Bolt tightening angle accessibility prediction system for breakaway torque wrench

Abstract

The application discloses a bolt tightening rotation angle accessibility prediction system for a tripping type torque wrench and relates to the technical field of industrial intelligence.The system is used to solve the problem of rotation angle accessibility prediction under nonlinear friction interference.Firstly, the system obtains a dynamic tightening sequence and a static working condition context through a wrench communication interface, uses entity embedding and a sliding window algorithm to construct a composite heterogeneous input tensor, and realizes multi-source data fusion; subsequently, local spatial features and time-dependent features are extracted in parallel through a convolutional neural network and a long short-term memory network, a high-dimensional spatiotemporal feature vector sequence is generated through cascaded fusion to capture hidden layer friction states; then, an attention weight matrix is calculated based on a self-attention mechanism, a context vector is generated, an estimated end point torque and a confidence probability value are solved, and rotation angle accessibility is quantitatively evaluated; finally, a confidence is mapped to a discrete state code by using a probability decoding function, and an output is encapsulated, and a closed-loop system from data perception to state prediction is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial intelligence, in particular to a bolt tightening angle accessibility prediction system for a tripping torque wrench. BACKGROUND

[0002] With the development of precision in high-end equipment manufacturing and rail transportation industry, bolt connection as a key fastening process, its assembly quality directly determines the operation safety of the equipment. In the tightening process of torque method and angle method, the bolt needs to continue to rotate a specific angle to generate the expected preload after reaching the fitting torque. In such high safety requirement assembly scenarios, ensuring that each bolt can smoothly reach the target angle (i.e. angle accessibility) within the specified torque window is the core indicator to guarantee the mechanical performance and fatigue life of the connected pair.

[0003] However, the existing tightening quality control method mainly relies on the post-threshold judgment, i.e. only after the tightening action is completed, whether the final torque and angle values are within the tolerance range is compared. This method cannot predict the future state trend in real time during the tightening process, and has a lag. Traditional physical model monitoring methods usually assume that the friction coefficient is a linear constant, but in actual working conditions, the friction coefficient between threads is affected by factors such as uneven lubrication distribution, micro roughness difference and rotation speed fluctuation, showing highly nonlinear time-varying characteristics. This makes it difficult for existing technologies to distinguish between normal friction fluctuations and potential mechanical jamming, and it is easy to produce false judgments when the bolt has not reached the target angle due to the torque exceeding the limit in advance, or it cannot identify the risk of false fastening although the angle is reached, causing the waste rate to rise and the production rhythm to be wasted. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a bolt tightening angle accessibility prediction system for a tripping torque wrench, which solves the problems of the above background art.

[0005] To achieve the above object, the bolt tightening angle accessibility prediction system for the tripping type torque wrench comprises the following modules: a data tensorization module, which is used to obtain a multi-dimensional dynamic tightening sequence and a static working condition context of a current operation, map the static working condition context to a high-dimensional embedding vector through an entity embedding algorithm, transform the multi-dimensional dynamic tightening sequence into a dynamic time sequence feature matrix using a sliding window algorithm, concatenate the high-dimensional embedding vector and the dynamic time sequence feature matrix in the channel dimension, and construct a composite heterogeneous input tensor; a feature mapping module, which is used to input the composite heterogeneous input tensor into a convolutional neural network branch to perform convolution operation, extract a spatial feature map of a local waveform, perform a recurrent operation through a long short-term memory network branch to calculate a time-dependent feature vector sequence of torque evolution with angle, perform cascade fusion on the spatial feature map and the time-dependent feature vector sequence to generate a high-dimensional spatio-temporal feature vector sequence representing the hidden layer state of the current friction working condition; an attention reasoning module, which is used to calculate an attention weight matrix of each time step in the high-dimensional spatio-temporal feature vector sequence based on a self-attention mechanism, perform weighted summation operation on the high-dimensional spatio-temporal feature vector sequence through the attention weight matrix to generate a context vector, input the context vector into a fully connected layer to calculate an estimated end point torque value, calculate a confidence probability value of quantifying the current process angle accessibility based on the current torque growth rate; and a prediction state mapping module, which is used to map the confidence probability value to a preset discrete state space, determine the working condition state through a probability decoding function to generate an angle accessibility state code, and encapsulate the angle accessibility state code as a digital prediction message for output.

[0006] Further, the specific process of obtaining the multi-dimensional dynamic tightening sequence and the static working condition context of the current operation, mapping the static working condition context to a high-dimensional embedding vector through an entity embedding algorithm, and transforming the multi-dimensional dynamic tightening sequence into a dynamic time sequence feature matrix using a sliding window algorithm is as follows: a discrete feature dictionary containing bolt material grade, target torque threshold and vehicle platform identifier is established, discrete category indexes in the static working condition context are mapped to a high-dimensional dense embedding vector of continuous numerical type through table lookup operation, and the high-dimensional dense embedding vector is subjected to normalization processing; real-time data streams in the operation process are intercepted at a fixed sampling frequency through a torque wrench communication interface, torque values and angle values containing the current time and historical time of a preset length are selected to form an original time sequence segment, a first-order differential feature sequence containing angular velocity and torque change rate is generated by performing difference operation on the original time sequence segment, and the original time sequence segment and the first-order differential feature sequence are stacked in the feature dimension to generate a dynamic time sequence feature matrix.

[0007] Further, the specific process of splicing the high-dimensional embedding vector and the dynamic time-series feature matrix in the channel dimension to construct a composite heterogeneous input tensor is as follows: analyzing the length of the time step dimension of the dynamic time-series feature matrix, performing a broadcast replication operation along the time axis on the high-dimensional dense embedding vector, expanding to generate a static context matrix isomorphic to the dynamic time-series feature matrix in the time dimension; performing a parallel splicing operation on the static context matrix and the dynamic time-series feature matrix in the feature channel dimension to construct a composite heterogeneous input tensor.

