Assembly tool health degree assessment method based on tightening data

Abstract

The invention relates to the technical field of strategy optimization, in particular to an assembly tool health degree evaluation method based on tightening data, which comprises the following steps: acquiring 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, and realizing adaptive adjustment of sampling frequency; performing order tracking processing on the current sequence, calculating mutual information in combination with vibration features, and extracting residual feature vectors; constructing the multi-dimensional features into a symmetric positive definite matrix manifold, projecting the symmetric positive definite matrix manifold to a Riemannian tangent space, eliminating global distribution offset caused by environmental fluctuation through manifold alignment operation, and obtaining an alignment feature vector; and inputting the alignment vector and the environment observation value into a reinforcement learning evaluation model, outputting a correction coefficient according to the geodesic distance to execute error compensation, and outputting a health level result. According to the method, through multi-dimensional calibration and physical constraint, high-confidence robust grading of the tool health state under the heterogeneous material fastening working condition is achieved.

Description

Technical Field

[0001] This invention relates to the field of strategy optimization technology, specifically to a method for assessing the health of assembly tools based on tightening data. Background Technology

[0002] With the development of industrial digitalization and intelligent monitoring technologies, analyzing multi-source digital sequences generated by power tool operation using computers to automatically classify equipment health status has become a key aspect of ensuring production line stability. Within the existing technological framework, this typically involves collecting digital observation data such as current, mechanical feedback, and structural response, and then using pre-defined data processing models to quantitatively evaluate the tool's performance evolution.

[0003] However, existing digital processing methods face technical bottlenecks when dealing with industrial data streams with strong random interference: due to the complexity of tight working conditions, the digitally acquired time series contains a large amount of environmental noise and non-stationary components affected by the processed object. In the digital feature space, these non-stationary components and the feature components that characterize the performance degradation of the equipment itself highly overlap in the time-frequency domain and amplitude dimension, making it difficult for traditional decoupling algorithms to accurately extract the clean features of the tool end in multi-dimensional and high-dimensional data streams.

[0004] Existing assessment algorithms often map data based on observations at a single moment, lacking effective constraints on the temporal logic of data evolution. When the digital observation stream is affected by transient random disturbances (such as signal distortion caused by abrupt changes in material interfaces), the calculated health indicators exhibit drastic non-physical fluctuations. This instability at the data processing level directly leads to false alarms or false negatives in the assessment system, making it difficult to meet the stringent requirements for robust data processing in industrial scenarios.

[0005] Therefore, designing a digital data processing method with adaptive feature enhancement, multi-dimensional spatial alignment, and temporal logic calibration capabilities to achieve efficient feature stripping and robust evaluation of constrained observation sequences under complex working conditions is a key problem that urgently needs to be solved.

[0006] To address this, a method for assessing the health of assembly tools based on tightening data is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a method for assessing the health of assembly tools based on tightening data. Through multi-dimensional calibration and physical constraints, it achieves a high-confidence, robust rating of the tool's health status under heterogeneous material tightening conditions. The method includes collecting multi-source data such as torque, angle, and current during the tightening process; using a reinforcement learning sampling model to identify process mode switching points and adaptively adjust the sampling frequency; subsequently, performing order tracking processing on the current sequence and calculating mutual information based on vibration characteristics to extract residual feature vectors; constructing a symmetric positive definite matrix manifold from the multi-dimensional features and projecting it onto the Riemann tangent space; eliminating global distribution offsets caused by environmental fluctuations through manifold alignment operations to obtain aligned feature vectors; inputting the aligned vectors and environmental observations into the reinforcement learning evaluation model; performing error compensation based on geodesic distance correction coefficients; and outputting the health level result.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for assessing the health of assembly tools based on tightening data includes: Acquire multi-source data during the fastening process of heterogeneous materials, including torque sequence, angle sequence, drive motor current sequence, ultrasonic stress pulse signal, and structural response vibration characteristics; The torque sequence, angle sequence, and structural response vibration features are constructed into a three-dimensional state vector, which is input into the reinforcement learning sampling model to identify the process mode switching point and output sampling adjustment command. The multi-source data is then resampled to obtain the resampled sequence. The drive motor current sequence in the resampled sequence is subjected to order tracking processing to extract energy features; the mutual information between the energy features and the structural response vibration features in the resampled sequence is calculated to obtain the residual feature vector; The torque component, angle component, and ultrasonic stress pulse signal in the resampled sequence are combined with the residual feature vector to construct a healthy state manifold. The healthy state manifold is then projected onto the tangent space of the Riemann manifold to obtain the aligned feature vector. The alignment feature vector is input into the reinforcement learning evaluation model, and the gradient components of the structural response vibration features in the resampled sequence are used as environmental observations to input into the evaluation model. The evaluation strategy is optimized based on the geodesic distance of the alignment feature vector relative to the standard state manifold and combined with the environmental observations, and the health level result is output.

[0009] Preferably, the process of acquiring multi-source data during the fastening of heterogeneous materials includes: simultaneously acquiring the torque sequence and the angle sequence during the fastening process using a torque sensor and an angular displacement sensor integrated inside the assembly tool; capturing the drive motor current sequence characterizing the dynamic load state of the motor using a current sensor in the electronic control unit; acquiring the ultrasonic stress pulse signal through the pulse reflection effect using a piezoelectric transducer; capturing the structural response vibration characteristics characterizing the material interface interaction using an acoustic emission sensor attached to the fastening area; and connecting to an edge acquisition terminal to perform synchronous alignment and construct multi-source data based on a unified time reference.

[0010] Preferably, the process of identifying the process mode switching point and outputting the sampling adjustment command includes: extracting the first derivative features of the torque sequence, the angular acceleration features of the angle sequence, and the short-time energy entropy features of the structural response vibration features; constructing a three-dimensional state vector representing the fastening state through normalized alignment and splicing; inputting the three-dimensional state vector into the reinforcement learning sampling model; identifying the elastic contact initiation point, interlayer flattening point, and material yield switching point during the contact process between the fastener and the heterogeneous material through nonlinear mapping, as the process mode switching point; the reinforcement learning sampling model calculates the feature information abundance requirement of the current stage based on the identified process mode switching point, and outputs a sampling adjustment command containing the sampling period scaling factor; calculating the signal reconstruction gain after resampling in real time and feeding it back to the reinforcement learning sampling model as a reward signal; adjusting the parameters of the policy network in the reinforcement learning sampling model through a near-end policy optimization algorithm; and correcting the timing of the sampling adjustment command output.

[0011] Preferably, the reinforcement learning sampling model includes: a feature-aware network: using a one-dimensional convolutional neural network to receive the three-dimensional state vector and extract local interaction features of multi-source data streams in the temporal domain; a temporal dependency network: using gated recurrent units to perform sequence modeling on the local interaction features, generating a long-range dependency feature vector containing the evolution trend of the fastening process; an Actor execution branch: acting as the execution entity of the policy network, using fully connected layers to perform nonlinear mapping on the long-range dependency feature vector, outputting the sampling period scaling factor conforming to a Gaussian distribution, and generating sampling adjustment instructions; a Critic evaluation branch: sharing the feature-aware network and the temporal dependency network with the Actor execution branch, used to calculate the state value estimate corresponding to the current fastening state; and a policy constraint optimizer: using a proximal policy optimization algorithm to calculate the probability ratio between the current policy and the historical policy, and forcibly constraining the update step size of the neural network according to the pruning probability ratio function.

[0012] Preferably, the process of obtaining the resampling sequence includes: determining the target sampling step size for each channel according to the scaling factor in the sampling adjustment instruction; performing linear interpolation on the original data stream according to the target sampling step size to fill in the missing data points of the low-frequency sampling channels; and mapping the aligned multi-source data to a unified discrete time axis through a timestamp matching algorithm to construct a multi-dimensional synchronization feature matrix as the resampling sequence.

[0013] Preferably, the process of obtaining the aligned feature vector includes: using the angular components in the resampled sequence as a reference, performing equal-angle resampling on the drive motor current sequence to map the time-domain current signal to the angular domain space; performing a fast Fourier transform on the angular domain current signal to obtain the order spectrum, extracting the characteristic order energy components that are integer multiples of the motor rotation frequency, and constructing the energy feature; estimating the marginal probability density distributions of the energy feature and the structural response vibration feature respectively, and constructing the joint probability density distribution between the two; calculating the mutual information between the energy feature and the structural response vibration feature based on the marginal probability density distribution and the joint probability density distribution, using the normalized weighting factor obtained based on the mutual information calculation, removing the interference information related to the structural response vibration feature from the energy feature, and obtaining the residual feature vector characterizing the degradation state of the tool transmission chain.

[0014] Preferably, the process of obtaining the alignment feature vector includes: synchronously fusing the torque component, angle component, ultrasonic stress pulse signal, and residual feature vector; constructing a symmetric positive definite matrix representing the tightness state space by calculating the covariance matrix, which serves as the healthy state manifold; performing a logarithmic-Euclidean transformation on the symmetric positive definite matrix using a logarithmic mapping function, projecting it from a nonlinear Riemannian space to an Euclidean tangent space to obtain a high-dimensional feature tangent vector; retrieving the mean of the reference manifold under the pre-stored tool standard healthy state and mapping it to the same tangent space to obtain a reference center vector; calculating the relative offset of the high-dimensional feature tangent vector relative to the reference center vector, and eliminating the global distribution offset caused by batch differences in heterogeneous materials and environmental fluctuations through a centering translation operation to obtain the alignment feature vector.

