CN-122020178-A - Method for evaluating health degree of assembly tool based on tightening data

CN122020178ACN 122020178 ACN122020178 ACN 122020178ACN-122020178-A

Abstract

The invention relates to the technical field of strategy optimization, in particular to an assembly tool health evaluation method based on tightening data, which comprises the steps of collecting multi-source data such as torque, angle and current in the tightening process, utilizing a reinforcement learning sampling model to identify a process mode switching point to realize self-adaptive adjustment of sampling frequency, then executing order tracking processing on a current sequence, combining vibration characteristics to calculate mutual information and extract residual characteristic vectors, constructing a multi-dimensional characteristic into a symmetrical positive matrix manifold, projecting the symmetrical positive matrix manifold to Li Manqie space, eliminating global distribution offset caused by environmental fluctuation through manifold alignment operation to obtain an alignment characteristic vector, inputting the alignment vector and an environmental observation value into the reinforcement learning evaluation model, outputting correction coefficients according to geodesic distances to execute error compensation, and outputting health grade results. The invention realizes the high-confidence steady grading of the tool health state under the heterogeneous material fastening working condition through multidimensional calibration and physical constraint.

Inventors

LU YUNHUA
WU BIN
XU SHUFENG

Assignees

无锡麦克马丁定扭矩装配工具有限公司

Dates

Publication Date: 20260512
Application Date: 20260210

Claims (8)