[0008] Further, the specific process of inputting the composite heterogeneous input tensor into the convolutional neural network branch to perform convolution operation, extracting the spatial feature map of the local waveform, and through the long short-term memory network branch to perform recurrent operation, calculating the time-dependent feature vector sequence of the torque evolution with angle is as follows: in the convolutional neural network branch, a set of preset one-dimensional convolution kernels are used to perform sliding convolution filtering operation on the time axis of the composite heterogeneous input tensor, to extract the local morphological features representing the mutation of the fitting point and the microscopic jamming in the tightening curve, and the maximum pooling layer is used to reduce the dimension of the convolution output to generate the local spatial feature map; in the long short-term memory network branch, the composite heterogeneous input tensor is input into the recurrent unit in time step sequence, the memory cell state in the unit is recursively updated through the forgetting gate and the input gate mechanism, the long-term dependence and nonlinear trend in the cumulative process of the torque with the rotation angle are captured, and the hidden layer output of the last time step is extracted as the time-dependent feature vector sequence.

[0009] Further, the specific process of performing cascade fusion on the spatial feature map and the time-dependent feature vector sequence to generate a high-dimensional spatio-temporal feature vector sequence representing the current friction condition hidden layer state is as follows: performing time dimension alignment operation on the local spatial feature map and the time-dependent feature vector sequence; constructing a feature fusion layer to perform cascade splicing on the aligned local spatial feature map and the time-dependent feature vector sequence in the vector dimension, and inputting the spliced joint vector sequence into the full connection layer to perform linear transformation and nonlinear activation operation, weighting and combining the local morphological features and global evolution trend features step by step through the network weights, to generate a high-dimensional spatio-temporal feature vector sequence that can implicitly represent the actual friction coefficient and contact stiffness state of the current bolted joint pair.

[0010] Further, the specific process of calculating the attention weight matrix of each time step in the high-dimensional spatiotemporal feature vector sequence based on the self-attention mechanism and generating the context vector by performing a weighted sum operation on the high-dimensional spatiotemporal feature vector through the attention weight matrix is as follows: three linear projection layers of a query matrix, a key matrix and a value matrix are constructed to map the high-dimensional spatiotemporal feature vector sequence to different subspaces to generate a query vector sequence, a key vector sequence and a value vector sequence; the dot product similarity of the query vector sequence and the key vector sequence is calculated, a scaling factor is introduced to scale the dot product result, and the attention weight matrix representing the feature correlation strength of each time step is generated by processing through a Softmax normalization function; the attention weight matrix and the value vector sequence are subjected to matrix multiplication operation to identify and aggregate the key features of the yield point and the friction mutation point in the tightening process, and generate a context vector containing global dependency information.

[0011] Further, the specific process of passing the context vector into the fully connected layer to calculate the estimated end point torque value and calculating the confidence probability value quantifying the current process corner reachability based on the current torque growth rate is as follows: first-order differential operation is performed on the end data of the multi-dimensional dynamic tightening sequence to extract the instantaneous torque growth rate at the current time, the instantaneous torque growth rate and the context vector are spliced and fused in the feature dimension to construct an enhanced feature vector containing physical trend constraints; a multi-task prediction head is constructed, a linear regression operation is performed on the enhanced feature vector through a regression branch fully connected layer to calculate the estimated end point torque value; and in parallel, a nonlinear transformation is performed on the enhanced feature vector through a classification branch fully connected layer, the transformation result is mapped to a continuous interval of zero to one through a Sigmoid activation function, and a confidence probability value quantifying the current process corner reachability is generated.

[0012] Further, the prediction state mapping module comprises the following steps: dividing the probability interval boundary corresponding to the reachable state, the blocked state and the overload risk state in the preset discrete state space, and constructing a probability-state mapping table; performing numerical comparison between the calculated confidence probability value and the probability interval boundary through a probability decoding function to determine the probability interval hit by the current working condition, indexing the probability-state mapping table to obtain the corresponding working condition category label; generating a corresponding hexadecimal corner reachability state code according to the working condition category label, assembling the corner reachability state code and the estimated end point torque value into a data payload according to a preset communication protocol format, and serializing the digital prediction message after adding a frame header and a check bit.

[0013] The present application has the following beneficial effects:

[0014] (1) The bolt tightening angle accessibility prediction system for the tripping type torque wrench effectively solves the problems of multi-source heterogeneous data fusion difficulty and insufficient nonlinear friction feature extraction by constructing a composite heterogeneous input tensor and adopting a double-flow feature mapping mechanism. The data tensorization module uses an entity embedding algorithm to map the static working condition context into a dense vector, and splices it with the dynamic time series feature matrix in the channel dimension, so that the neural network can perceive the physical property constraints of the work object. The feature mapping module adopts a parallel architecture, uses a convolutional neural network branch to capture the microscopic morphological mutations of local waveforms, and uses a long short-term memory network branch to capture the long-term dependence trend of the angle evolution of the torque. This design enables the system to accurately extract high-dimensional spatiotemporal features representing the hidden state of the current working condition in a complex nonlinear friction environment, significantly improving the model's generalization ability and the robustness of feature expression under different working conditions.

[0015] (2) The bolt tightening angle accessibility prediction system for the tripping type torque wrench realizes forward-looking and accurate prediction and digital output of the tightening process angle accessibility through a self-attention mechanism and a probability state mapping strategy. The attention deduction module uses a self-attention mechanism to dynamically calculate the attention weights at different time steps, automatically focusing on key moments such as yield points or friction mutation points, avoiding dilution of effective information in long sequence data. By combining the regression and classification capabilities of the fully connected layer, the system can not only estimate the final torque value, but also output quantitative confidence probability values, thereby continuously assessing the risk of whether the angle is accessible. The prediction state mapping module further converts the continuous probability values into deterministic discrete state codes through a probability decoding function, and encapsulates them into standardized digital prediction messages. This mechanism converts ambiguous trend prediction into executable digital instructions.

[0016] Of course, implementing any product of the present application does not necessarily require all the advantages described above to be achieved simultaneously. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 Flowchart of the bolt tightening angle accessibility prediction system for the tripping type torque wrench of the present application. DETAILED DESCRIPTION

[0018] The bolt tightening angle accessibility prediction system for the tripping type torque wrench of the present application solves the problem that traditional methods cannot accurately predict whether the bolt can safely reach the target angle in advance under nonlinear friction interference and complex working conditions.