[0015] Preferably, the process of outputting the health level result includes: inputting the alignment feature vector as the core evaluation variable into the reinforcement learning evaluation model, and simultaneously injecting the gradient components of the structural response vibration characteristics in the resampled sequence as environmental observations in real time to construct a dynamic observation space; calculating the geodesic distance of the alignment feature vector relative to the preset standard state manifold in Riemann space to obtain an initial deviation scalar characterizing the degree of deviation of the tool state from the baseline; the reinforcement learning evaluation model uses the geodesic distance and the environmental observations as input states, and uses a policy network to generate decision threshold correction coefficients for different process modes, and performs evaluation strategy optimization; using the decision threshold correction coefficients to perform error compensation calculations on the geodesic distance, removing the unsteady components affected by the gaps in the heterogeneous material stacks, and extracting the corrected feature deviation reflecting the degree of performance degradation of the tool body; mapping the corrected feature deviation to a preset health evaluation index system, and outputting the health level result characterizing the tool's operating accuracy and mechanical fatigue state.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By employing a multi-source sensor fusion and reinforcement learning sampling model adaptive sampling mechanism, this application achieves accurate identification of process mode switching points during the fastening of heterogeneous materials. This mechanism can dynamically increase the sampling frequency at key physical nodes such as elastic contact, interlayer flattening, and yield deformation based on the instantaneous changes in torque slope and short-term energy entropy, thereby capturing more weak high-frequency transient features containing tool degradation information. This process improves the abundance of feature information from the data source, effectively alleviating the problem of feature overlap caused by data gaps.

[0017] 2. By employing angular domain order tracking processing and a feature component stripping algorithm based on mutual information, a residual feature vector reflecting the degradation state of the tool drivetrain was constructed. Utilizing the nonlinear correlation quantification between current energy characteristics and structural response vibration characteristics, the energy response components caused by the damping characteristics of heterogeneous materials and load fluctuations can be accurately estimated and reduced. This allows the electromechanical degradation signal of the tool body to be independently extracted from complex background noise, improving the sensitivity of the evaluation system in identifying the early wear state of the tool body. By introducing a causal mapping mechanism, the problem of distinguishing between environmental interference and equipment degradation in traditional signal processing is solved, ensuring that the residual feature vector has a clear physical orientation.

[0018] 3. By constructing a symmetric positive definite manifold and performing alignment processing in the Riemann tangent space, the nonlinear health state features are projected to the linear tangent space using the logarithmic-Euclidean transformation. Combined with benchmark distribution extraction and centering translation operations, the consistency calibration of feature distribution under different assembly conditions is achieved, which enhances the generalization ability of the algorithm in multi-condition scenarios and ensures the uniformity and stability of the evaluation criteria in heterogeneous material assembly tasks.

[0019] 4. By combining reinforcement learning evaluation models with geodesic distance and environmental gradient observations for strategy optimization, geodesic distance is used to quantify the length of the essential degradation path within the manifold space, and structural response gradient is introduced as an environmental observation. The model can perceive the impact of operating condition fluctuations on the evaluation results in real time and adaptively adjust the judgment weights and threshold matrices, providing a scientific and reliable quantitative decision-making basis for the predictive maintenance of precision assembly tools. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a method for assessing the health of assembly tools based on tightening data according to the present invention. Figure 2 This is a schematic diagram of the process for obtaining the aligned feature vector according to the present invention; Figure 3 This is a schematic diagram of the output health level result flow of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1 to 3 This invention provides a method for assessing the health of assembly tools based on tightening data, the technical solution of which is as follows: Example 1: A method for assessing the health of assembly tools based on tightening data, the specific process of which is as follows: Figure 1 As shown, it includes: Acquire multi-source data during the fastening process of heterogeneous materials, including torque sequence, angle sequence, drive motor current sequence, ultrasonic stress pulse signal, and structural response vibration characteristics; The torque sequence, angle sequence, and structural response vibration features are constructed into a three-dimensional state vector, which is input into the reinforcement learning sampling model to identify the process mode switching point and output sampling adjustment command. The multi-source data is then resampled to obtain the resampled sequence. The drive motor current sequence in the resampled sequence is subjected to order tracking processing to extract energy features; the mutual information between the energy features and the structural response vibration features in the resampled sequence is calculated to obtain the residual feature vector; The torque component, angle component, and ultrasonic stress pulse signal in the resampled sequence are combined with the residual feature vector to construct a healthy state manifold. The healthy state manifold is then projected onto the tangent space of the Riemann manifold to obtain the aligned feature vector. The alignment feature vector is input into the reinforcement learning evaluation model, and the gradient components of the structural response vibration features in the resampled sequence are used as environmental observations to input into the evaluation model. The evaluation strategy is optimized based on the geodesic distance of the alignment feature vector relative to the standard state manifold and combined with the environmental observations, and the health level result is output.

[0023] Furthermore, the process of acquiring multi-source data during the fastening of heterogeneous materials includes: simultaneously acquiring the torque sequence and angle sequence during the fastening process using torque sensors and angular displacement sensors integrated within the assembly tool; capturing the drive motor current sequence characterizing the dynamic load state of the motor using a current sensor in the electronic control unit; acquiring the ultrasonic stress pulse signal through the pulse reflection effect using a piezoelectric transducer; capturing the structural response vibration characteristics characterizing the material interface interaction using an acoustic emission sensor attached to the fastening area; and connecting to an edge acquisition terminal to perform synchronous alignment and construct multi-source data based on a unified time reference.

[0024] In the fastening scenario of carbon fiber composite materials and aluminum alloy laminates, based on the physical integration at the hardware level of the assembly tool, the torque sensor and angular displacement sensor integrated inside the output end of the assembly tool are used to synchronously collect the torque sequence and angular sequence during the fastening process at a fixed frequency of 1kHz at the same time as the tightening program starts. The torque sensor is used to capture the evolution process of thread engagement resistance and final preload, while the angular displacement sensor records the number of rotations and the axial displacement of the fastener. These two constitute the most basic macroscopic mechanical and kinematic observation data of the fastening process.

[0025] Meanwhile, using the current sensor integrated within the assembly tool's electronic control unit, the drive motor current sequence, which characterizes the dynamic load state of the motor, is captured in real time. In the specific execution process, the current sensor monitors changes in the stator current or bus current of the drive motor, converts it into a discrete current amplitude sequence through analog-to-digital conversion, and ensures that the sampling trigger signal of this sequence is synchronized with the clock source for acquiring mechanical data such as torque and angle. Based on the linear physical mapping relationship between current magnitude and motor output electromagnetic torque, the micro-pulsation characteristics in the current waveform are monitored in real time. When gear wear, abnormal bearing load, or lubrication failure occur within the transmission chain, the corresponding damping fluctuations will cause high-frequency modulation changes in the amplitude of the current sequence. By capturing these electrical signal components directly related to the mechanical state of the tool, observational data reflecting the degree of degradation of the internal transmission system of the tool is obtained, providing feature input for subsequent health assessment.

[0026] To directly obtain the internal axial load of the fastener, a high-frequency piezoelectric transducer is arranged on the bolt head end face through an acoustic coupling medium. During the fastening process, the transducer receives electrical excitation to generate ultrasonic longitudinal waves. The propagation time difference of the sound wave back and forth in the bolt shank is obtained by using the pulse reflection effect. According to the principle of acoustoelasticity, this propagation time shifts linearly with the increase of bolt axial stress, thereby generating the ultrasonic stress pulse signal that characterizes the actual stress state of the fastener in real time. This signal is not affected by fluctuations in the surface friction coefficient, providing an objective physical benchmark for evaluating the loading accuracy of the tool.

[0027] To address the dynamic response of heterogeneous materials under fastening pressure, an acoustic emission sensor attached to the material surface near the fastening area is used to capture the structural response vibration characteristics that characterize the material interface interaction. In the specific implementation process, the frequency response range of the acoustic emission sensor is set to 100kHz to 1MHz to ensure complete coverage of the stress wave frequency band excited by the interface friction, lamination extrusion, and micro-plastic deformation of the heterogeneous materials during the fastener screwing process. The acoustic emission sensor converts the captured high-frequency mechanical vibration signal into an electrical signal, which is then pre-amplified and filtered by the signal processing unit to extract the feature sequence reflecting the evolution of key working conditions such as lamination gap flattening and interface contact stress surge. Through this signal, the system perceives the physical state switching of the fastening working condition from the dimension of high-frequency micro-vibration, providing working condition background data with high spatiotemporal resolution for subsequent identification of process mode switching points.

[0028] Finally, heterogeneous signals from different sensors are all connected to the edge acquisition terminal through shielded cables. Given that the sampling rate of acoustic emission signals and ultrasonic signals is much higher than that of torque and current signals, the edge acquisition terminal performs multi-rate synchronization alignment through its built-in hardware clock distributor. It assigns a high-precision timestamp based on a global synchronization pulse to each raw data stream. Through this timestamp matching mechanism, the high-speed sampling and low-speed sampling data are logically correlated. Finally, a multi-source data matrix based on a unified time reference is constructed on the discrete time axis. This process eliminates the phase lag between channels and ensures that the subsequent resampling sequence and feature extraction are closely aligned in physical time.

[0029] Before inputting the structural response vibration characteristics into the reinforcement learning sampling model, the method further includes: performing time-frequency domain analysis on the acoustic emission signal using a spectral kurtosis algorithm to identify and extract the characteristic frequency band with the greatest non-stationarity in the signal; extracting the transient pulse component of the characteristic frequency band through a bandpass filter, and using the energy change rate of the transient pulse component as the core feature that triggers the reinforcement learning sampling model to adjust the sampling frequency.

