Intelligent detection and automatic reset control method for locking defects based on displacement sensing

Abstract

This invention relates to the field of workpiece fastening technology, and particularly to an intelligent detection and automatic reset control method for fastening defects based on displacement sensing. The method includes: analyzing fastening depth signals at multiple scales to extract abrupt and steady-state features, providing a comprehensive and crucial data foundation for defect identification; generating feature vectors reflecting the fastening process state by identifying fastening stages and rationally fusing features, thus improving the reliability of defect judgment; identifying multiple fastening defect types by employing models and attention mechanisms, thereby improving the accuracy and comprehensiveness of defect detection; controlling the electric screwdriver's actions by generating a reverse compensation strategy based on steady-state offset features, ensuring fastening quality; determining a reasonable screw removal mode by comprehensively analyzing abrupt energy characteristics and monitoring and judging screw loosening in real time, achieving efficient handling and refeeding of stripped thread defects; and verifying reset results in real time and triggering intervention by accumulating failure counts, ensuring a stable and reliable fastening process.

Description

Technical Field

[0001] This invention relates to the field of workpiece fastening technology, and in particular to a method for intelligent detection and automatic reset control of fastening defects based on displacement sensing. Background Technology

[0002] Automated screw fastening equipment is widely used in product assembly in fields such as precision electronic assembly, automotive manufacturing, and aerospace. The quality of screw fastening directly affects the structural reliability and electrical performance of the product; therefore, real-time detection and correction of common fastening defects are crucial.

[0003] However, existing methods have fundamental limitations: the detection logic largely relies on static threshold comparisons of final depth or torque, failing to perform multi-scale analysis and coupling analysis of dynamic temporal signals during the fastening process. This results in insufficient sensitivity for identifying early minor defects and complex defects, leading to a high false positive rate. Existing automatic reset strategies lack differentiation and specificity for defect types, typically employing fixed reverse rotation angles or simple repeated fastening, which cannot adapt to the physical causes and repair mechanisms of different defects such as floating, sinking, and stripping, resulting in low reset success rates and potentially exacerbating damage. Therefore, this invention proposes an intelligent detection and automatic reset control method for fastening defects based on displacement sensing. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the background art, and to propose a method for intelligent detection and automatic reset control of locking defects based on displacement sensing.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent detection and automatic reset control of locking defects based on displacement sensing includes: S1. The locking depth time-series signal during the locking process is acquired in real time by a displacement sensor, and the torque-angle curve of the electric screwdriver is acquired simultaneously. The locking depth time-series signal is decomposed into multiple scales to extract its abrupt energy characteristics and steady-state offset characteristics. S2. Identify the locking stage based on the torque-angle curve of the electric screwdriver, and fuse the sudden energy characteristics and steady-state offset characteristics according to the locking stage to generate a fused defect feature vector. S3. Input the fused defect feature vector into the pre-trained locking defect classification model to identify at least one type of locking defect among floating, sinking, slipping, and missing lock. S4. If a floating or sinking defect is identified, a reverse compensation strategy is generated based on the steady-state offset characteristics to control the electric screwdriver to perform reverse rotation and re-forward locking actions. S5. If a stripped tooth defect is identified, a screw removal and refeeding strategy is generated based on the mutation energy characteristics. The electric screwdriver is controlled to remove the screw and trigger the feeder to refeed the screw. S6. During the automatic reset process, the locking depth timing signal and the electric screwdriver torque-angle curve are verified in real time. If they meet the preset compliance conditions, the reset is determined to be successful. Otherwise, the number of failures is accumulated and the manual intervention mechanism is triggered.

[0006] Further, S1 includes: A dual-tree complex wavelet transform is performed on the locked depth time series signal to decompose the locked depth time series signal into multiple sub-band signals of different scales, wherein the sub-band signal includes at least one high-frequency sub-band signal and one low-frequency sub-band signal; Calculate the integral of the squared modulus of the wavelet coefficients of each high-frequency sub-band signal in the time domain, and use the integral value as the abrupt energy characteristic. Empirical mode decomposition is performed on the low-frequency subband signal, and the first intrinsic mode function is obtained by selection. Calculate the mean of the first intrinsic mode function over the entire time window and the final value at the end of the time window corresponding to the locking depth timing signal. Subtract the mean and the final value from the theoretical locking depth to obtain the corresponding deviations, namely the mean deviation and the final value deviation. The two deviations are combined by weighted summation, and the combined result is used as the steady-state offset feature.