1. An assembly tool health assessment method based on tightening data, comprising: The method comprises the steps of obtaining multi-source data in the heterogeneous material fastening process, wherein the multi-source data comprise a torque sequence, an angle sequence, a driving motor current sequence, an ultrasonic stress pulse signal and a structural response vibration characteristic; The torque sequence, the angle sequence and the structural response vibration characteristic are constructed into three-dimensional state vectors, a reinforcement learning sampling model is input, a process mode switching point is identified, a sampling adjustment instruction is output, multi-source data are resampled, and a resampling sequence is obtained; performing order tracking processing on a current sequence of a driving motor in the resampling sequence to extract energy characteristics, and calculating mutual information of the energy characteristics and structural response vibration characteristics in the resampling sequence to obtain residual characteristic vectors; combining torque components, angle components and ultrasonic stress pulse signals in the resampling sequence with residual characteristic vectors to construct a health state manifold, and projecting the health state manifold to a cutting space of the Riemann manifold to obtain an alignment characteristic vector; Inputting the alignment feature vector into a reinforcement learning evaluation model, inputting gradient components of structural response vibration features in the resampling sequence into the evaluation model as environment observation values, and performing evaluation strategy optimization according to the geodesic distance of the alignment feature vector relative to the standard state manifold and the environment observation values, and outputting a health grade result.
2. The method for evaluating the health of the assembly tool based on the tightening data, which is characterized in that the process of acquiring the multi-source data in the tightening process of the heterogeneous materials comprises the steps of synchronously acquiring a torque sequence and an angle sequence in the tightening process by utilizing a torque sensor and an angle displacement sensor which are integrated in the assembly tool, acquiring the driving motor current sequence representing the dynamic load state of a motor by utilizing a current sensor in an electric control unit, acquiring the ultrasonic stress pulse signal by utilizing a piezoelectric transducer through a pulse reflection effect, capturing the structural response vibration characteristic representing the interaction of a material interface by utilizing an acoustic emission sensor attached to a tightening area, and executing synchronous alignment by utilizing an access edge acquisition terminal to construct the multi-source data based on a uniform time reference.
3. The assembly tool health evaluation method based on tightening data according to claim 1, wherein the process of identifying a process mode switching point and outputting a sampling adjustment instruction comprises the steps of extracting first derivative characteristics of a torque sequence, angular acceleration characteristics of an angle sequence and short-time energy entropy characteristics of a structural response vibration characteristic, constructing a three-dimensional state vector representing a tightening state through normalization alignment and splicing, inputting the three-dimensional state vector into a reinforcement learning sampling model, identifying an elastic contact starting point, a gap leveling point and a material yield switching point in the process of contacting a fastener with a heterogeneous material through nonlinear mapping, and using the elastic contact starting point, the gap leveling point and the material yield switching point as the process mode switching point, calculating characteristic information requirements of a current stage by the reinforcement learning sampling model, outputting the sampling adjustment instruction comprising sampling period scaling, calculating a re-sampled signal reconstruction gain in real time, feeding the re-sampled signal back to the reinforcement learning sampling model as a reward signal, adjusting parameters of a strategy network in the reinforcement learning sampling model through a near-end strategy optimization algorithm, and correcting sampling adjustment instruction output time.
4. The assembly tool health evaluation method based on tightening data according to claim 3, wherein the reinforcement learning model comprises a feature perception network, a time sequence dependence network, an Actor execution branch, a strategy constraint optimizer and a strategy constraint optimizer, wherein the feature perception network is used for receiving the three-dimensional state vector by a one-dimensional convolution neural network, extracting local interaction features of a multi-source data stream on a time domain, the time sequence dependence network is used for carrying out sequence modeling on the local interaction features by a gating circulation unit to generate a long-range dependence feature vector containing a tightening process evolution trend, the Actor execution branch is used as an execution entity of the strategy network, the non-linear mapping is carried out on the long-range dependence feature vector by a full connection layer, the sampling period scaling factor conforming to Gaussian distribution is output, and a sampling adjustment instruction is generated, the Critic evaluation branch is used for sharing the feature perception network and the time sequence dependence network with the Actor execution branch and is used for calculating a state value estimated value corresponding to a current tightening state, and the strategy constraint optimizer is used for calculating probability ratio of a current strategy and forcing update step length of the neural network according to a clipping probability ratio function.
5. The assembly tool health evaluation method based on tightening data according to claim 1, wherein the process of obtaining the resampling sequence comprises the steps of determining a target sampling step length of each channel according to a scaling factor in the sampling adjustment instruction, performing linear interpolation operation on an original data stream according to the target sampling step length, filling in data missing points of a low-frequency sampling channel, mapping the aligned multi-source data to a uniform discrete time axis through a timestamp matching algorithm, and constructing a multi-dimensional synchronous feature matrix to serve as the resampling sequence.
6. The fitting tool health evaluation method based on tightening data according to claim 1, wherein the process of obtaining the alignment feature vector comprises performing equal-angle resampling on the driving motor current sequence by using an angle component in the resampling sequence as a reference, mapping a time domain current signal to an angle domain space, performing fast fourier transform on the angle domain current signal to obtain an order spectrum, extracting feature order energy components which are related to integer multiples of a motor rotation frequency, constructing the energy feature, respectively estimating edge probability density distribution of the energy feature and the structural response vibration feature, and constructing joint probability density distribution between the energy feature and the structural response vibration feature, calculating mutual information of the energy feature and the structural response vibration feature based on the edge probability density distribution, and stripping interference information related to the structural response vibration feature from the energy feature by using a normalized weight factor obtained based on the mutual information calculation, thereby obtaining the residual feature vector which characterizes a degradation state of a tool driving chain.
7. The fitting tool health evaluation method based on tightening data according to claim 1, wherein the process of obtaining an alignment feature vector comprises the steps of synchronously fusing the torque component, the angle component, the ultrasonic stress pulse signal and the residual feature vector, constructing a symmetrical positive definite matrix representing a tightening state space by calculating a covariance matrix to serve as the health state manifold, performing log-Euclidean transformation on the symmetrical positive definite matrix by utilizing a log mapping function, projecting the non-linear Riemann space to the Euclidean space to obtain a high-dimensional feature vector, retrieving a pre-stored reference manifold mean value under a tool standard health state, mapping the reference manifold mean value to the same cutting space to obtain a reference center vector, calculating the relative offset of the high-dimensional feature vector relative to the reference center vector, and eliminating global distribution offset caused by heterogeneous material batch difference and environment fluctuation through a centering translation operation to obtain the alignment feature vector.
8. The fitting tool health evaluation method based on tightening data according to claim 1 is characterized in that the process of outputting health grade results comprises the steps of inputting the alignment feature vector as a core evaluation variable into the reinforcement learning evaluation model, simultaneously injecting gradient components of structural response vibration features in the resampling sequence as environment observation values in real time to construct a dynamic observation space, calculating a geodesic distance of the alignment feature vector in a Riemann space relative to a preset standard state manifold, obtaining an initial deviation scalar representing tool state deviation reference degree, using the geodesic distance and the environment observation values as input states, generating judging threshold correction coefficients for different process modes by using a strategy network for executing evaluation strategy optimization, using the judging threshold correction coefficients for executing error compensation operation on the geodesic distance, stripping non-steady-state components influenced by the heterogeneous material lamination gap, extracting correction feature deviation reflecting tool body performance degradation degree, mapping the correction feature deviation to a preset health evaluation index, and outputting the health grade results representing tool operation and mechanical fatigue state.