[0030] Specifically, the process begins by acquiring the raw acoustic emission signal sequence and performing time-frequency domain decomposition on it using a fast kurtosis algorithm. In this process, the preset decomposition level is 6 levels, and a 1 / 3 binary tree structure is used to divide the full-band signal into multiple sub-band windows. For the signal samples within each sub-band window, the fourth-order cumulant is calculated to obtain the corresponding kurtosis value. This kurtosis value is used to quantify the degree of non-stationarity of the signal in a specific frequency band, thereby identifying transient pulse components masked by broadband mechanical noise. The internal search operator performs a global scan on the generated two-dimensional kurtosis matrix to find and lock the coordinates of the extreme point with the largest kurtosis value. Once the coordinates are locked, the corresponding center frequency and decomposition bandwidth are automatically extracted and mapped in real time to the upper and lower cutoff frequency boundaries of the physical bandpass filter. Subsequently, based on the above frequency boundaries, the 64th-order FIR filter coefficients are calculated in real time using the window function method and dynamically loaded into the register of the digital signal processing unit to complete the adaptive filter configuration for the sensitive characteristic frequency band under the current operating conditions.

[0031] After the original acoustic emission signal stream passes through an adaptive bandpass filter, steady-state interference related to mechanical transmission and motor electromagnetics is filtered out, retaining only the transient pulse component reflecting the state changes of the heterogeneous material interface. Then, a Hilbert transform is performed on the filtered signal to obtain its analytical envelope, and the sum of squares of the envelope signal's energy is calculated by sliding along a preset time step to obtain an instantaneous energy sequence reflecting the intensity of interface frictional abrupt changes. A first-order difference operation is performed on the instantaneous energy sequence to calculate the rate of change of instantaneous energy over time. This rate of change, as a core state feature, is injected into the observation space of the reinforcement learning sampling model in real time. When the reinforcement learning model detects that the rate of change of energy exceeds a preset physical criterion threshold, it determines that the current fastening process has entered a non-steady-state stage of interlayer flattening or frictional abrupt changes. Based on the input energy rate of change feature, the reinforcement learning sampling model outputs a corresponding frequency adjustment command. The system drives the front-end acquisition circuit to switch the sampling rate according to the instruction, instantly switching the sampling frequency of the acoustic emission signal from low-frequency monitoring mode to high-frequency capture mode (e.g., from 500kHz to over 2MHz in this embodiment). This process ensures that the system can achieve high-fidelity capture of the physical abrupt changes at the interface of heterogeneous materials while eliminating background noise interference. By identifying and extracting transient pulse components in the high-frequency sensitive band through the spectral kurtosis algorithm, the technical problem of strong background noise masking weak friction signals at heterogeneous interfaces is solved, significantly improving the sensitivity and triggering accuracy of the reinforcement learning sampling model for key physical nodes.

[0032] By integrating torque, angle, and current sensors, as well as ultrasonic and acoustic emission detection units, a sensing system covering multiple modes of electricity, force, and sound was constructed. The introduction of ultrasonic stress pulses and acoustic emission signals effectively supplemented the details of axial force and interface interaction during the fastening process of heterogeneous materials. With the help of synchronous alignment processing of edge acquisition terminals, the physical consistency of multi-dimensional features under the same time reference was ensured, laying a high-fidelity observation foundation for subsequent deep decoupling of execution features and accurate health assessment.

[0033] Furthermore, the process of identifying the process mode switching point and outputting the sampling adjustment command includes: extracting the first derivative features of the torque sequence, the angular acceleration features of the angle sequence, and the short-time energy entropy features of the structural response vibration features; constructing a three-dimensional state vector representing the fastening state through normalized alignment and splicing; inputting the three-dimensional state vector into the reinforcement learning sampling model; identifying the elastic contact initiation point, interlayer flattening point, and material yield switching point during the contact process between the fastener and the heterogeneous material through nonlinear mapping, as the process mode switching point; the reinforcement learning sampling model calculates the feature information abundance requirement of the current stage based on the identified process mode switching point, and outputs a sampling adjustment command containing the sampling period scaling factor; calculating the signal reconstruction gain after resampling in real time and feeding it back to the reinforcement learning sampling model as a reward signal; adjusting the parameters of the policy network in the reinforcement learning sampling model through a near-end policy optimization algorithm; and correcting the timing of the sampling adjustment command output.

[0034] The specific calculation method of the signal reconstruction gain is as follows: calculate the cross-correlation coefficient R between the original high-frequency acoustic emission signal envelope and the resampled signal envelope within the current process mode window, and take the absolute value of the difference between R and 1 as the negative reward. Through this short-term feedback, the reinforcement learning sampling model can perceive in real time whether the current sampling frequency is sufficient to capture the transient pulse generated by the friction of heterogeneous materials, thereby achieving a dynamic balance between ensuring information integrity and reducing data redundancy.

[0035] Specifically, firstly, feature extraction is performed on the collected multi-source data to construct a state space. For the torque sequence, the first derivative feature is calculated using the sliding window difference method to reflect the rate of change of fastening resistance. For the angle sequence, the angular acceleration feature is calculated using the second-order difference method to capture the transient fluctuations of the fastener rotation dynamics. For the structural response vibration features obtained by the acoustic emission sensor, the short-time energy entropy feature is calculated using the short-time energy analysis method to quantify the microscopic activity of the interface interaction. Subsequently, the above three feature quantities are respectively processed by Min-Max normalization to the [0, 1] interval, and aligned and spliced ​​according to the temporal logic to construct a three-dimensional state vector representing the current fastening state space. The specific data dimension of the three-dimensional state vector is 3 times 200, where 3 represents the three channels of torque first derivative, angular acceleration and short-time energy entropy, and 200 represents the number of continuous sampling points in the current sliding observation window. This dimensional design ensures that the model can simultaneously perceive the coupling characteristics of different physical quantities in the time dimension.

[0036] Next, the constructed three-dimensional state vector is input into the pre-trained reinforcement learning sampling model in real time. In this embodiment, the reinforcement learning sampling model performs nonlinear mapping through a multilayer perceptron to identify inflection points in the feature trajectory. Specifically: when the first derivative of the torque suddenly increases from zero and the angular acceleration shows a negative pulse, it is identified as the elastic contact initiation point of the fastener contacting the laminated surface; when the acoustic emission short-time energy entropy shows a significant peak and the torque slope undergoes a second abrupt change, it is identified as the lamination gap being completely closed by the fastening force and flattened; when the first derivative of the torque gradually decreases and tends to level off, it is identified as the material yield switching point where the material enters the plastic deformation stage. The key physical feature points identified above are defined as the process mode switching points.

[0037] Subsequently, the reinforcement learning sampling model calculates the feature information abundance requirement for the current stage based on the current process mode and its evolution trend. During the interlayer flattening stage, due to the drastic signal fluctuations and the inclusion of critical interlayer components, the model determines that the information abundance requirement is high. Consequently, it outputs a sampling adjustment command containing a sampling period scaling factor (e.g., setting the scaling factor to 0.1, i.e., increasing the sampling frequency by 10 times), driving the edge acquisition terminal to adjust the hardware sampling timer. Specifically, the sampling adjustment command is sent in real time to the hardware clock management module of the edge acquisition terminal via the fieldbus. By modifying the value of the clock divider register inside the FPGA, the hardware-level switching of the data acquisition trigger frequency within a nanosecond delay is achieved.

[0038] Finally, by comparing the envelope similarity between the original high-frequency reference signal and the resampled signal, the reconstructed signal gain is calculated in real time and fed back as a reward signal to the reinforcement learning sampling model. In the training iteration process, the KL divergence between the current policy and the old policy is calculated using the near-end policy optimization algorithm. The parameters of the policy network are updated by pruning the objective function. Through this closed-loop feedback mechanism, the model can adaptively correct the timing and scaling of the sampling adjustment command output, ensuring that complete degradation feature information can be captured at the optimal sampling rate under the assembly conditions of composite materials of different thicknesses.

[0039] By using multidimensional feature extraction and reinforcement learning to identify process switching points, the sampling frequency can be adaptively adjusted. This mechanism accurately captures the key nodes of heterogeneous material fastening, optimizes data redundancy while ensuring the abundance of key information, and enhances the dynamic matching between the sensing system and physical conditions by combining real-time correction with the PPO algorithm, providing a high-fidelity data foundation for the accurate decoupling of subsequent degradation features.

[0040] Furthermore, the reinforcement learning model includes: a feature-aware network: using a one-dimensional convolutional neural network to receive the three-dimensional state vector and extract local interaction features of multi-source data streams in the temporal domain; a temporal dependency network: using gated recurrent units to perform sequence modeling on the local interaction features, generating a long-range dependency feature vector containing the evolution trend of the fastening process; an Actor execution branch: acting as the execution entity of the policy network, using fully connected layers to perform nonlinear mapping on the long-range dependency feature vector, outputting the sampling period scaling factor conforming to a Gaussian distribution, and generating sampling adjustment instructions; a Critic evaluation branch: sharing the feature-aware network and the temporal dependency network with the Actor execution branch, used to calculate the state value estimate corresponding to the current fastening state; and a policy constraint optimizer: using a proximal policy optimization algorithm to calculate the probability ratio between the current policy and the historical policy, and forcibly constraining the update step size of the neural network according to the pruning probability ratio function.

[0041] In the specific data processing flow, the core of the reinforcement learning sampling model consists of a parameterized policy network (i.e., the Actor execution branch). This policy network receives high-dimensional feature vectors from the output of a time-dependent network and performs feature dimensionality reduction and nonlinear activation mapping through two layers of fully connected neural networks. Specifically, the output layer of the policy network uses Softmax to transform the feature vectors into a sampling probability distribution, and samples from this distribution to obtain the specific sampling period scaling factor. In this way, the policy network achieves end-to-end automatic mapping from process modal features (data space) to hardware sampling frequency (physical execution space). During the model update phase, the policy constraint optimizer uses a proximal policy optimization algorithm to calculate the gradient of the current policy network. By maximizing the reward signal (such as signal reconstruction gain), the parameters of the policy network are continuously adjusted, enabling the output of the sampling command to adapt to the transient changes of the heterogeneous material interface.