[0007] Further, S2 includes: Calculate the first and second derivative sequences of the torque-angle curve; From the first derivative sequence, the point where its value first exceeds the preset first slope threshold is identified as the starting point of the locking process from the unloaded to the pre-tightening stage; the point corresponding to the global maximum value in the first derivative sequence is identified as the starting point of the locking process from the pre-tightening stage to the main locking stage. From the second derivative sequence, find the zero-crossing point where its value changes from positive to negative, and after the zero-crossing point, identify the point where the value of the first derivative sequence first falls below the preset second slope threshold as the starting point of the locking process from the main locking stage to the final tightening stage. The hysteresis loop area sequence of torque as a function of angle is calculated synchronously. If the increment of the hysteresis loop area sequence in the adjacent calculation window exceeds the preset area mutation threshold, the identification position of the starting point of the main lock stage is corrected forward by a preset time compensation amount.

[0008] Furthermore, S2 also includes: Establish a three-row, two-column weight allocation matrix. The three rows of the matrix correspond to the pre-tightening stage, the master locking stage, and the final tightening stage, respectively, and the two columns correspond to the mutation energy characteristics and the steady-state offset characteristics, respectively. The matrix elements are assigned values ​​according to the preset rules. The weight of the mutation energy feature is higher in the main locking stage than in the pre-tightening stage and the final tightening stage. The weight of the steady-state offset feature is higher in the final tightening stage than in the pre-tightening stage and the main locking stage. The mutation energy feature and steady-state shift feature corresponding to each stage are multiplied by the corresponding weights in the matrix and summed to obtain the fusion intermediate value of that locking stage. The fusion intermediate values ​​of the pre-tightening stage, the main locking stage, and the final tightening stage, as well as the proportion of the duration of each locking stage to the total locking time, are sequentially concatenated into a feature array. The feature array is subjected to max-min normalization, and the normalized feature array is output as the fusion defect feature vector.

[0009] Further, S3 includes: The locking defect classification model consists of a sequentially connected one-dimensional convolutional layer, pooling layer, gated recurrent unit layer, and fully connected classification layer; the fused defect feature vector is reshaped into a two-dimensional tensor with a time step, input into the one-dimensional convolutional layer for convolution operation, and dynamic pattern features of the multi-scale locking process are extracted. The dynamic pattern features of the multi-scale locking process are input into the gated recurrent unit layer to extract the evolution correlation features of its locking stage. Before the fully connected classification layer, an attention mechanism layer is introduced to assign dynamic weights to different time steps of the lock-attachment stage evolution associated features, generating weighted lock-attachment stage attention features. The weighted attention features of the locking stage are input into the fully connected classification layer, and the output is the probability distribution representing four locking defect types: floating, sinking, slipping, and missing lock.

[0010] Further, S4 includes: Based on the sign of the steady-state offset feature, determine whether the type of locking defect is floating or sinking; if it is identified as a floating defect, the value of the steady-state offset feature is positive; if it is identified as a sinking defect, the value of the steady-state offset feature is negative. Based on the absolute value of the steady-state offset characteristic, query the preset deviation-compensation angle mapping table to obtain the basic reverse compensation angle. The reverse compensation strategy consists of the superposition of feedforward control and feedback control: the feedforward control is determined by the basic reverse compensation angle; the feedback control angle is obtained by performing proportional-integral calculation on the difference between the real-time acquired locking depth timing signal and the preset target locking depth value during the compensation process, and is used as the feedback control. The total compensation angle command is determined based on the sum of the feedforward control quantity and the feedback control quantity. The electric screwdriver is controlled to rotate in the reverse direction according to the total compensation angle command, and then performs a re-forward locking action according to the preset forward locking parameters.