[0019] The general idea of the scheme in the embodiments of the present application is as follows: first, the system obtains the real-time dynamic sequence uploaded by the wireless torque wrench and the static parameters of the management system through the data quantization module, and converts them into a composite heterogeneous input tensor that can be processed by a computer by using entity embedding and sliding window technology; second, the spatial local features and time-dependent features of the data are extracted by parallel convolutional neural networks and long short-term memory networks respectively, and are cascaded and fused to represent the hidden layer friction state; third, the key time steps are weighted by using the self-attention mechanism, and the estimated end torque and the confidence probability of the reachable angle are calculated by the full connection layer; finally, the calculation results are mapped to the standard angle reachability state code by probability decoding, and output in the form of digital message to the terminal for visual warning or device control.

[0020] Please refer to Figure 1 The embodiments of the present application provide a technical scheme: a bolt tightening angle reachability prediction system for a tripping torque wrench, comprising the following modules: a data tensorization module for obtaining a multi-dimensional dynamic tightening sequence and a static working condition context of the current operation, mapping the static working condition context into a high-dimensional embedding vector by an entity embedding algorithm, transforming the multi-dimensional dynamic tightening sequence into a dynamic time sequence feature matrix by a sliding window algorithm, splicing the high-dimensional embedding vector and the dynamic time sequence feature matrix in the channel dimension, and constructing a composite heterogeneous input tensor; a feature mapping module for inputting the composite heterogeneous input tensor into a convolutional neural network branch to perform convolution operation, extracting the spatial feature map of the local waveform, performing cyclic operation through a long short-term memory network branch to calculate the time-dependent feature vector sequence of the torque evolution with angle, and performing cascaded fusion on the spatial feature map and the time-dependent feature vector sequence to generate a high-dimensional spatio-temporal feature vector sequence representing the current friction working condition hidden layer state; an attention reasoning module for calculating the attention weight matrix of each time step in the high-dimensional spatio-temporal feature vector sequence based on the self-attention mechanism, performing weighted summation operation on the high-dimensional spatio-temporal feature vector sequence through the attention weight matrix to generate a context vector, inputting the context vector into a full connection layer to calculate the estimated end torque value, and calculating the confidence probability value of the current process angle reachability based on the current torque growth rate; a prediction state mapping module for mapping the confidence probability value to a preset discrete state space, determining the working condition state by a probability decoding function, generating an angle reachability state code, and outputting the angle reachability state code as a digital prediction message.

[0021] In this embodiment, the data quantization module is mainly responsible for converting the heterogeneous data sensed by the torque wrench into a unified mathematical expression that can be processed by the neural network. This module obtains the multi-dimensional dynamic tightening sequence through the wireless communication interface, which refers to the torque and angle values collected by the built-in sensors in real time during the operation of the torque wrench. Especially for the torque wrench, the sequence implies the vibration characteristics of the mechanical mechanism action (such as the "click" moment of disengagement). The disengagement action can be regarded as a physical marker of the attachment point. The system uses the detected disengagement vibration characteristics as the starting anchor point for angle measurement and predicts the reachability of the angle by analyzing the subsequent rotation data. At the same time, the static working condition context, such as the vehicle platform and bolt grade, is obtained, which does not change with time. The entity embedding algorithm used here is a technology that maps discrete symbols or categories to continuous numerical vectors, which can convert text labels such as CR400AF vehicles into high-dimensional embedding vectors that can be operated by computers. At the same time, the sliding window algorithm discretizes the continuous data stream into a dynamic time series feature matrix by cutting a fixed length of data slices on the time axis. The technical role of this module is to solve the problem of mismatch between the dimensions of static attribute data and dynamic process data uploaded by the torque wrench. By mathematical splicing in the channel dimension, a composite heterogeneous input tensor containing all the operation information is constructed, laying a data foundation for the operation of the subsequent deep learning model. The feature mapping module is mainly responsible for extracting deep abstract features that can represent the physical friction state from the original data. This module inputs the composite heterogeneous input tensor into a parallel computing architecture, where the convolutional neural network branch uses convolution kernels to perform sliding operations on the data, and is specifically used to extract spatial feature maps of local waveforms, i.e., to identify small wave peaks, troughs, or abrupt shapes in the data, which usually correspond to mechanical attachment or microscopic jamming moments. At the same time, the long short-term memory network branch uses a recursive structure to process sequence data, maintains the accumulation of historical information through memory cells, and calculates the time-dependent feature vector sequence of the torque evolution with the angle, i.e., captures the long-term trend of torque growth. The concatenation of the two is to splice the local morphology in space and the long-term trend in time in terms of feature dimension, which generates a high-dimensional spatio-temporal feature vector sequence that can implicitly represent the current complex friction working condition, solving the technical problem that a single model cannot consider both local anomalies and global trends. The attention inference module calculates the attention weight matrix based on the self-attention mechanism, which is an algorithm that measures the correlation strength between different time steps in the sequence. It can automatically identify key moments such as yield points or friction coefficient mutation points and give them higher computational weights. The context vector generated by the weighted sum operation is essentially a numerical vector that condenses the essence of the entire sequence. Subsequently, the fully connected layer calculates the estimated end torque value and the confidence probability value based on this vector through linear transformation and nonlinear activation function in parallel.The confidence probability value here is a value between zero and one, which quantitatively represents the possibility of safely reaching the target rotation angle under the current working condition. The technical role of this module is to overcome the problem that effective information in long sequence data is easily diluted, and to realize accurate quantitative evaluation of rotation angle accessibility. The prediction state mapping module is mainly responsible for converting the probability output of the model into deterministic instructions that can be executed by the industrial system. This module maps continuous confidence probability values to a pre-defined discrete state space through a probability decoding function, i.e., a pre-set logical decision rule or threshold comparison algorithm. The discrete state space refers to a pre-defined set of working states of the system, such as accessibility, obstruction, overload, etc. The rotation angle accessibility state code generated on this basis is a computer-recognizable hexadecimal or binary code. The technical role of this module is to complete the conversion from uncertain probability reasoning to deterministic logical control, and by encapsulating the state code into a standard format of digital prediction message, the lower computer or interactive terminal can directly parse and execute the corresponding alarm or blocking action, realizing seamless integration of intelligent algorithms and industrial control protocols.