[0042] Specifically, in the feature perception stage, the feature perception network directly receives a three-dimensional state vector sequence composed of the first derivative of torque, angular acceleration, and short-time energy entropy. The one-dimensional convolutional layer of this network is configured with 32 convolutional kernels with a size of 3 and a stride of 1. It performs sliding convolution operation on the input sequence to extract the local interaction features of multi-source data in the micro-temporal sequence. Through the nonlinear mapping of the convolutional layer, the discrete physical quantity features are transformed into feature maps in a high-dimensional space, thus completing the preliminary feature recognition of waveform distortion in the early contact stage of fastening.

[0043] Subsequently, the extracted local interaction features are input into the temporal dependency network, which adopts a gated recurrent unit structure with 128 hidden layer neurons. The GRU unit performs temporal accumulation modeling of local features through internal update and reset gate mechanisms, captures the dynamic evolution features of the fastening process from the elastic contact point to the yield switching point, and generates a long-range dependency feature vector containing the current fastening condition context information. This vector not only contains the current physical state, but also encodes the evolution trend of the process history.

[0044] In the decision output and evaluation phase, the model processes data in parallel using an Actor-Critic architecture. The Actor execution branch receives the long-range dependent feature vector, performs nonlinear mapping through two fully connected layers with 64 and 32 neurons respectively, and outputs the mean and standard deviation parameters representing the sampling distribution. Based on the output parameters, a Gaussian probability distribution is constructed, and a specific sampling period scaling factor is obtained from it. This generates a sampling adjustment instruction to be sent to the edge acquisition terminal. The Critic evaluation branch shares all weight parameters of the feature perception network and the temporal dependency network with the Actor execution branch. Its output maps the long-range dependent feature vector to a scalar value through a fully connected layer, namely the state value estimate corresponding to the current tight state. This score is used to evaluate the contribution of the current sampling step size selection to the signal reconstruction gain.

[0045] Finally, the policy-constrained optimizer updates the network parameters using a near-end policy optimization algorithm. In the specific update process, the optimizer calculates the probability ratio between the current policy and historical policies in the state space in real time. By introducing a pruning probability ratio function, the fluctuation range of the policy update is forcibly limited to a certain limit. Within the interval, where the clipping factor is... In this embodiment, the value is set to 0.2. This constraint mechanism ensures that the neural network will not generate excessively large gradient step size updates when faced with signal abrupt changes caused by gaps in heterogeneous material stacks, thereby ensuring the stability of the sampling adjustment command output and the continuity of physical logic.

[0046] The process of extracting the residual feature vector includes: using the angle sequence as a reference, performing angular domain resampling on the drive motor current sequence; removing non-stationary random noise unrelated to the shaft rotation frequency by using the angular domain synchronous averaging operator, extracting deterministic component features characterizing the mechanical state of the tightening shaft transmission chain, and using the deterministic component features as input for constructing the symmetric positive definite matrix manifold.

[0047] Specifically, the drive motor current sequence and output shaft angle sequence are first acquired synchronously throughout the entire tightening cycle. Due to the resistance of the heterogeneous material during the loading process, the rotational speed of the tightening tool fluctuates nonlinearly, causing frequency modulation of mechanical fault characteristics (such as gear meshing frequency) in the time-domain current signal. The angle sequence is used as an independent variable to establish a one-to-one mapping relationship between the current sampling points and the absolute rotation angle position. Using the angle sequence as a sampling reference, non-uniform interpolation resampling is performed on the time-domain current sequence. Specifically, a cubic spline interpolation algorithm is used to convert the original current sequence from equal time intervals to equal angular intervals. In this embodiment, the resampling accuracy is set to extract N sampling points for every 360 degrees (one revolution) (N=1024 in this embodiment) to generate the angular domain current sequence. This step eliminates the influence of rotational speed fluctuations on the spectral distribution, so that the periodic characteristics of the mechanical transmission components are stabilized in the angular domain.

[0048] Based on the reduction ratio parameters of the internal transmission chain of the tightening shaft, the resampled angular domain current sequence is divided into multiple segments. The length of each segment precisely corresponds to one rotation cycle of the monitored target component (such as a planetary gear or sun gear). By identifying the zero-position pulse of the angle encoder or the preset angle increment as the starting trigger point, it is ensured that each segment is strictly aligned in phase. The angular domain synchronous averaging operator is called to perform an arithmetic average operation on the multiple phase-aligned angular domain current segments. According to the signal statistical characteristics, the periodic mechanical features related to the shaft rotation frequency (such as gear meshing and bearing eccentricity) are preserved and enhanced during the averaging process, while the non-stationary random noise generated by friction of heterogeneous materials and gap closure, as well as the background interference unrelated to rotation, tend to cancel each other out as the number of averaging times increases.

[0049] After processing by the synchronous averaging operator, deterministic component features characterizing the mechanical state of the tightening shaft drive chain are extracted. These components eliminate random load fluctuations caused by external operating conditions and reflect the fatigue, wear, or assembly clearance state within the reduction mechanism. Finally, the extracted deterministic component features (i.e., the processed angular domain current vector) are used as the core input, combined with the synchronously acquired torque residual features, to construct a symmetric positive definite matrix manifold by calculating the covariance matrix. This process ensures that the input features entering the Riemann manifold alignment stage have extremely high signal-to-noise ratio and are not correlated with the body degradation, thus physically eliminating interface interference during the fastening process of heterogeneous materials. Through angular domain resampling and synchronous averaging, random interference unrelated to shaft rotation is eliminated, and the influence of speed fluctuations is removed, solving the technical problem of aliasing between mechanical body degradation features and interface friction noise, and ensuring that the features entering the health assessment manifold have extremely high physical purity.

[0050] By constructing a deep reinforcement learning architecture that integrates local feature extraction from a one-dimensional convolutional neural network with long-range temporal modeling using GRU, this application enhances the ability to capture microscopic jumps and evolutionary trends in tight signals. The Actor-Critic architecture utilizes a shared perceptual layer to improve feature utilization efficiency, and, in conjunction with the pruning probability ratio constraint of the PPO algorithm, forcibly limits the policy update step size, thereby effectively suppressing oscillations during training. This ensures that the sampling adjustment command maintains high output stability and real-time response accuracy even when the physical modes of heterogeneous materials switch drastically, providing robust dynamic sampling support for subsequent refined feature decoupling.

[0051] Furthermore, the process of obtaining the resampling sequence includes: determining the target sampling step size for each channel according to the scaling factor in the sampling adjustment instruction; performing linear interpolation on the original data stream according to the target sampling step size to fill in the missing data points of the low-frequency sampling channels; and mapping the aligned multi-source data to a unified discrete time axis through a timestamp matching algorithm to construct a multi-dimensional synchronization feature matrix, which serves as the resampling sequence.

[0052] Specifically, firstly, based on the sampling adjustment instructions output by the reinforcement learning sampling model, the resampling parameters for each data channel are determined. In the specific execution process, the scaling factor in the terminal parsing instruction is collected and multiplied with the basic sampling period of each sensor to determine the target sampling step size for each channel. During specific execution, the system maps the target sampling step size to the count value of the FPGA's internal clock management module via the fieldbus. By rewriting the frequency divider register, the system directly intervenes in the toggle cycle of the hardware sampling clock to achieve nanosecond-level sampling frequency switching response. When the scaling factor is less than 1, the target sampling step size decreases, and the corresponding sampling frequency increases to capture the high-frequency transient characteristics at the moment of solid mode switching. When the scaling factor is greater than 1, the target sampling step size increases to reduce data redundancy during the process stabilization phase.

[0053] Next, linear interpolation is performed on the original data stream. Since the original sampling frequencies of the torque, angle, and current sequences are low, while acoustic emission and ultrasonic stress pulse signals have high physical bandwidth, when the system performs high-frequency resampling adjustment, data gaps will occur in the low-frequency channel at the target sampling time point. At this time, the time-domain position of the missing point in the original sequence is located according to the target sampling step size, and the two adjacent original discrete sample values ​​before and after the missing point are retrieved. An estimated value of the missing point is calculated using a first-order linear proportional relationship, thereby filling in the data gaps in the low-frequency sampling channel. This process ensures that all sensing dimensions have the same temporal resolution after resampling.

[0054] Subsequently, the resampled asynchronous data stream is integrated using a timestamp matching algorithm. In this embodiment, the edge acquisition terminal uses the trigger clock of the ultrasonic stress pulse signal as the main reference and maps the aligned torque, angle, current and structural response vibration characteristics to a unified discrete time axis defined by the main reference. The algorithm performs time-domain rearrangement of each dimension feature by searching the overlap between the timestamp of each channel and the reference time axis, ensuring that each set of sampling points physically and logically represents the system state at the same fastening instant.

[0055] Finally, the calibrated multidimensional feature data are stacked in chronological order to construct a multidimensional synchronization feature matrix, which serves as the resampling sequence. In this matrix, the row dimension represents discrete time steps, and the column dimension corresponds to the synchronized torque component, angle component, current component, ultrasonic stress pulse signal, and structural response vibration characteristics, respectively. This synchronized data structure eliminates the sampling phase deviation caused by the differences in heterogeneous sensor hardware, providing rigorous data support for subsequent cross-modal feature decoupling and high-dimensional manifold alignment.

[0056] By executing linear interpolation and timestamp matching algorithms, resampling alignment of heterogeneous sensing data was achieved, effectively supplementing the information gaps of low-frequency sensors and constructing a multi-dimensional synchronous feature matrix under a unified time axis. This process ensures the tight correlation of multi-source data of electricity, force, and sound in the time domain, providing a high-precision aligned data carrier for subsequent cross-modal feature extraction and information decoupling, and improving the accuracy of calculation.

