CN-122020085-A - Real-time detection method for assembly quality based on time sequence data in screwing process

CN122020085ACN 122020085 ACN122020085 ACN 122020085ACN-122020085-A

Abstract

The invention relates to the technical field of industrial data processing, in particular to an assembly quality real-time detection method based on time sequence data of a screwing process, which comprises the steps of collecting time sequence data of torque, angle and current; the method comprises the steps of constructing a dynamic model comprising a physical evolution branch and a transient perception branch, dividing data into a screwing section, a fitting section, an elastic section and a yield rear section based on the abrupt change characteristic of a second derivative of torque with respect to angles, extracting a whole process time sequence dependent characteristic vector by utilizing a long-period memory network, constructing an intelligent body based on deep reinforcement learning, observing an environment state space by utilizing a strategy network, extracting Gao Weiyin state vectors as local transient form characteristic vectors, generating a multi-mode fusion characteristic vector by utilizing a multi-head attention mechanism, and outputting an assembly defect type label and a physical stage index of abnormality occurrence by utilizing a full-connection layer classifier. The invention solves the problems of difficult identification of the micro defects and low positioning precision.

Inventors

XU SHUFENG
WU BIN
HUA JIAXIN

Assignees

无锡麦克马丁定扭矩装配工具有限公司

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (7)