[0022] Specifically, the specific process of obtaining the multi-dimensional dynamic tightening sequence and the static working condition context of the current job, mapping the static working condition context to a high-dimensional embedding vector through an entity embedding algorithm, and transforming the multi-dimensional dynamic tightening sequence into a dynamic time series feature matrix using a sliding window algorithm is as follows: a discrete feature dictionary containing bolt material grade, target torque threshold, and vehicle platform identifier is established, the discrete category index in the static working condition context is mapped to a continuous numerical high-dimensional dense embedding vector through table lookup operation, and the high-dimensional dense embedding vector is normalized; the real-time data stream in the job process is intercepted through the torque wrench communication interface at a fixed sampling frequency, the torque and rotation angle values containing the current time and the historical time of a preset length are selected to form an original time series segment, the original time series segment is subjected to difference operation to generate a first-order differential feature sequence containing angular velocity and torque change rate, and the original time series segment and the first-order differential feature sequence are stacked in the feature dimension to generate a dynamic time series feature matrix.

[0023] In this embodiment, the data quantization module first processes the static context, which refers to the physical property constraint data that remains unchanged in a single tightening task. Since the bolt material, vehicle platform, and other information are represented as discrete symbolic identifiers in the raw data, they do not have mathematical continuity and cannot be directly used as input to the neural network. Therefore, a discrete feature dictionary is established, and a lookup table is used to convert these discrete features into numerical values that can be understood by the computer. The lookup table is essentially an index-based matrix addressing process that maps each discrete category to a high-dimensional dense embedding vector, which can capture the potential semantic relationships between different process parameters. To eliminate the dimensional differences between different features and accelerate network convergence, the extracted vectors need to be normalized, and the core calculation process is as follows: ; wherein, : represents the high-dimensional dense embedding vector after normalization; : represents the pre-constructed entity embedding lookup table, which is obtained by pre-training using unsupervised learning methods on large-scale historical process data; : represents the index value of the target discrete category in the static context; : represents the mean statistical quantity of all feature vectors in the lookup table; : represents the standard deviation statistical quantity of all feature vectors in the lookup table; : represents a numerical stability constant to prevent division by zero errors, with a value of 1e-6. Subsequently, the module processes the multi-dimensional dynamic tightening sequence using a sliding window algorithm. The sliding window algorithm is a data processing technique that extracts local time series segments by moving a fixed-length window along the time axis to construct the original time series segments. To further explore the dynamic characteristics of the tightening process, the system performs a difference operation on the torque and angle data in the original time series segments to generate a first-order differential feature sequence. The first-order differential feature sequence contains angular velocity and torque rate, which can sensitively reflect the instantaneous changes in bolt contact stiffness. The calculation process for generating a single-step dynamic vector containing differential features is as follows: ; wherein, : represents the dynamic feature vector at the t-th time step; : represents the real-time torque value at the current sampling time; : represents the real-time angle value at the current sampling time; : represents the torque value at the previous sampling time; : represents the angle value at the previous sampling time; : represents the fixed sampling time interval of the sensor. By continuously arranging the calculated dynamic feature vectors along the time axis, a dynamic time series feature matrix containing position information and velocity information is generated.

[0024] Specifically, the specific process of splicing the high-dimensional embedding vector and the dynamic time sequence feature matrix in the channel dimension to construct a composite heterogeneous input tensor is as follows: analyzing the length of the time step dimension of the dynamic time sequence feature matrix, performing a broadcast replication operation along the time axis on the high-dimensional dense embedding vector, and expanding to generate a static context matrix isomorphic to the dynamic time sequence feature matrix in the time dimension; performing a parallel splicing operation on the static context matrix and the dynamic time sequence feature matrix in the feature channel dimension to construct a composite heterogeneous input tensor.

[0025] In the embodiment, the core of constructing the composite heterogeneous input tensor is to solve the inconsistency of static data and dynamic data in the dimension structure. The dynamic time sequence feature matrix has a time dimension, which describes the evolution of the process; while the high-dimensional dense embedding vector is a static global attribute and does not change with time. In order to fuse the two, the length of the time step dimension of the dynamic time sequence feature matrix must be analyzed first, that is, the total number of sampling points contained in the sequence is obtained. Then, a broadcast replication operation is performed on the high-dimensional dense embedding vector. The broadcast replication operation is a tensor transformation technique, which aims to expand a low-dimensional tensor into a high-dimensional tensor by repeating and copying data along a specific axis, so that it is aligned in geometry with the target tensor. The calculation process of generating the isomorphic static context matrix is as follows: ; wherein, : represents the static context matrix generated after expansion, which is consistent with the dynamic sequence in the time axis; : represents the normalized high-dimensional dense embedding vector generated in the foregoing step; : represents the time step length of the dynamic time sequence feature matrix; Broadcast: represents the replication function performed along the time axis. Finally, a parallel splicing operation in the feature channel dimension is performed. The feature channel dimension is the dimension in the data tensor used to distinguish different physical attributes (such as torque channel, angle channel, vehicle type channel). The parallel splicing operation merges the static context matrix and the dynamic time sequence feature matrix side by side in the feature channel, thereby generating the final composite heterogeneous input tensor. The construction process of the tensor is as follows: ; wherein, : represents the constructed composite heterogeneous input tensor, which is used as the input of the subsequent neural network; : represents the dynamic time sequence feature matrix stacked by the original time sequence segment and the first-order differential feature sequence; represents the matrix splicing operation performed in the feature channel dimension. Through the above steps, the neural network can simultaneously perceive the current static process constraint when processing the dynamic tightening data of each time step, thereby realizing the deep fusion of heterogeneous data.

[0026] Specifically, the composite heterogeneous input tensor is input into the convolutional neural network branch to perform convolution operation, extract the spatial feature map of the local waveform, and through the long short-term memory network branch to perform the specific process of the recurrent operation, the calculation of the time-dependent feature vector sequence of the moment of force evolution with angle as follows: in the convolutional neural network branch, a set of preset one-dimensional convolution kernels are used to perform sliding convolution filtering operation on the time axis of the composite heterogeneous input tensor, to extract the local morphological features in the tightening curve that represent the mutation of the fitting point and the microscopic jamming, and the maximum pooling layer is used to reduce the dimension of the convolution output to generate the local spatial feature map; in the long short-term memory network branch, the composite heterogeneous input tensor is input into the recurrent unit in time step sequence, the memory cell state in the unit is recursively updated through the forgetting gate and the input gate mechanism, the long-term dependence and nonlinear trend in the cumulative process of the torque with the rotation angle are captured, and the hidden layer output at the last time step is extracted as the time-dependent feature vector sequence.