[0057] Further, the process of obtaining the aligned feature vector includes: using the angular components in the resampled sequence as a reference, performing equal-angle resampling on the drive motor current sequence to map the time-domain current signal to the angular domain space; performing a fast Fourier transform on the angular domain current signal to obtain the order spectrum, extracting the characteristic order energy components that are integer multiples of the motor rotation frequency, and constructing the energy feature; estimating the marginal probability density distributions of the energy feature and the structural response vibration feature respectively, and constructing the joint probability density distribution between the two; calculating the mutual information between the energy feature and the structural response vibration feature based on the marginal probability density distribution and the joint probability density distribution, and using the normalized weighting factor obtained based on the mutual information calculation to remove the interference information related to the structural response vibration feature from the energy feature, thereby obtaining the residual feature vector characterizing the degradation state of the tool transmission chain.

[0058] Specifically, a coordinate space transformation from the time domain to the angular domain is performed. Using the angular components in the resampled sequence as a reference, the current sequence of the drive motor is resampled at equal angles. During the process, discrete scales are established on the angle axis with a constant angular increment of 0.1 degrees. The current amplitude corresponding to each angular scale is calculated using a linear interpolation algorithm. The time-varying current signal is mapped to the angular domain space with the rotation angle as the independent variable to obtain the angular domain current signal. Then, the tool body features are extracted through order analysis. A fast Fourier transform is performed on the angular domain current signal to generate an order spectrum with the order as the abscissa. Based on the number of pole pairs and transmission ratio parameters of the tool drive motor, the 1st, 2nd, and 3rd order characteristic energy components that are integer multiples of the motor rotation frequency are located and extracted from the order spectrum. By performing square root calculation on the amplitudes of the above characteristic orders, the energy features characterizing the tool's dynamic state are constructed.

[0059] Subsequently, the coupling strength between the quantification tool characteristics and the operating condition interference is determined by retrieving the energy characteristics and the structural response vibration characteristics within the same observation window. The marginal probability density distributions of both are estimated using the Gaussian kernel density estimation method, and a joint probability density distribution is constructed based on a two-dimensional normal kernel function. Integration is then performed using the aforementioned probability distribution function to calculate the mutual information between the energy characteristics and the structural response vibration characteristics. This value serves as a quantification index, characterizing the proportion of the component in the motor current characteristics affected by the interaction interference of the material layer interface. In estimating the marginal probability density distribution and the joint probability density distribution, this embodiment employs the Gaussian kernel density estimation method. The calculation rule for the kernel function with h is set as follows: 1.06 multiplied by the sample standard deviation, then multiplied by the sample size N raised to the power of -0.2. This adaptive bandwidth setting ensures that the probability distribution curve can effectively smooth random noise while accurately locking the energy fluctuation characteristics at the moment of heterogeneous material fastening. It should be noted that the mutual information is numerically equal to the self-information entropy of the energy characteristic minus its conditional entropy under known vibration characteristics. This definition ensures that the weighting factor can accurately quantify the coupling components between the two nonlinear signals.

[0060] In this step, the following physical causal assumption is established: the unsteady frictional vibration generated during the fastening of heterogeneous materials will be transformed into an additional damping load on the transmission chain. The load pulsation directly causes the stator current of the drive motor to generate excited fluctuations. Therefore, the current energy characteristics contain two parts: one part is the steady component reflecting the state of the transmission components of the tool body, and the other part is the disturbance component induced by the above-mentioned vibration load. Mutual information is defined here as a measure of the strength of the causal relationship, used to quantify the proportion of the disturbance component in the total energy.

[0061] To ensure that the mutual information calculation results have a clear physical orientation, this embodiment first verified the unidirectional causal chain between vibration and current through controlled variable experiments: In a laboratory environment, by keeping the assembly tool rotation speed constant and changing the interface of heterogeneous materials with different damping characteristics, it was observed that the change in the vibration energy of the structure response preceded the envelope modulation of the current waveform of the drive motor. Furthermore, the Granger causality test was used to confirm that the vibration characteristics have a significant unidirectional time-series predictive ability for the current energy characteristics, thereby eliminating the spurious correlation between the two due to interference from a common third-party factor.

[0062] To achieve feature stripping, a process is executed that maps mutual information to weighting factors. Specifically, the self-information entropy of the energy feature is first calculated using the constructed marginal probability density distribution of the energy feature through integration. Then, the mutual information between the energy feature and the structural response vibration feature is divided by this self-information entropy to calculate the normalized proportion of the overlapping information between the two to the total energy feature. This proportion is defined as a weighting factor, with a value ranging from 0 to 1, precisely quantifying the proportion of the interference component in the motor current signal caused by the interaction vibration of the heterogeneous material interface.

[0063] In the specific calculation of mutual information, this embodiment adopts a 64-by-64 discrete grid summation method: the value ranges of energy features and vibration features are divided into 64 equal intervals. At each grid point, the joint probability density distribution value is divided by the product of the corresponding two marginal probability density distribution values. The logarithm of the quotient is taken to the base 2, and then the logarithm is multiplied by the joint probability of that point. Finally, the summation is performed on all grid points. If the denominator is close to 0 during the calculation, the lower limit threshold is set to 0.000001 to ensure the stability of the numerical calculation.

[0064] Finally, a feature weighted subtraction operation is performed on the energy feature based on the weighting factor. The specific operation process includes: First, a unit constant with a value of 1 is retrieved and subtracted from the weighting factor to calculate the compensation coefficient representing the proportion of the tool body feature; then, the values ​​of each dimension in the energy feature vector are multiplied element-wise with the compensation coefficient. This process proportionally subtracts the energy component affected by material stack gaps and frictional fluctuations at the physical level, and removes the background noise caused by the heterogeneity of heterogeneous materials by suppressing signal components strongly correlated with environmental vibrations.

[0065] The specific logic of the feature weighted subtraction operation is as follows: the compensation coefficient is obtained by subtracting the weight factor from the value 1, and each element in the original energy feature vector is multiplied by the compensation coefficient. This operation physically realizes the precise deduction of the interference energy strongly correlated with the interface vibration from the total energy based on the interference ratio quantized by mutual information, thereby extracting the pure residual feature vector.

[0066] Through the above process, the residual feature vector characterizing the degradation state of the tool transmission chain is finally extracted. This vector has undergone information subtraction processing based on weighting factors, which eliminates the interference contribution of physical differences of the assembly object to the current signal. This ensures that the numerical offset in the feature space only points to the degradation of the body performance caused by mechanical wear or electrical performance decline inside the tool, providing an environment-independent feature benchmark for subsequent cross-working condition alignment and evaluation.

[0067] By performing angular domain reconstruction and mutual information feature decoupling, the tool degradation component was effectively extracted from complex physical disturbances. Order tracking ensured the stability of energy features during dynamic processes. Mutual information was used to quantify the degree of nonlinear coupling between motor current and structural vibration, and based on this, energy disturbances caused by the damping characteristics of heterogeneous materials were removed. Thus, residual vectors characterizing the degradation state of the tool drive chain were extracted from strong background noise, significantly reducing the interference contribution of material heterogeneity to feature observation and improving the sensitivity and analytical accuracy of feature signals to tool wear.

[0068] Further, the process of obtaining the alignment feature vector includes: synchronously fusing the torque component, angle component, ultrasonic stress pulse signal, and residual feature vector; constructing a symmetric positive definite matrix representing the tightness state space by calculating the covariance matrix, which serves as the healthy state manifold; performing a logarithmic-Euclidean transformation on the symmetric positive definite matrix using a logarithmic mapping function, projecting it from a nonlinear Riemannian space to an Euclidean tangent space to obtain a high-dimensional feature tangent vector; retrieving the mean of the reference manifold under the pre-stored tool standard healthy state and mapping it to the same tangent space to obtain a reference center vector; calculating the relative offset of the high-dimensional feature tangent vector relative to the reference center vector, and eliminating the global distribution offset caused by batch differences in heterogeneous materials and environmental fluctuations through a centering translation operation to obtain the alignment feature vector. The specific process is as follows: Figure 2 As shown.

[0069] First, spatial dimensional fusion of multi-source heterogeneous data is performed. The synchronized and aligned torque component, angle component, ultrasonic stress pulse signal, and decoupled residual feature vector are vertically stacked. In the specific execution process, the system constructs a multi-dimensional observation matrix with a sliding calculation window of 200 sampling points. Subsequently, the covariance matrix of this observation matrix is ​​calculated, and the second-order coupling statistical features between different physical quantities are extracted, thereby constructing a one-dimensional symmetric positive definite matrix. This matrix serves as the healthy state manifold, and its internal elements accurately characterize the correlation distribution between the tool transmission state and the fastening process parameters in Riemannian geometry. Specifically, the symmetric positive definite matrix is ​​a 4x4 covariance matrix. Its construction process is as follows: within a sliding window of 200 sampling points, data from four channels—torque component T, angle component A, ultrasonic stress pulse signal U, and residual feature vector R—are extracted to form a 200x4 observation matrix. Calculate the sample covariance of the observation matrix to obtain a 4x4 symmetric positive definite matrix M. To ensure the strict positive definiteness of the matrix, a minimum value of 0.000001 is uniformly added to the diagonal elements of M as a regularization term.