1. The real-time assembly quality detection method based on the time sequence data of the screwing process is characterized by comprising the following steps of: The method comprises the steps of collecting torque time sequence data, angle time sequence data and current time sequence data in the screwing process, constructing a dynamic model comprising a physical evolution branch and a transient perception branch which are arranged in parallel; Inputting the torque time sequence data and the angle time sequence data into the physical evolution branch, calculating a second derivative change curve of the torque with respect to the angle, determining a time index based on abrupt change feature points of the second derivative change curve, dividing the torque time sequence data and the angle time sequence data into a screwing section, a fitting section, an elastic section and a yielding rear section by using the time index, inputting the divided data segments into a long-period memory network and extracting a whole-process time sequence dependency feature vector; Inputting current time sequence data into the transient sensing branch, constructing an intelligent agent based on deep reinforcement learning, mapping the current time sequence data into an environment state space, observing the environment state space by using a strategy network, outputting motion probability distribution, and extracting Gao Weiyin state vectors of the strategy network as local transient morphological feature vectors; Calculating an association weight matrix between the whole process time sequence dependency feature vector and the local transient state form feature vector by utilizing a multi-head attention mechanism, weighting and splicing the two feature vectors according to the association weight matrix to generate a multi-mode fusion feature vector, inputting the multi-mode fusion feature vector into a full-connection layer classifier, and outputting an assembly defect type label and a physical stage index of abnormality occurrence.
2. The real-time assembly quality detection method based on the time sequence data of the tightening process is characterized by further comprising a data preprocessing process after the time sequence data, the angle time sequence data and the current time sequence data in the tightening process are collected, wherein the data preprocessing process specifically comprises the steps of taking the time sequence data with highest sampling frequency as a reference time axis, resampling the collected time sequence data, angle time sequence data and current time sequence data by using a linear interpolation algorithm to enable all the time sequence data to be aligned in a time dimension, conducting smoothing denoising processing on the time sequence data and the angle time sequence data input into the physical evolution branch by using a Savitzky-Golay filter, keeping high-frequency quantization noise while keeping peak characteristics of signal waveforms, and respectively calculating statistical mean values and standard differences of the time sequence data, the angle time sequence data and the current time sequence data, and mapping each time sequence data to a standard normal distribution space by using a Z-score normalization formula.
3. The method for detecting the assembly quality based on the time sequence data of the tightening process in real time is characterized by comprising the steps of constructing a discrete mapping sequence of torque changing along with angles, calculating local curvature of the discrete mapping sequence by utilizing a second-order center difference operator to generate a second derivative sequence of an angle domain, traversing the second derivative sequence, identifying a position of the second derivative value, which is positively risen from zero and exceeds a preset positive threshold value for the first time, as a fitting start characteristic point, identifying a position of the second derivative value, which is firstly fallen back and passes through a zero axis after undergoing a positive interval, as an elastic start characteristic point, identifying a position of the second derivative value, which is negatively fallen from zero and is lower than the preset negative threshold value for the first time, as a yield start characteristic point, retrieving angle values corresponding to the fitting start characteristic point, the elastic start characteristic point and the yield start characteristic point in the discrete mapping sequence, performing reverse searching in the angle time sequence, locating time stamps matched with the angle values, and marking the located time stamps as a first time index for dividing a screw-in segment and a fitting segment, a second time index for fitting a fitting segment and an elastic segment, and a second time index for dividing a third time segment after the elastic segment and a yield index.
4. The real-time detection method for assembly quality based on time sequence data of a screwing process is characterized by comprising the steps of constructing a multi-dimensional input vector sequence, combining a current torque time sequence data value and an angle time sequence data value into a time sequence input vector for each time sequence of a screwing section, a fitting section, an elastic section and a yield back section, sequentially inputting the time sequence input vectors of each section into a long-short-term memory network according to the sequence of physical processes, respectively extracting hidden state vectors output by the long-short-term memory network when the last time sequence of the screwing section, the fitting section, the elastic section and the yield back section is processed, obtaining a screwing feature vector, a fitting feature vector, an elastic feature vector and a yield feature vector, performing time sequence feature aggregation, and splicing the screwing feature vector, the fitting feature vector, the elastic feature vector and the yield feature vector according to time dimensions to obtain the whole-process time sequence dependency feature vector.
5. The method for real-time detection of assembly quality based on time series data of tightening process according to claim 1, wherein the construction of the specific operation flow of the intelligent body based on deep reinforcement learning comprises intercepting a current time series data segment by utilizing a sliding window mechanism, calculating the mean value, variance and statistical characteristic quantity of crest factor of the current time series data segment, splicing the statistical characteristic quantity with an action vector output by the intelligent body at the last moment to construct an observation state vector at the current moment, inputting the observation state vector into a strategy network, updating an internal hidden memory state of the intelligent body by utilizing a circulating neural network layer integrated in the strategy network according to the observation state vector, outputting action probability distribution at the current moment by a full connection layer based on the internal hidden memory state, generating discrete monitoring actions according to the action probability distribution sampling, wherein the discrete monitoring actions comprise maintaining the observation actions, focusing actions and triggering pre-warning actions, the intelligent body dynamically adjusts sliding window step length parameters of the sliding window mechanism at the next moment according to the generated discrete monitoring actions, extracting the sliding window step length parameters of the sliding window mechanism at the final moment after the full sequence observation of the current time series data is completed, outputting the final circulating neural network layer at the moment at the time to serve as a time of the sparse response, and performing a sparse response function in response on the transient state or a transient state, and giving a sparse response on the transient state in response, or a training and a gain, and a sparse form is configured in response to the transient state, and a sparse form, based on the transient state is optimized, and giving false alarm punishment return when the key focusing action is executed and/or the early warning action is triggered in the undisturbed interval.
6. The method for detecting the assembly quality based on the time sequence data of the tightening process in real time is characterized by comprising the steps of constructing a group of learnable linear mapping matrixes, mapping the whole process time sequence dependent feature vectors into query matrixes, mapping the local transient state form feature vectors into key matrixes and value matrixes respectively, carrying out dot product operation on the query matrixes and the transposed key matrixes, dividing an operation result by a scaling factor for normalization, converting the normalized result into an associated weight matrix in a probability distribution form by using a Softmax function, multiplying the associated weight matrix by the value matrixes to obtain transient state perception feature vectors, carrying out feature channel splicing operation, splicing the whole process time sequence dependent feature vectors and the transient state perception feature vectors in channel dimension, carrying out dimension fusion through a full connection layer, and outputting the multimode fusion feature vectors.
7. The real-time assembly quality detection method based on time sequence data of a tightening process is characterized by comprising the steps of inputting a multi-modal fusion feature vector into a full-connection layer classifier, outputting an assembly defect type label and a physical stage index of abnormal occurrence, and specifically comprises the steps of configuring the full-connection layer classifier to a multi-task parallel output framework comprising a shared feature extraction layer, a defect classification branch and a stage positioning branch, inputting the multi-modal fusion feature vector into the shared feature extraction layer, performing high-dimensional mapping by utilizing a nonlinear activation function, extracting an implicit public semantic feature vector, inputting the public semantic feature vector into the defect classification branch, mapping feature dimensions into the number of preset assembly defect types through the full-connection layer, calculating probability distribution of samples belonging to normal assembly, sliding teeth, false buckles, floating screws and abnormal occurrence of materials by utilizing a Softmax normalization function, selecting the type corresponding to the maximum probability value as the assembly defect type label, inputting the public semantic feature vector into the stage positioning branch, mapping the feature dimensions into a screwing-in section, a fitting section, an elastic section and the number of physical stages after the physical stages by utilizing the full-connection layer, and calculating confidence coefficient of abnormal occurrence of each abnormal occurrence stage by utilizing the confidence coefficient of the physical stage normalization function.