[0027] In the embodiment, a double-flow parallel architecture is adopted to process the input composite heterogeneous input tensor to adapt to the complex mechanical characteristics of the tripping torque wrench. First, in the convolutional neural network branch, one-dimensional convolution kernels are used to perform sliding convolution filtering operation. The core purpose of this step is to capture the high-frequency local features in the tightening data, especially the mechanical vibration waveform generated when the tripping mechanism acts. Since the arrival of the fitting point is usually accompanied by instantaneous mutation of the torque slope during the bolt tightening process, and the microscopic jamming or slip tooth is manifested as jagged fluctuations on the curve, convolution operation can effectively identify these non-stationary signals through local receptive fields. The maximum pooling layer then down-samples the convolution output, reduces the data dimension while retaining the most significant feature responses, and enhances the model's robustness to subtle time shifts. The core calculation process of the convolution operation is as follows: ; wherein, : represents the feature response value output by the kth convolution kernel at time step t; : represents the input composite heterogeneous input tensor; : represents the weight matrix of the kth one-dimensional convolution kernel, which is automatically optimized and updated in the training process through the backpropagation algorithm; represents the time window size of the convolution kernel, which is usually set to cover a length of 5 to 10 sampling points; : represents the bias term parameter corresponding to the kth convolution kernel; represents the rectified linear unit activation function, which is used to introduce a nonlinear factor to eliminate negative response signals. In parallel, in the long short-term memory network branch, the system processes the long-term dependence of the data. Since the final pretightening force of the bolt depends not only on the current torque but also on the energy accumulation in the entire tightening history process, a simple fully connected network is difficult to remember this long sequence information. The long short-term memory network adjusts the information flow by introducing a gating mechanism. The forget gate determines how much historical state to retain, and the input gate determines how much current information to update. This mechanism enables the model to simulate the cumulative effect of stress in the physical process. The recursive update of the memory cell state and the calculation process of the hidden layer output are as follows: ; ; wherein, : represents the time-dependent feature vector output at sequence time step : represents the memory cell state at time step : represents the memory cell state at time step : represents the memory cell state at the previous time step : represents the activation output value of the forget gate, which determines the retention proportion of historical information : represents the activation output value of the input gate, which determines the writing proportion of current new information : represents the candidate memory state at the current time : represents the activation output value of the output gate : represents the Hadamard product, which is a point-by-point multiplication operation of matrix elements : represents the hyperbolic tangent activation function, which is used to normalize the state value to the interval of negative one to positive one.

[0028] Specifically, the specific process of performing cascade fusion on the spatial feature map and the sequence of time-dependent feature vectors to generate a high-dimensional spatiotemporal feature vector sequence representing the current friction condition hidden layer state is as follows: performing a time dimension alignment operation on the local spatial feature map and the sequence of time-dependent feature vectors; constructing a feature fusion layer to perform cascade splicing on the aligned local spatial feature map and the sequence of time-dependent feature vectors in the vector dimension, and inputting the spliced joint vector sequence into a fully connected layer to perform linear transformation and nonlinear activation operation. Through network weights, local morphological features and global evolution trend features are weighted and combined step by step to generate a high-dimensional spatiotemporal feature vector sequence that can implicitly represent the current bolted joint actual friction coefficient and contact stiffness state.

[0029] In this embodiment, the key to generating high-dimensional spatio-temporal feature vector sequence lies in the effective fusion of local microscopic features and global macroscopic trends. Since the output of the convolutional neural network branch after the pooling operation may have a time dimension smaller than the original sequence length, a time dimension alignment operation is first needed, usually using upsampling interpolation or linear projection to synchronize the spatial feature map with the time-dependent feature vector sequence on the time axis. Subsequently, a feature fusion layer is constructed for cascading concatenation, which is equivalent in physical meaning to binding the local waveform state (such as the presence of oscillation) at a certain time with the global evolution state (such as whether in the elastic or plastic region) at that time. Finally, a linear transformation and nonlinear activation are performed through a fully connected layer, and the model automatically balances the importance of the two types of features using learnable weights. For example, in unstable friction coefficient conditions, the model may give higher weights to local waveform features to identify stiction; while in normal tightening conditions, it focuses more on global trends. The feature vector generated by this weighted combination, although abstract in mathematics, implicitly contains information about physical parameters such as actual friction coefficient and contact stiffness that are difficult to measure directly. The fusion calculation process is as follows: ; wherein, : represents the feature value of the fusion-generated high-dimensional spatio-temporal feature vector sequence at time step t; : represents the feature vector of the local spatial feature map after time alignment at time step t; : represents the feature vector of the time-dependent feature vector sequence output by the long short-term memory network at time step t; : represents the vector concatenation operation, which combines two vectors into a longer joint vector; : represents the weight matrix of the fully connected layer, used for feature reorganization and dimension mapping of the joint vector; : represents the bias vector of the fully connected layer; : represents the parameterized rectified linear unit activation function, used to enhance the nonlinear expression ability of feature fusion and avoid feature degradation.

[0030] Specifically, the specific process of calculating the attention weight matrix of each time step in the high-dimensional spatiotemporal feature vector sequence based on the self-attention mechanism and generating the context vector by performing a weighted sum operation on the high-dimensional spatiotemporal feature vector through the attention weight matrix is as follows: three linear projection layers of a query matrix, a key matrix and a value matrix are constructed to map the high-dimensional spatiotemporal feature vector sequence to different subspaces to generate a query vector sequence, a key vector sequence and a value vector sequence; the dot product similarity of the query vector sequence and the key vector sequence is calculated, a scaling factor is introduced to scale the dot product result, and the Softmax normalization function is used to process the dot product result to generate an attention weight matrix representing the correlation strength of each time step feature; the matrix multiplication operation is performed on the attention weight matrix and the value vector sequence to identify and aggregate the key features of the yield point and the friction mutation point in the tightening process, and a context vector containing global dependency information is generated.