[0070] Next, a nonlinear projection from Riemannian space to Euclidean tangent space is performed. Since the symmetric positive definite matrix does not constitute a vector space, a logarithmic-Euclidean transformation is performed on the symmetric positive definite matrix using a logarithmic mapping function. The specific execution process is as follows: eigenvalue decomposition is performed on the symmetric positive definite matrix to obtain the eigenvalue matrix and eigenvector matrix; the natural logarithm function is used to perform logarithmic operations on the diagonal elements of the eigenvalue matrix, and then the symmetric matrix in the logarithmic space is reconstructed through the eigenvector matrix; finally, the upper triangular elements of the symmetric matrix are extracted and arranged in row order to obtain the high-dimensional eigenvector.

[0071] The specific calculation steps of the logarithmic-Euclidean transformation are as follows: The 4x4 symmetric positive definite matrix M is decomposed into an orthogonal matrix U and a diagonal matrix L, where the diagonal elements of L are eigenvalues ​​L1, L2, L3, and L4. The natural logarithmic values ​​of the four eigenvalues ​​are calculated to obtain a new diagonal matrix Log_L, whose diagonal elements are ln(L1), ln(L2), ln(L3), and ln(L4). The matrix Log_M is calculated as U multiplied by Log_L and then multiplied by the transpose of U, resulting in a 4x4 symmetric matrix in logarithmic space. The upper triangular portion of Log_M, consisting of 10 elements, is extracted row-wise, arranged as follows: the 4 elements in the first row, followed by the last 3 elements in the second row, the last 2 elements in the third row, and the last element in the fourth row. These are then combined to form a 10-dimensional high-dimensional eigenvector.

[0072] Subsequently, the baseline state extraction and mapping are performed. The baseline manifold mean of the tool under the standard healthy state is retrieved from the database. The baseline manifold mean is obtained by collecting 50 sample matrices under the standard healthy state, performing arithmetic mean calculation on the sample vectors after projection to the tangent space, and mapping them back to the manifold space to obtain the statistical center. Using the same eigenvalue decomposition and logarithmic operation logic as the previous steps, the baseline manifold mean is projected to the same tangent space to generate the corresponding baseline center vector.

[0073] It should be noted that, for a symmetric positive definite matrix manifold, its theoretical statistical center should be the Riemann geometric center. In this embodiment, considering that the 50 standard health samples were all collected from new tools of the same model and batch, their original distribution in the manifold space has a high degree of clustering, satisfying the local neighborhood assumption of the tangent space. Therefore, the arithmetic mean in the tangent space is used as an approximate estimate of the Riemann geometric center. This approach avoids complex recursive iterative algorithms while maintaining the representativeness of the physical center, significantly reducing the cost of generating the benchmark vector. Furthermore, experimental results show that the Euclidean distance deviation between the approximate center and the theoretical geometric center is less than 0.01, fully meeting the engineering accuracy requirements for tool health rating.

[0074] Finally, global distribution correction is performed by linear displacement within the tangent space. The vector difference between the high-dimensional feature tangent vector and the reference center vector is calculated to obtain the relative offset. Subsequently, a centering translation operation is performed using this relative offset to translate the origin of the coordinate system of the current observation point to the reference center. From a physical and logical perspective, this translation operation cancels out the overall translation of the feature space caused by differences in the hardness of different batches of materials, fluctuations in ambient temperature, or initial drift of the sensor by subtracting the reference deviation, thereby eliminating the global distribution offset. Finally, the aligned feature vector representing the true degradation increment of the tool body is obtained. This vector ensures that the health assessment has a unified observation benchmark under different material batches and environmental fluctuations.

[0075] By constructing a symmetric positive definite matrix and performing a logarithmic-Euclidean transformation, linearization projection of nonlinear manifold features is achieved. By utilizing the centralized translation in the tangent space, the global distribution drift caused by material batch differences and environmental fluctuations is effectively corrected, ensuring the comparability of heterogeneous features under a unified benchmark. This enhances the robustness of feature expression, allows the evaluation model to focus more on the geometric shift caused by tool ontology degradation, and significantly improves the algorithm's generalization ability in complex working conditions.

[0076] Further, the process of outputting the health level result includes: inputting the alignment feature vector as the core evaluation variable into the reinforcement learning evaluation model, and simultaneously injecting the gradient components of the structural response vibration characteristics in the resampled sequence as environmental observations in real time to construct a dynamic observation space; calculating the geodesic distance of the alignment feature vector relative to the preset standard state manifold in Riemann space to obtain an initial deviation scalar characterizing the degree of deviation of the tool state from the baseline; the reinforcement learning evaluation model uses the geodesic distance and the environmental observations as input states, and uses a policy network to generate decision threshold correction coefficients for different process modes, and performs evaluation strategy optimization; using the decision threshold correction coefficients to perform error compensation calculations on the geodesic distance, removing the unsteady components affected by the gaps in the heterogeneous material stacks, and extracting the corrected feature deviation reflecting the degree of performance degradation of the tool body; mapping the corrected feature deviation to a preset health evaluation index system, and outputting the health level result characterizing the tool's operating accuracy and mechanical fatigue state, the specific process is as follows. Figure 3 As shown.

[0077] Specifically, firstly, a dynamic observation space for evaluation is constructed, with the aligned feature vector as the core evaluation variable. Simultaneously, the gradient component of the structural response vibration characteristics (acoustic emission signal) in the resampled sequence is extracted. Specifically, this gradient component is obtained by calculating the first-order rate of change of the acoustic emission envelope signal within the current fastening window, which is used to reflect the instantaneous fluctuation of the friction state of the heterogeneous material. The aligned feature vector is then concatenated with this gradient component to form an input tensor characterizing the coupling features of the tool state-environment disturbance.

[0078] Next, a geometric measure characterizing the degree of degradation is calculated. Since the features have been projected onto the Euclidean tangent space and a centering translation has been performed, the geodesic distance between the symmetric positive definite matrix manifolds is equivalently transformed in the tangent space into the calculation of the Frobenius norm (L2 norm) of the aligned feature vectors. By calculating the magnitude of this norm, an initial deviation scalar characterizing the degree of deviation of the current tool state from the standard reference manifold is obtained. This scalar initially reflects the total deviation, including tool wear and steady-state environmental disturbances. It should be clarified that the geodesic distance in this embodiment is defined based on the logarithmic-Euclidean transformation. According to the logarithmic-Euclidean metric principle, the nonlinear geodesic distance on a symmetric positive definite matrix manifold, after logarithmic mapping, is equivalent to its Euclidean distance (i.e., L2 norm) in the tangent space (logarithmic space). Therefore, by calculating the L2 norm of the aligned eigenvector, the geodesic distance value representing the degree of deviation of the current tool state from the standard reference manifold can be accurately obtained. This processing method simplifies the complex Riemannian geometric distance calculation into an efficient vector norm calculation, while maintaining the rigor of the geometric measurement of physical quantities in the manifold space. By calculating the L2 norm, the 10-dimensional aligned eigenvector is compressed into a 1-dimensional initial deviation scalar, achieving feature dimensionality reduction while retaining essential degenerative information and eliminating computational redundancy in high-dimensional space.

[0079] Subsequently, a reinforcement learning evaluation model is used to perform decision optimization. This model employs a deep neural network with a three-layer fully connected structure. Using the initial deviation scalar and the environmental observation (vibration gradient) as input states, the model performs nonlinear mapping through its policy network, outputting a judgment threshold correction coefficient for the current fastening process mode (such as the interlayer flattening stage). This coefficient is a continuous value in the interval [0.82, 1.15], used to quantify the contribution weight of the current environmental disturbance to the deviation scalar.

[0080] First, the reinforcement learning evaluation model is constructed. In this embodiment, the model adopts a near-end policy optimization architecture and uses offline pre-training to enable the model to acquire adaptive compensation capabilities for non-steady-state disturbances of heterogeneous materials. The core logic of the model is to use the external environmental gradient to correct feature deviations in real time, thereby stripping away non-entity degradation interference components. Next, the dynamic observation space is constructed. The model receives a two-dimensional state vector as input in real time: the first dimension is set as the initial deviation scalar converted from the geodesic distance in Riemann space to reflect the initial deviation of the tool performance; the second dimension is set as the environmental observation value, that is, the instantaneous gradient component obtained by performing sliding window first-order difference calculation on the resampled acoustic emission signal. The two-dimensional vector constitutes the feature benchmark for the model to perceive the coupled state of tool degradation and interface disturbance.

[0081] Subsequently, feature mapping is performed through a policy network. The evaluation model integrates a multilayer perceptron with a three-layer neuron structure. The input layer contains two neurons, corresponding to a two-dimensional state vector. The two-dimensional state vector is formed by recombining the 1-dimensional initial bias scalar obtained by the aforementioned dimensionality reduction process with the 1-dimensional environmental observation (i.e., the gradient component of the acoustic emission signal). This design completes the mapping from the original 11-dimensional features to the model's 2-dimensional input, significantly improving the convergence speed of the reinforcement learning model under varying conditions. The hidden layer consists of two fully connected layers: the first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. All hidden layers use the ReLU activation function to ensure that the model can effectively capture the nonlinear mapping relationship between tool bias and interface vibration gradient. Furthermore, a parameterized linear scaling operator is added to the output layer, which contains one neuron. The original output is restricted to the [0, 1] interval by the Sigmoid activation function. In this operator, the range of the correction coefficient is automatically determined by the model during the pre-training stage. In this embodiment, after reinforcement learning training on 10,000 sets of heterogeneous material fastening samples, the operator automatically converges to a specific compensation interval [0.82, 1.15]. When the environmental observation shows that the interface vibration gradient increases abnormally, the policy network drives the operator output to tend to the lower limit of the interval, realizing the proportional stripping of the interference components. In addition, the model's reward mechanism and policy optimization are executed.