Description

Method for evaluating health degree of assembly tool based on tightening data Technical Field The invention relates to the technical field of policy optimization, in particular to an assembly tool health evaluation method based on tightening data. Background Along with the development of industrial digital and intelligent monitoring technologies, a computer is utilized to analyze a multi-source digital sequence generated by the operation of a power tool, so that the automatic grading of the health state of equipment is realized, and the automatic grading becomes a key link for guaranteeing the stability of a production line. Under the prior art frame, the performance evolution of the tool is quantitatively evaluated by utilizing a preset data processing model through collecting digital observation data such as current, mechanical feedback, structural response and the like. However, when processing such industrial data streams with strong random interference properties, the existing digital processing method faces a technical bottleneck that due to the complexity of the fastening working condition, the digitally acquired time sequence contains a large amount of environmental noise and nonstationary components influenced by the processed object, and in the digital feature space, the nonstationary components and feature components representing the self performance degradation of the equipment are highly overlapped in time-frequency domain and amplitude dimension, so that the traditional decoupling algorithm is difficult to realize accurate extraction of the pure features of the tool end in the multidimensional high-dimensional data stream. The existing evaluation algorithm is often mapped based on the observation value at a single moment, and lacks effective constraint on the data evolution sequential logic. When the digital observation flow is affected by transient random interference (such as signal distortion caused by abrupt interface change of materials), the calculated health index can have severe non-physical fluctuation. This data processing-level instability directly leads to false positives or false negatives in the evaluation system, which makes it difficult to meet the stringent requirements of industrial scenarios for data processing robustness. Therefore, how to design a digital data processing method with the capabilities of self-adaptive feature enhancement, multidimensional space alignment and time sequence logic calibration, and achieve efficient feature stripping and robust evaluation of a limited observation sequence under a complex working condition is a key problem to be solved currently. For this purpose, an assembly tool health evaluation method based on tightening data is proposed. Disclosure of Invention The invention aims to provide an assembly tool health degree assessment method based on tightening data, which realizes high-confidence and steady grading of tool health states under heterogeneous material tightening working conditions through multidimensional calibration and physical constraint. The method comprises the steps of collecting multi-source data such as torque, angle and current in a fastening process, identifying a process mode switching point by using a reinforcement learning sampling model to realize self-adaptive adjustment of sampling frequency, then performing order tracking processing on a current sequence, calculating mutual information by combining vibration characteristics, extracting residual characteristic vectors, constructing multi-dimensional characteristics into symmetrical positive definite matrix manifold, projecting the manifold to Li Manqie space, eliminating global distribution offset caused by environmental fluctuation through manifold alignment operation to obtain alignment characteristic vectors, inputting the alignment vectors and an environmental observation value into a reinforcement learning evaluation model, performing error compensation according to a geodesic distance output correction coefficient, and outputting health grade results. The invention provides the following technical scheme for realizing the purposes: An assembly tool health assessment method based on tightening data, comprising: The method comprises the steps of obtaining multi-source data in the heterogeneous material fastening process, wherein the multi-source data comprise a torque sequence, an angle sequence, a driving motor current sequence, an ultrasonic stress pulse signal and a structural response vibration characteristic; The torque sequence, the angle sequence and the structural response vibration characteristic are constructed into three-dimensional state vectors, a reinforcement learning sampling model is input, a process mode switching point is identified, a sampling adjustment instruction is output, multi-source data are resampled, and a resampling sequence is obtained; performing order tracking processing on a current sequence of a driving motor in the resamplin