[0011] Further, S5 includes: The phase space of the numerical sequence of mutation energy characteristics during the master-locking phase is reconstructed, and its maximum Lyapunov exponent is calculated. Simultaneously, the centroid frequency of the power spectrum of this numerical sequence is calculated; The maximum Lyapunov exponent and the power spectrum centroid frequency are input into a preset linear discriminant function; the coefficients of the linear discriminant function are obtained by logistic regression training on historical slipped tooth sample data; the linear discriminant function outputs the corresponding screw removal mode command, which includes high-speed pulse screw removal mode, medium-speed oscillation screw removal mode and low-speed constant torque screw removal mode. The electric screwdriver is controlled to perform a reverse rotation action according to the screw ejection mode command. During the screw ejection process, a fast Fourier transform is performed on the real-time monitored output torque signal of the electric screwdriver to calculate the proportion of signal energy in the characteristic frequency band related to the screw thread meshing vibration to the total energy. When this proportion is continuously lower than the preset energy proportion threshold, it is determined that the screw has come loose, the screw ejection is stopped, and the feeder is triggered to perform screw feeding and air blowing cleaning operations.

[0012] Further, S6 includes: After the reverse rotation and re-forward locking action or the pin ejection and refeeding operation is completed, the verification time window is activated. Within this verification time window, the locking depth timing signal and the electric screwdriver torque-angle curve are simultaneously acquired; after applying a moving average filter to the locking depth timing signal, the steady-state data is extracted and calculated to obtain the reset verification depth value; For the torque-angle curve of the electric screwdriver, extract the torque data segment after entering the final tightening stage, and calculate the dynamic time warping distance between this data segment and its ideal torque profile curve, which serves as the torque curve matching index.

[0013] Furthermore, S6 also includes: Set the dynamic acceptable depth range and the consistency threshold; Compare the reset verification depth value with the qualified depth range; compare the torque curve matching index with the preset matching threshold. The reset operation is considered successful if and only if the reset verification depth value falls within the dynamic qualified depth range and the torque curve matching index is less than the matching threshold.

[0014] Furthermore, S6 also includes: If the reset operation is deemed a failure, the type of locking defect identified in this operation is obtained. Based on the probability distribution output by the locking defect classification model, a confidence-weighted entropy value is calculated to characterize the severity of the defect. The cumulative number of failures at the current locking point is increased by an increment positively correlated with the confidence-weighted entropy value. It is then determined whether the cumulative number of failures exceeds the set repair threshold. If it exceeds the set repair threshold, the intervention process is executed.

[0015] Compared with existing technologies, the beneficial effects of this invention in providing a method for intelligent detection and automatic reset control of locking defects based on displacement sensing are as follows: 1) By using a multi-scale decomposition method, the sudden energy features and steady-state offset features of the locking depth time series signal are fully extracted to reflect the abnormal impacts during the locking process and the overall offset situation, laying the foundation for accurate identification of defect types in the future. By using the derivative information of the torque-angle curve to identify the locking stage, and then assigning weights to fuse features according to preset rules, the importance differences of features in different stages are considered, and the stage duration information is combined, so that the fused defect feature vector can more comprehensively and accurately represent the state of the locking process. 2) By employing a locking defect classification model that includes one-dimensional convolutional layers and gated recurrent unit layers, combined with an attention mechanism, it can effectively extract multi-scale dynamic pattern features and locking stage evolution correlation features, and dynamically allocate weights to improve the recognition accuracy of floating, sinking, slipping, and missing locking defect types; by generating a reverse compensation strategy based on steady-state offset features, and superimposing feedforward and feedback control quantities to determine the total compensation angle command, it can accurately control the electric screwdriver to perform reverse rotation and re-forward locking actions, effectively correcting floating or sinking defects and improving locking quality; 3) By performing phase space reconstruction analysis on the characteristics of mutation energy and combining it with a linear discriminant function to determine the screw removal mode, and by monitoring the energy ratio of the characteristic frequency band of the torque signal to determine the screw loosening situation, reasonable screw removal and refeeding of stripped teeth defects are achieved to ensure production continuity. Real-time verification is performed during the automatic reset process. The success of the reset is judged by setting a dynamic qualified range and a matching degree threshold. The number of failures is accumulated and a manual intervention mechanism is triggered to ensure stable and reliable fastening quality and prevent defective products from flowing into the next process. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent detection and automatic reset control method for locking defects based on displacement sensing proposed in this invention. Detailed Implementation

[0017] 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.