[0082] In the offline pre-training phase, to overcome the feedback delay problem caused by long-term mechanical wear, this embodiment adopts a hierarchical reward mechanism: the short-term reward is set as the consistency score between the current cycle output level and the predicted trend of the Hidden Markov Model, which is used to constrain the stability of the evaluation results; the long-term reward is set as the negative mean square error between the evaluation level curve and the actual wear evolution curve measured manually over the entire life cycle. Through this composite reward function of short-term consistency and long-term accuracy, the evaluation model is guided to output a health level that conforms to the physical degradation logic in each tightening cycle. By constraining the update step size through the pruning objective function of the PPO algorithm, the model finally learns a steady-state control strategy: performing negative feedback compensation when the environment fluctuates drastically to offset the initial deviation jitter caused by material gap fluctuations; and making the correction coefficient tend to 1.0 when the environment is stable, thereby maintaining the original feature strength.

[0083] Next, error compensation calculation is performed on the geodetic distance using the judgment threshold correction coefficient. Specifically, the geodetic distance is multiplied and compensated by the judgment threshold correction coefficient. In this process, the correction coefficient serves as a dynamic adjustment factor to scale the original geodetic distance proportionally. The geometric deviation is calibrated in real time using the AI ​​decision results, thereby removing the unsteady components affected by the gaps between the heterogeneous materials. Through this calculation, false deviation pulses caused by external factors such as gasket compression and abrupt changes in interlayer friction are effectively eliminated, ensuring the purity of the evaluation data.

[0084] Finally, through error compensation calculation, an index reflecting only the degree of performance degradation of the tool's transmission chain (such as wear of the planetary reducer and decrease in motor efficiency) is extracted, namely the corrected characteristic deviation. The index mapping is performed and the result is output. The corrected characteristic deviation is input into a preset health evaluation index system. In this system, multiple deviation threshold ranges are set: when the deviation is below 0.3, the output is health level I (excellent operating accuracy, no obvious fatigue); when the deviation is in [0.3, 0.7], the output is health level II (warning state, risk of mechanical fatigue); when the deviation exceeds 0.7, the output is health level III (performance degradation, maintenance recommended). Through this process, the quantitative classification and output of tool operating accuracy and mechanical fatigue status are realized.

[0085] After outputting the health level result, the method further includes: performing time-series calibration on the correction feature deviation of the current fastening cycle using a hidden Markov model; the state transition probability matrix of the hidden Markov model is preset as a lower triangular matrix according to the irreversible unidirectional evolution law of mechanical damage, so as to constrain the unidirectional degradation trend of the correction feature deviation in the time dimension and eliminate the non-physical grade jump caused by the single heterogeneous material assembly anomaly.

[0086] Specifically, a Hidden Markov Model (HMM) is first constructed, whose hidden state space S corresponds to the physical health level of the assembly tool. In this embodiment, S = {S1, S2, S3}, representing "healthy", "warning", and "fault" respectively. The observed values ​​of the model are set as the corrected feature deviation extracted by the current fastening cycle. Through a preset probability density function (such as a Gaussian distribution), an emission probability distribution between the corrected feature deviation and the hidden state is established to characterize the probability of the deviation value occurring under a specific health state. Specifically, the observed values ​​are set as the corrected feature deviation values ​​output by each fastening cycle, and their values ​​are usually between 0 and 2. The hidden state space S corresponds one-to-one with health levels I, II, and III, where S1 represents health, S2 represents warning, and S3 represents fault.

[0087] According to the irreversible physical degradation law of mechanical damage, the state transition probability matrix A of the hidden Markov model is preset. Specifically, this matrix is constructed in the form of a lower triangular matrix: the element a in the matrix ij represents the probability that the tool transitions from state S i to state S j . Since the element representing reverse degradation in the matrix (i.e., when j < i) a ij is forced to be preset to 0, the unidirectionality of state transition is forcibly constrained, allowing only the tool to transition from "healthy" to "warning" or to maintain the current state from "warning" or transition to "fault", while strictly prohibiting the probability path of reverse transition from "warning" to "healthy".

[0088] In this embodiment, the specific values of the state transition probability matrix A are set as follows: the probability of remaining in the healthy state S1 is 0.95, and the probability of transitioning to the warning state S2 is 0.05; the probability of remaining in the warning state S2 is 0.90, and the probability of transitioning to the fault state S3 is 0.10; the probability of remaining in the fault state S3 is 1.00, and transitions to any other state are prohibited. The above lower triangular matrix parameters are determined by statistical modeling of a large amount of historical wear data, which can forcibly constrain the evaluation results to follow the physical unidirectional degradation trend.

[0089] To avoid false alarms caused by instantaneous noise or random interference of sensors, an observation smoothing and threshold hysteresis mechanism is introduced in this embodiment when performing state transition determination. Specifically, although the state transition matrix physically prohibits reverse degradation, a confidence check rule for the observation sequence is set: only when the corrected feature deviation is continuously in the probability interval of the higher level for N fastening cycles (N = 3 is set in this embodiment), the state transition update is officially triggered. If it is only a single pulse-like deviation jump, the Viterbi algorithm will automatically identify it as observation noise through the path cost function and smooth it, so as to ensure the robustness of the evaluation conclusion to random perturbations while maintaining the physical law of unidirectional evolution of mechanical degradation.

[0090] The comprehensive probabilities of each health state at time t are recorded and compared, and the state with the highest probability is selected as the optimal hidden state at the current time. The emission probability is defined as the probability distribution of specific observation deviations occurring under different health states. In this embodiment, it is assumed to follow a Gaussian distribution with the following parameter configurations: the mean of the observations corresponding to state S1 is 0.15, and the standard deviation is 0.10; the mean of the observations corresponding to state S2 is 0.50, and the standard deviation is 0.15; the mean of the observations corresponding to state S3 is 0.90, and the standard deviation is 0.20. This parameter setting establishes a probabilistic correlation between the deviation value and the hidden state, enabling the Viterbi algorithm to decode the health level sequence that best conforms to physical logic based on a continuous deviation sequence. Through this path-locking mechanism, it is ensured that the output of the health level depends not only on the current instantaneous observation but also on the physical continuity of the historical states. This effectively eliminates non-physical grade jumps caused by single heterogeneous material assembly anomalies, ensuring the unidirectional degradation trend of the health evaluation results in the time dimension.

[0091] Finally, the stable health level result after decoding by the Viterbi algorithm is output. Since this result has been filtered by the physical logic of the lower triangular transition matrix, it reflects the actual wear accumulation process of the tool transmission chain. Based on this rating result, subsequent early warning instructions are issued or maintenance plans are scheduled, achieving high-confidence fault prediction in complex interference environments. Temporal calibration is performed by a hidden Markov model constrained by the lower triangular transition matrix, and the evolution trend is constrained by the irreversible law of mechanical degradation. This solves the problem of non-physical jumps in the evaluation level caused by single assembly anomalies, ensuring the logical robustness and temporal continuity of the health assessment conclusion.

[0092] By combining geodesic distance quantification with environmental gradient observation, dynamic optimization of assessment decisions was achieved. The reinforcement learning model, through threshold compensation, effectively eliminated unsteady-state interference caused by material gaps, accurately extracting the performance degradation of the tool itself. This process reduced assessment bias caused by environmental coupling, ensuring the physical consistency and reliability of health level outputs, and providing quantitative support for predictive maintenance of precision assembly tools.

[0093] By employing multi-source sensing fusion and deep reinforcement learning mechanisms, dynamic characterization of the entire process of heterogeneous material fastening is achieved. An adaptive sampling strategy effectively balances the abundance of key feature information with system data redundancy. Combined with order tracking and mutual information stripping techniques, deep decoupling of tool degradation features and complex material background interference is realized. Riemannian manifold space alignment eliminates global distribution shifts caused by material batch and environmental fluctuations, significantly enhancing the robustness of the sensing system under heterogeneous assembly conditions. This lays a solid technical foundation for improving the accuracy of performance degradation assessment and the reliability of predictive maintenance for precision assembly tools.

[0094] Example 2: This invention provides a method for assessing the health of assembly tools based on tightening data, which is applied to the automated fastening process of the bottom shell and internal frame of a power battery pack for new energy vehicles. This scenario involves a typical heterogeneous material stack: the upper layer is a 2mm thick aluminum alloy shell, the middle layer is a 1.5mm thick elastic composite sealing gasket, and the bottom layer is a steel internal support frame. The tool used is a servo tightening shaft fixed to the end of a six-axis collaborative robot.

[0095] During the tightening operation, the torque sequence and angle sequence are acquired at a frequency of 1.2kHz by the sensor built into the servo tightening shaft. At the same time, the current sequence of the drive motor is captured in real time using a current transformer. For the compression state of the sealing gasket in the laminated material, an acoustic emission sensor is placed on the contact surface between the fixture and the aluminum alloy shell to capture the structural response vibration characteristics in the range of 100kHz to 800kHz. Through the edge acquisition terminal, the above force, electrical and acoustic signals are mapped to a unified discrete time axis to construct a multi-source data matrix.

[0096] Because the torque increases slowly but the acoustic emission signal is intense during the process from initial contact to complete flattening of the gasket, the extracted first derivative of the torque, angular acceleration, and short-time energy entropy are input into the reinforcement learning sampling model. In the specific scenario of this embodiment, when the model detects a continuous increase in short-time energy entropy and steady-state fluctuations in angular acceleration, it determines that the interlayer flattening point (gasket compression stage) has been reached. At this time, the reinforcement learning model outputs a sampling adjustment command with a scaling factor of 0.15, instantly increasing the system sampling frequency by 6.6 times to capture the micro-modulation information of the tool load caused by the nonlinear deformation of the gasket with high fidelity.