[0031] In the embodiment, the core of the attention inference module is to dynamically evaluate the importance of each time step in the high-dimensional spatiotemporal feature vector sequence to the final prediction result using the self-attention mechanism. Since the tightening using a torque wrench is a continuous physical process, some specific moments (the moment when the wrench is about to reach the tripping threshold, the moment when the bolt reaches the yield point or the moment when microslip occurs) contain much more information than the moments of smooth rotation. In order to capture these key information, the module first constructs three linear projection layers, which are essentially three independent weight matrices, respectively used to generate a query vector sequence, a key vector sequence and a value vector sequence. The query vector represents the information being sought at the current moment, the key vector represents the index feature of the queried moment, and the value vector contains the actual content information. The mapping process is as follows: ; wherein, represents the generated query vector sequence matrix; represents the generated key vector sequence matrix; represents the generated value vector sequence matrix; represents the high-dimensional spatiotemporal feature vector sequence generated in the foregoing step, i.e., the input of the attention mechanism; represents the learnable weight matrix of the query projection layer, which is optimized by the back propagation algorithm to extract the features for matching; represents the learnable weight matrix of the key projection layer; The learned weight matrix of the value projection layer. Then, the dot product similarity between the query vector sequence and the key vector sequence is calculated. The dot product operation can measure the alignment of two vectors in direction, and the larger the value, the closer the association between the two. In order to prevent the dot product result from being too large and causing gradient disappearance, a scaling factor is introduced for numerical scaling, usually taking the square root of the inverse of the vector dimension. The Softmax normalization function then converts the scaled value into a probability distribution with a sum of one, i.e. the attention weight matrix. Each element in this matrix represents the degree of attention that the model should give to the data at other time points when processing the data at a certain time. The calculation process is shown in the following formula: ; wherein, : represents the calculated attention weight matrix; : represents the transpose matrix of the key vector sequence matrix; : represents the dimension of the feature vector, which is used as the denominator of the scaling factor to stabilize the numerical range; Softmax: represents the normalization exponential function used to convert the original score into a probability form of weight. Finally, the attention weight matrix is multiplied with the value vector sequence. This step is essentially a weighted sum process, in which the model extracts and aggregates those features that are most valuable to the corner reachability judgment (such as the torque climb feature before the tripping mechanism acts) according to the calculated weights, and suppresses irrelevant noise, thereby generating a context vector containing global dependency information. This vector is a highly condensed representation of the entire tightening process. The calculation process is shown in the following formula: ; wherein, : represents the generated context vector, which integrates the information of all key time steps in the sequence; : represents the attention weight matrix calculated in the previous step; represents the value vector sequence matrix.

[0032] Specifically, the context vector is passed into a fully connected layer to calculate the estimated final torque value, and the specific process of calculating the confidence probability value of the current process corner reachability based on the current torque growth rate is as follows: a first-order differential operation is performed on the end data of the multi-dimensional dynamic tightening sequence to extract the instantaneous torque growth rate at the current time, the instantaneous torque growth rate is spliced and fused with the context vector in the feature dimension to construct an enhanced feature vector containing physical trend constraints; a multi-task prediction head is constructed, a linear regression operation is performed on the enhanced feature vector through a regression branch fully connected layer to calculate the estimated final torque value; and in parallel, a nonlinear transformation is performed on the enhanced feature vector through a classification branch fully connected layer, and the transformed result is mapped to a continuous interval of zero to one through a Sigmoid activation function to generate the confidence probability value quantifying the current process corner reachability.

[0033] In this embodiment, after generating the context vector, the system needs to transform it into specific physical prediction values and risk assessment indicators. In order to improve the physical consistency of the prediction, the module first extracts the instantaneous torque growth rate at the current time. This is because in the later stage of tightening, the rising rate of torque directly reflects the stiffness state of the bolt, which is an important physical basis for judging whether it will soon exceed the limit. The explicit physical feature is spliced and fused with the implicit context vector extracted by the neural network in the feature dimension to construct an enhanced feature vector, so that the subsequent prediction contains both the global pattern of historical data and the constraint of the current physical trend. Then, the module constructs a multi-task prediction head to handle the regression task and the classification task respectively. In the regression branch, the fully connected layer acts as a linear converter to map the high-dimensional enhanced feature vector to a one-dimensional continuous numerical value, which is the estimated end-point torque value. This value represents the final torque that the bolt will bear when reaching the target angle. The calculation process is as follows: ; wherein, : represents the calculated estimated end-point torque value; : represents the weight matrix of the regression branch fully connected layer, which is obtained by training the least square error loss function; : represents the context vector generated in the previous step; : represents the instantaneous torque growth rate at the current time; : represents the vector splicing operation; : represents the bias term of the regression branch. In parallel, in the classification branch, the system needs to assess the risk of reaching the angle. The fully connected layer performs a nonlinear transformation on the enhanced feature vector, compressing it into a scalar, and then maps it to the interval of zero to one through the Sigmoid activation function. This output value is the confidence probability value, and the closer the value is to 1, the greater the possibility of reaching the angle, and the closer the value is to 0, the greater the risk of not reaching or being blocked. The calculation process is as follows: ; wherein, : represents the confidence probability value quantifying the reachability of the current process angle; : represents the weight matrix of the classification branch fully connected layer, which is obtained by training the cross-entropy loss function; : represents the enhanced feature vector after splicing, i.e. ; : represents the bias term of the classification branch; : represents the natural constant.

[0034] Specifically, the predicted state mapping module comprises the following steps: dividing the probability interval boundaries corresponding to the reachable state, the blocked state and the overload risk state in the preset discrete state space, constructing a probability-state mapping table; performing numerical comparison between the calculated confidence probability value and the probability interval boundaries through a probability decoding function, determining the probability interval hit by the current working condition, indexing the probability-state mapping table to obtain the corresponding working condition category label; generating a corresponding hexadecimal corner reachability state code according to the working condition category label, assembling the corner reachability state code and the estimated endpoint torque value into a data payload according to a preset communication protocol format, and serializing the digital prediction message after adding a frame header and a check bit.

