability of the mode; Skewness γ m : To measure the asymmetry of torque distribution and identify whether there is a one-sided tailing phenomenon; Kurtosis κ m It reflects the sharpness of the torque distribution and detects whether there is abnormal concentration or dispersion.

[0029] The four statistics mentioned above constitute a 4×3 characteristic matrix F. m (4 statistics × 3 parameter types) This matrix is ​​updated every 10 seconds and serves as the input for subsequent diagnostic models.

[0030] Step 3: Construct a data-driven statistical baseline model based on historical qualified data, and calculate the overall mean of each tightening parameter. with standard deviation Establish dynamic confidence intervals And construct an objective offset scoring function: in, These are real-time measured values; This is the objective deviation score for the j-th parameter.

[0031] Step 4: Construct an autoencoder-long short-term memory hybrid model using TensorFlow, input the statistical feature matrix F, train it unsupervised with historical normal data, and calculate the deep learning anomaly score using the reconstruction error. in An adaptive threshold is calculated based on the median error and absolute median difference of the reconstruction based on historical data. To reconstruct the root mean square error; in, This represents the absolute deviation of the median.

[0032] Step 5: Calculate the statistical baseline score S o b and deep learning score The information entropy is used to dynamically weight and fuse the entropy values ​​to obtain the final diagnostic score. in For dynamic fusion weights, when When the value is below the threshold τ, an anomaly is identified and a risk point matrix R is generated. in, For the entropy of the statistical model, Entropy for deep learning models.

[0033] In summary, this invention, by constructing a purely objective statistical baseline model and an unsupervised deep learning model, and employing a dynamic fusion strategy of information entropy, thoroughly solves the core pain points of traditional bolt tightening quality control, such as strong subjectivity, large lag, low coverage, and poor operability. Practical application results show that this method can significantly improve the anomaly detection rate, reduce quality costs, and empower non-technical personnel, providing an innovative and feasible technical solution for threaded connection quality assurance in the context of intelligent manufacturing.

[0034] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0035] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0036] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0037] If the integrated unit 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 technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for diagnosing bolt tightening anomalies based on historical data and deep learning, characterized in that, Includes the following steps: Multi-source heterogeneous time-series data of bolt tightening process is captured in real time. The raw data is cleaned, outliers are removed, and dimensions are standardized to obtain a standardized tightening dataset. ; The dataset D is automatically clustered into multiple monitoring subsets according to potential failure modes. A sliding time window mechanism is used to calculate the statistical feature matrix F for each subset in real time. The statistical features include mean, standard deviation, skewness and kurtosis. A purely data-driven statistical baseline model is constructed based on historical compliance data to calculate the overall mean of each tightening parameter. with standard deviation Establish dynamic confidence intervals And construct the objective offset scoring function: in, These are real-time measured values; For the first j Objective deviation score for each parameter; A hybrid model of autoencoder-long short-term memory network was constructed using TensorFlow. The statistical feature matrix F was input, and the model was trained unsupervised using historical normal data. The deep learning anomaly score was calculated using the reconstruction error. in An adaptive threshold is calculated based on the median error and absolute median difference of the reconstruction based on historical data. To reconstruct the root mean square error; Calculate the statistical baseline score S o b and deep learning score The information entropy is used to dynamically weight and fuse the entropy values ​​to obtain the final diagnostic score. in For dynamic fusion weights, when When the value is below the threshold τ, an anomaly is identified and a risk point matrix R is generated.

2. The bolt tightening anomaly diagnosis method based on historical data and deep learning according to claim 1, characterized in that, Multi-source heterogeneous timing data includes torque value, rotation angle, speed, timestamp, tool number, workpiece serial number, and controller error information.

3. The bolt tightening anomaly diagnosis method based on historical data and deep learning according to claim 1, characterized in that, The monitoring subset includes a qualified steady-state set, a set with high torque, a set with low torque, a set with out-of-tolerance angle, and a set with time-series fluctuations. Each column of the statistical feature matrix F is a statistical feature vector of a monitoring subset.

4. The bolt tightening anomaly diagnosis method based on historical data and deep learning according to claim 1, characterized in that, The dynamic confidence interval The objective offset score Take each parameter The minimum value is used as the overall score.

5. The bolt tightening anomaly diagnosis method based on historical data and deep learning according to claim 1, characterized in that, The adaptive threshold values ​​are as follows: in, This represents the absolute deviation of the median.

6. The bolt tightening anomaly diagnosis method based on historical data and deep learning according to claim 1, characterized in that, The values ​​for the dynamic fusion weights are as follows: in, For the entropy of the statistical model, Entropy for deep learning models.

7. The bolt tightening anomaly diagnosis method based on historical data and deep learning according to claim 1, characterized in that, It also includes the following steps: The risk point matrix R is mapped to a failure mode-response knowledge graph, and structured solution suggestions are automatically generated and pushed to the production site terminal through the PyQT front-end interface.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the bolt tightening anomaly diagnosis method based on historical data and deep learning as described in any one of claims 1 to 7.

9. A processor, characterized in that, The processor is used to run a program, wherein the program executes the bolt tightening anomaly diagnosis method based on historical data and deep learning as described in any one of claims 1 to 7.

Images