[0097] After acquiring the resampled sequence, feature decoupling is performed to address the characteristic that the aluminum alloy shell is prone to structural resonance under high-frequency rotation. Using the angular component as a reference, the drive motor current signal is mapped to the angular domain space, and the energy components of the first to third orders are extracted. The mutual information between this energy feature and the vibration feature of the aluminum alloy shell is calculated using kernel density estimation. Based on physical modeling analysis, the structural resonance of the aluminum alloy shell at the moment of tightening will act in reverse on the motor through the output shaft, generating a current ripple at the synchronous frequency. In this embodiment, mutual information calculation identifies that approximately 18% of the energy fluctuation in the current signal has a statistical causal correlation with the shell resonance feature. Subsequently, a weighted subtraction operation is performed using the calculated normalized weighting factor (0.18). The physical essence of this operation is to remove the stimulated current response caused by shell resonance from the total energy, thereby extracting the pure residual feature vector that only characterizes the wear state of the planetary reducer body of the tightening shaft.

[0098] Given the differences in hardness and heat treatment conditions among different batches of aluminum alloy housings, torque, ultrasonic stress pulses, and residual eigenvectors are fused to construct a symmetric positive definite matrix characterizing the current fastening state space. The mean value of the reference manifold under the pre-stored standard working condition (laboratory standard part fastening) is retrieved, and the current observed manifold and the reference center are mapped to the same space through logarithmic-Euclidean transformation. By performing a centering translation operation, the global torque reference offset caused by the higher hardness of this batch of aluminum alloy housings is eliminated, and the eigenvectors are realigned to the standard base, ensuring that the evaluation results only reflect the performance drift of the tightening shaft body.

[0099] Finally, the aligned feature vector and vibration gradient components are input into the reinforcement learning evaluation model to calculate the geodesic distance (L2 norm) between the current feature point and the standard manifold, obtaining the initial deviation scalar. In this embodiment, the policy network detects that the environmental observation value (vibration gradient) is at a low level, determines that the current fastening process is stable, and outputs a correction coefficient of 1.02. This coefficient reflects that under a low vibration gradient environment, the masking effect of the material stack gap on the tool performance deviation is extremely low, and only a very small gain compensation is required. After error compensation calculation, a corrected feature deviation of 0.459, reflecting the fatigue degree of the tool reducer tooth surface, is obtained. After mapping by the index system, a health level of II is output, and the lubrication performance of the reducer is indicated as declining. It is recommended to perform maintenance after 500 cycles, achieving high-precision health decision-making under complex material interference.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the health of assembly tools based on tightening data, characterized in that, include: Acquire multi-source data during the fastening process of heterogeneous materials, including torque sequence, angle sequence, drive motor current sequence, ultrasonic stress pulse signal, and structural response vibration characteristics; The torque sequence, angle sequence, and structural response vibration features are constructed into a three-dimensional state vector, which is input into the reinforcement learning sampling model to identify the process mode switching point and output sampling adjustment command. The multi-source data is then resampled to obtain the resampled sequence. The drive motor current sequence in the resampled sequence is subjected to order tracking processing to extract energy features; The mutual information between the energy characteristics and the structural response vibration characteristics in the resampled sequence is calculated to obtain the residual feature vector; The torque component, angle component, and ultrasonic stress pulse signal in the resampled sequence are combined with the residual feature vector to construct a healthy state manifold. The healthy state manifold is then projected onto the tangent space of the Riemann manifold to obtain the aligned feature vector. The alignment feature vector is input into the reinforcement learning evaluation model, and the gradient components of the structural response vibration features in the resampled sequence are used as environmental observations to input into the evaluation model. The evaluation strategy is optimized based on the geodesic distance of the alignment feature vector relative to the standard state manifold and combined with the environmental observations, and the health level result is output.

2. The method for assessing the health of assembly tools based on tightening data according to claim 1, characterized in that, The process of acquiring multi-source data during the fastening of heterogeneous materials includes: simultaneously acquiring the torque sequence and angle sequence during the fastening process using torque and angular displacement sensors integrated within the assembly tool; capturing the drive motor current sequence characterizing the dynamic load state of the motor using a current sensor in the electronic control unit; acquiring the ultrasonic stress pulse signal through the pulse reflection effect using a piezoelectric transducer; capturing the structural response vibration characteristics characterizing the material interface interaction using an acoustic emission sensor attached to the fastening area; and connecting to an edge acquisition terminal to perform synchronous alignment and construct multi-source data based on a unified time reference.

3. The method for assessing the health of assembly tools based on tightening data according to claim 1, characterized in that, The process of identifying process mode switching points and outputting sampling adjustment commands includes: extracting the first derivative features of the torque sequence, the angular acceleration features of the angle sequence, and the short-time energy entropy features of the structural response vibration features; constructing a three-dimensional state vector representing the fastening state through normalized alignment and splicing; inputting the three-dimensional state vector into the reinforcement learning sampling model; identifying the elastic contact initiation point, interlayer flattening point, and material yield switching point during the contact process between the fastener and the heterogeneous material through nonlinear mapping, as the process mode switching points; calculating the feature information abundance requirement for the current stage based on the identified process mode switching points; outputting a sampling adjustment command containing the sampling period scaling factor; calculating the signal reconstruction gain after resampling in real time and feeding it back to the reinforcement learning sampling model as a reward signal; adjusting the parameters of the policy network in the reinforcement learning sampling model through a near-end policy optimization algorithm; and correcting the timing of the sampling adjustment command output.

4. The method for assessing the health of assembly tools based on tightening data according to claim 3, characterized in that, The reinforcement learning model includes: a feature-aware network: using a one-dimensional convolutional neural network to receive the three-dimensional state vector and extract local interaction features of multi-source data streams in the time domain; a temporal dependency network: using gated recurrent units to perform sequence modeling on the local interaction features, generating a long-range dependency feature vector containing the evolution trend of the fastening process; an Actor execution branch: acting as the execution entity of the policy network, using fully connected layers to perform nonlinear mapping on the long-range dependency feature vector, outputting the sampling period scaling factor conforming to a Gaussian distribution, and generating sampling adjustment instructions; a Critic evaluation branch: sharing the feature-aware network and the temporal dependency network with the Actor execution branch, used to calculate the state value estimate corresponding to the current fastening state; and a policy constraint optimizer: using a proximal policy optimization algorithm to calculate the probability ratio between the current policy and the historical policy, and forcibly constraining the update step size of the neural network according to the pruning probability ratio function.

5. The method for assessing the health of assembly tools based on tightening data according to claim 1, characterized in that, The process of obtaining the resampling sequence includes: determining the target sampling step size for each channel according to the scaling factor in the sampling adjustment instruction; performing linear interpolation on the original data stream according to the target sampling step size to fill in the missing data points of the low-frequency sampling channels; and mapping the aligned multi-source data to a unified discrete time axis through a timestamp matching algorithm to construct a multi-dimensional synchronization feature matrix, which serves as the resampling sequence.

6. The method for assessing the health of assembly tools based on tightening data according to claim 1, characterized in that, The process of obtaining the aligned feature vector includes: using the angular components in the resampled sequence as a reference, performing equal-angle resampling on the drive motor current sequence to map the time-domain current signal to the angular domain space; performing a fast Fourier transform on the angular domain current signal to obtain the order spectrum, extracting the characteristic order energy components that are integer multiples of the motor rotation frequency, and constructing the energy feature; estimating the marginal probability density distributions of the energy feature and the structural response vibration feature respectively, and constructing the joint probability density distribution between the two; calculating the mutual information between the energy feature and the structural response vibration feature based on the marginal probability density distribution and the joint probability density distribution, and using the normalized weighting factor obtained based on the mutual information calculation to remove the interference information related to the structural response vibration feature from the energy feature, thereby obtaining the residual feature vector characterizing the degradation state of the tool transmission chain.

7. The method for assessing the health of assembly tools based on tightening data according to claim 1, characterized in that, The process of obtaining the alignment feature vector includes: synchronously fusing the torque component, angle component, ultrasonic stress pulse signal, and residual feature vector; constructing a symmetric positive definite matrix representing the tightness state space by calculating the covariance matrix, which serves as the healthy state manifold; performing a logarithmic-Euclidean transformation on the symmetric positive definite matrix using a logarithmic mapping function, projecting it from a nonlinear Riemannian space to an Euclidean tangent space to obtain a high-dimensional feature tangent vector; retrieving the mean of the reference manifold under the pre-stored tool standard healthy state and mapping it to the same tangent space to obtain the reference center vector; calculating the relative offset of the high-dimensional feature tangent vector relative to the reference center vector, and eliminating the global distribution offset caused by batch differences in heterogeneous materials and environmental fluctuations through a centering translation operation to obtain the alignment feature vector.

8. The method for assessing the health of assembly tools based on tightening data according to claim 1, characterized in that, The process of outputting the health level result includes: inputting the alignment feature vector as the core evaluation variable into the reinforcement learning evaluation model, and simultaneously injecting the gradient components of the structural response vibration characteristics in the resampled sequence as environmental observations in real time to construct a dynamic observation space; calculating the geodesic distance of the alignment feature vector relative to the preset standard state manifold in Riemann space to obtain an initial deviation scalar characterizing the degree of deviation of the tool state from the baseline; the reinforcement learning evaluation model uses the geodesic distance and the environmental observations as input states, and uses a policy network to generate decision threshold correction coefficients for different process modes to perform evaluation strategy optimization; using the decision threshold correction coefficients to perform error compensation calculations on the geodesic distance, removing the unsteady components affected by the gaps in the heterogeneous material stacks, and extracting the corrected feature deviation reflecting the degree of performance degradation of the tool body; mapping the corrected feature deviation to a preset health evaluation index system, and outputting the health level result characterizing the tool's operating accuracy and mechanical fatigue state.

Images