[0035] In the embodiment, the core task of the predicted state mapping module is to convert the continuous probability value output by the neural network into discrete control instructions that can be directly executed by the trip-type wireless digital display torque wrench and its matching terminal. First, the system presets a discrete state space, which is a set of a limited number of mutually exclusive working states divided according to the tightening process requirements, usually including a reachable state representing normal, a blocked state representing mechanical jamming, and an overload risk state representing that the torque may exceed the standard. In order to realize the mapping from probability to state, a probability-state mapping table needs to be constructed, which defines the distribution range of different states on the probability axis. The setting of the probability interval boundary (i.e. threshold value) is crucial, and through the numerical comparison and indexing operation of the probability decoding function, the category of the current working condition is determined. The logical decision process is as follows: ; wherein, : represents the working condition category label of the decision output, which is used to indicate the final predicted state of the current bolt connection; : represents the confidence probability value output by the previous module; : represents the lower threshold value of the safe confidence for determining the reachable state; : represents the upper threshold value of the fault confidence for determining the blocked state; CLASS_OK, CLASS_RISK, CLASS_FAIL: respectively represent the reachable, risk and blocked category indexes in the mapping table. Regarding the key parameters and The determination method is as follows: a receiver operating characteristic curve (ROC curve) analysis method is used, all possible threshold points are traversed on the verification data set, and a cut-off point is selected as a boundary value, so that the recall rate of the model meets a preset safety standard (such as 99.9%) and the false positive rate is the lowest. Subsequently, the system performs state coding and message serialization. According to the obtained working condition category label, the system looks up to generate a corresponding hexadecimal corner reachability state code (for example, 0xA1 represents reachability, and 0xEE represents obstruction), which is a kind of machine language level instruction code. In order to return the state code and the estimated final end torque value to the digital screen or the sound and light alarm module of the wrench, it is necessary to assemble it into a data payload according to a preset communication protocol format (such as Modbus or a custom TCP protocol), and add a frame header for synchronization and a check bit for error correction. So that the intelligent torque wrench in the field can prompt the operator to realize closed-loop control according to the prediction result. This serialization process packs the dispersed information into a continuous binary stream, that is, a digital prediction message. The packaging calculation process is as follows: ; wherein, : represents the finally generated digital prediction message, which is the final data product output of the GO6N system; : represents a fixed frame header sequence specified by the communication protocol, used for data frame synchronization at the receiving end; represents a bit-level concatenation splicing operation of binary data, represents a state code conversion function that converts category labels into fixed hexadecimal instruction codes; : represents the estimated final end torque value; Quantize: represents a numerical quantization function that converts the floating-point torque value into a fixed-point integer to save bandwidth; CRC16: represents a cyclic redundancy check function, used to calculate the check code of the data payload to ensure the integrity of the transmission.

[0036] In summary, the present application has at least the following effects:

[0037] The bolt tightening corner reachability prediction system for the tripping type torque wrench realizes deep fusion and extraction of local microscopic waveform features and global time-dependent trends by constructing a complex heterogeneous input tensor of static context and dynamic sequence, using a dual-flow parallel architecture of convolutional neural networks and long short-term memory networks. At the same time, the self-attention mechanism is introduced to dynamically focus on key physical moments such as yield points and friction mutation points, and the continuous confidence deduction result is converted into a deterministic digital state code combined with a probability state mapping strategy, so as to upgrade the traditional passive threshold detection to an active forward-looking process prediction, significantly improving the identification accuracy and operation reliability of process risks such as false fastening and mechanical jamming in the precision assembly scene.

[0038] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In a software embodiment, the methods can be tangibly embodied in a machine-readable storage medium having stored thereon instructions that can be used to program a processing system to perform the methods. The program can be embedded in a computer program product which can be a storage medium or a memory, for example, a semiconductor memory, a floppy diskette, a CD-ROM, or an optical storage device, for example, a DVD or a Blu-ray Disc®. The storage medium can be loaded into a processing system, for example, a computer, to cause the processing system to perform the methods. The storage medium can be a computer readable storage medium.

[0039] The present application is described in reference to the drawings, which are as follows: Figure 1 Each flow or multiple flows and / or blocks Figure 1 Means for carrying out the function specified in the flow or flows and / or blocks

[0040] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flow Figure 1 Each flow or multiple flows and / or blocks Figure 1 Means for carrying out the function specified in the flow or flows and / or blocks

[0041] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 Each flow or multiple flows and / or blocks Figure 1 Means for carrying out the function specified in the flow or flows and / or blocks

[0042] While preferred embodiments of the application have been described, modifications and alterations thereto will occur to those skilled in the art upon reading the preceding description. In particular, it will be apparent to those skilled in the art that parts can be added to, or substituted for, parts of the disclosed embodiments. It is intended that the application susceptible to alterations and modifications of the preferred embodiments. Therefore, the application is to be construed as including all alternatives falling within the scope of the present application.

[0043] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A bolt tightening angle reachability prediction system for trip torque wrenches, characterized in that, Includes the following modules: The data tensor module is used to obtain the multidimensional dynamic tightening sequence and static working condition context of the current operation. The static working condition context is mapped into a high-dimensional embedding vector through the entity embedding algorithm. The multidimensional dynamic tightening sequence is transformed into a dynamic temporal feature matrix using the sliding window algorithm. The high-dimensional embedding vector and the dynamic temporal feature matrix are concatenated in the channel dimension to construct a composite heterogeneous input tensor. The feature mapping module is used to input the composite heterogeneous input tensor into the convolutional neural network branch to perform convolution operations, extract the spatial feature map of the local waveform, perform loop operations through the long short-term memory network branch to calculate the time-dependent feature vector sequence of torque evolution with angle, and perform cascade fusion of the spatial feature map and the time-dependent feature vector sequence to generate a high-dimensional spatiotemporal feature vector sequence representing the hidden state of the current friction condition. The attention inference module is used to calculate the attention weight matrix of each time step in the high-dimensional spatiotemporal feature vector sequence based on the self-attention mechanism. It generates a context vector by performing a weighted summation operation on the high-dimensional spatiotemporal feature vector sequence through the attention weight matrix. The context vector is then passed into the fully connected layer to calculate the estimated endpoint torque value. Based on the current torque growth rate, it calculates the confidence probability value of the current process corner reachability. The prediction state mapping module is used to map the confidence probability value to a preset discrete state space, determine the working condition state through a probability decoding function and generate a corner reachability state code, and encapsulate the corner reachability state code into a digital prediction message for output.

2. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 1, characterized in that: The specific process of obtaining the multidimensional dynamic tightening sequence and static working condition context of the current operation, mapping the static working condition context into a high-dimensional embedding vector through an entity embedding algorithm, and transforming the multidimensional dynamic tightening sequence into a dynamic temporal feature matrix using a sliding window algorithm is as follows: A discrete feature dictionary containing bolt material grade, target torque threshold and vehicle platform identifier is established. The discrete category index in the static working condition context is mapped to a continuous numerical high-dimensional dense embedding vector through table lookup operation. Normalization processing is performed on the high-dimensional dense embedding vector. The real-time data stream during the operation is captured at a fixed sampling frequency through the torque wrench communication interface. The torque and angle values ​​containing the current time and the previous preset length of historical time are selected to form the original time sequence segment. The original time sequence segment is subjected to differential operation to generate a first-order differential feature sequence containing angular velocity and torque change rate. The original time sequence segment and the first-order differential feature sequence are stacked in the feature dimension to generate a dynamic time sequence feature matrix.

3. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 2, characterized in that: The specific process of concatenating high-dimensional embedding vectors and dynamic temporal feature matrices along the channel dimension to construct a composite heterogeneous input tensor is as follows: The time step dimension length of the dynamic temporal feature matrix is ​​analyzed, and a broadcast copy operation along the time axis is performed on the high-dimensional dense embedding vector to expand it in the time dimension to generate a static context matrix isomorphic to the dynamic temporal feature matrix. The static context matrix and the dynamic temporal feature matrix are concatenated in parallel along the feature channel dimension to construct a composite heterogeneous input tensor.

4. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 1, characterized in that: The specific process of performing convolution operations on the branch of the composite heterogeneous input tensor convolutional neural network to extract the spatial feature map of the local waveform, and then performing loop operations through the long short-term memory network branch to calculate the time-dependent feature vector sequence of torque evolution with angle is as follows: In the convolutional neural network branch, a set of preset one-dimensional convolutional kernels are used to perform sliding convolution filtering operations on the time axis of the composite heterogeneous input tensor to extract local morphological features that represent abrupt changes and microscopic lags in the tightening curve. The convolutional output is then dimensionality-reduced through a max pooling layer to generate a local spatial feature map. In the Long Short-Term Memory (LSTM) branch, composite heterogeneous input tensors are input into the recurrent units in a time-step sequence. The memory cell states inside the units are recursively updated through forget gate and input gate mechanisms to capture the long-term dependence and nonlinear trend of torque accumulation with rotation angle. The hidden layer output of the last time step is extracted as a time-dependent feature vector sequence.

5. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 4, characterized in that: The specific process of performing concatenated fusion of spatial feature maps and time-dependent feature vector sequences to generate a high-dimensional spatiotemporal feature vector sequence representing the hidden state of the current friction condition is as follows: Perform temporal dimension alignment on the local spatial feature map and the temporally dependent feature vector sequence; A feature fusion layer is constructed, which concatenates the aligned local spatial feature map with the time-dependent feature vector sequence in the vector dimension. The concatenated joint vector sequence is then input into a fully connected layer to perform linear transformation and nonlinear activation operations. Through network weights, local morphological features and global evolution trend features are weighted and combined step by step to generate a high-dimensional spatiotemporal feature vector sequence that can implicitly represent the actual friction coefficient and contact stiffness state of the current bolted connection pair.

6. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 1, characterized in that: The specific process of calculating the attention weight matrix at each time step in the high-dimensional spatiotemporal feature vector sequence based on the self-attention mechanism, and then performing a weighted summation operation on the high-dimensional spatiotemporal feature vector sequence using the attention weight matrix to generate the context vector is as follows: Three linear projection layers—query matrix, key matrix, and value matrix—are constructed to map the high-dimensional spatiotemporal feature vector sequence to different subspaces to generate query vector sequence, key vector sequence, and value vector sequence. The dot product similarity between the query vector sequence and the key vector sequence is calculated. A scaling factor is introduced to numerically scale the dot product result. The attention weight matrix representing the feature association strength at each time step is generated by processing the result through the Softmax normalization function. The system will perform matrix multiplication on the weight matrix and value vector sequence to identify and aggregate key features of yield point and friction mutation point during tightening, and generate a context vector containing global dependency information.

7. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 6, characterized in that: The specific process of passing the context vector into the fully connected layer to calculate the estimated endpoint torque value, and calculating the confidence probability value of the current process corner reachability based on the current torque growth rate, is as follows: The instantaneous torque growth rate at the current moment is extracted by performing first-order differential operation on the end data of the multidimensional dynamic tightening sequence. The instantaneous torque growth rate is then concatenated and fused with the context vector in the feature dimension to construct an enhanced feature vector containing physical trend constraints. A multi-task prediction head is constructed, and linear regression is performed on the enhanced feature vector through a fully connected layer in the regression branch to calculate the estimated endpoint torque value. In parallel, a nonlinear transformation is performed on the enhanced feature vector through a fully connected layer in the classification branch. The transformation result is then mapped to a continuous interval from zero to one using a Sigmoid activation function to generate a confidence probability value that quantifies the reachability of the current process turn.

8. The bolt tightening angle reachability prediction system for tripping torque wrenches according to claim 1, characterized in that: The predictive state mapping module includes the following steps: In a pre-defined discrete state space, probability interval boundaries corresponding to reachable state, blocked state, and overload risk state are divided, and a probability-state mapping table is constructed. The calculated confidence probability value is compared with the probability interval boundary by using the probability decoding function to determine the probability interval of the current working condition, and the corresponding working condition category label is obtained by indexing the probability-state mapping table. Based on the working condition category label, a corresponding hexadecimal corner reachability status code is generated. The corner reachability status code and the estimated endpoint torque value are assembled into a data payload according to a preset communication protocol format. After adding a frame header and check bit, it is serialized into a digital prediction message.

Images