Description

Real-time detection method for assembly quality based on time sequence data in screwing process Technical Field The invention relates to the technical field of industrial data processing, in particular to a real-time detection method for assembly quality based on time sequence data in a screwing process. Background With the development of intelligent manufacturing, industrial assembly sites generate massive high-frequency time sequence data of tightening processes. The real-time analysis and modeling of the multi-source heterogeneous data by using a computer technology is a core requirement for realizing the digital monitoring and quality tracing of the assembly process. Currently, assembly quality testing relies primarily on statistical-based threshold decisions or supervised learning models. These methods usually use a fixed-step sliding window mechanism to passively process data, and it is difficult to combine long timing dependency with high-frequency transient characteristics. Although reinforcement learning has been applied in the field of industrial control, these applications are mostly limited to robot path planning or offline optimization of PID control parameters, lacking research on online adaptive adjustment of the signal sampling process itself by means of an agent. The existing data processing architecture often carries out hybrid processing on physical evolution trend and high-frequency disturbance signals, and multisource feature fusion is realized only through simple channel splicing, so that dynamic focusing and accurate positioning based on physical context cannot be realized when a model faces microscopic anomalies. Therefore, the technical problem to be solved is how to overcome the defects of the existing model such as missing modeling of the stage evolution logic and low capturing precision of transient anomalies in the multi-source time sequence data processing, and realize high-precision identification and positioning of the state of the assembly process. Therefore, a real-time detection method for assembly quality based on time sequence data of a screwing process is provided. Disclosure of Invention The invention aims to provide a real-time detection method for assembly quality based on time sequence data in a screwing process, which solves the problems of difficult recognition and low positioning accuracy of tiny assembly abnormality through deep fusion of physical evolution logic and transient active perception. In order to achieve the above purpose, the present invention provides the following technical solutions: a real-time detection method for assembly quality based on time sequence data of a tightening process comprises the following steps: The method comprises the steps of collecting torque time sequence data, angle time sequence data and current time sequence data in the screwing process, constructing a dynamic model comprising a physical evolution branch and a transient perception branch which are arranged in parallel; Inputting the torque time sequence data and the angle time sequence data into the physical evolution branch, calculating a second derivative change curve of the torque with respect to the angle, determining a time index based on abrupt change feature points of the second derivative change curve, dividing the torque time sequence data and the angle time sequence data into a screwing section, a fitting section, an elastic section and a yielding rear section by using the time index, inputting the divided data segments into a long-period memory network and extracting a whole-process time sequence dependency feature vector; Inputting current time sequence data into the transient sensing branch, constructing an intelligent agent based on deep reinforcement learning, mapping the current time sequence data into an environment state space, observing the environment state space by using a strategy network, outputting motion probability distribution, and extracting Gao Weiyin state vectors of the strategy network as local transient morphological feature vectors; Calculating an association weight matrix between the whole process time sequence dependency feature vector and the local transient state form feature vector by utilizing a multi-head attention mechanism, weighting and splicing the two feature vectors according to the association weight matrix to generate a multi-mode fusion feature vector, inputting the multi-mode fusion feature vector into a full-connection layer classifier, and outputting an assembly defect type label and a physical stage index of abnormality occurrence. The method comprises the steps of collecting torque time sequence data, angle time sequence data and current time sequence data in a screwing process, and then further comprising a data preprocessing process, wherein the data preprocessing process specifically comprises the steps of taking the time sequence data with highest sampling frequency as a reference time axis, resampling the collec