[0018] Reference Figure 1 A method for intelligent detection and automatic reset control of locking defects based on displacement sensing includes: S1. The locking depth time-series signal during the locking process is acquired in real time through a displacement sensor, and the electric screwdriver torque-angle curve is acquired simultaneously. The locking depth time-series signal is decomposed into multiple scales to extract its abrupt energy characteristics and steady-state offset characteristics. The process of synchronously acquiring the electric screwdriver torque-angle curve includes: the electric screwdriver controller acquires the output signals of its built-in torque sensor and rotary encoder in real time; the torque sensor samples and outputs the instantaneous torque value at a frequency of not less than 1kHz, and the rotary encoder synchronously records the rotation angle of the electric screwdriver spindle; the electric screwdriver controller pairs the torque value acquired at each moment with the corresponding angle value, and uploads it to the upper host control system in real time through RS-485 or CAN bus communication in a preset message format and period; the upper host system parses and buffers the data according to the received time sequence, and adds a timestamp synchronized with the displacement sensor data, finally forming a time-aligned electric screwdriver torque-angle curve. S2. Identify the locking stage based on the torque-angle curve of the electric screwdriver, and fuse the sudden energy characteristics and steady-state offset characteristics according to the locking stage to generate a fused defect feature vector. S3. Input the fused defect feature vector into the pre-trained locking defect classification model to identify at least one type of locking defect among floating, sinking, slipping, and missing lock. S4. If a floating or sinking defect is identified, a reverse compensation strategy is generated based on the steady-state offset characteristics to control the electric screwdriver to perform reverse rotation and re-forward locking actions. S5. If a stripped tooth defect is identified, a screw removal and refeeding strategy is generated based on the mutation energy characteristics. The electric screwdriver is controlled to remove the screw and trigger the feeder to refeed the screw. S6. During the automatic reset process, the locking depth timing signal and the electric screwdriver torque-angle curve are verified in real time. If they meet the preset compliance conditions, the reset is determined to be successful. Otherwise, the number of failures is accumulated and the manual intervention mechanism is triggered.

[0019] It should be further explained that, in the specific implementation process, the locking depth time-series signal during the locking process is acquired in real time through a displacement sensor, and the torque-angle curve of the electric screwdriver is acquired simultaneously. The locking depth time-series signal is then decomposed into multiple scales to extract its abrupt energy characteristics and steady-state offset characteristics, including: A dual-tree complex wavelet transform is performed on the locked depth time series signal to decompose the locked depth time series signal into multiple sub-band signals of different scales, wherein the sub-band signal includes at least one high-frequency sub-band signal and one low-frequency sub-band signal; Calculate the integral of the squared modulus of the wavelet coefficients of each high-frequency sub-band signal in the time domain, and use the integral value as the abrupt energy characteristic. Empirical mode decomposition is performed on the low-frequency subband signal, and the first intrinsic mode function is obtained by selection. The mean value of the first intrinsic mode function over the entire time window and the final value at the end of the corresponding time window of the locking depth timing signal are calculated. The mean value and the final value are then subtracted from the theoretical locking depth to obtain the corresponding deviations, namely the mean deviation and the final value deviation. The theoretical locking depth refers to the expected height of the screw head relative to the workpiece surface after the screw is screwed into the workpiece, which is calculated in advance based on the screw specifications, thread parameters and process requirements. The two deviations are combined by weighted summation, and the combined result is used as the steady-state offset feature. Specifically, the locking depth time-series signal is acquired in real time using a displacement sensor; the acquired locking depth time-series signal is preprocessed, including subtracting the initial offset and performing a 5-point moving average filter to suppress high-frequency noise; a dual-tree complex wavelet transform is performed on the preprocessed locking depth time-series signal: a Q-shift 14 filter bank is selected for 5-level decomposition; after decomposition, each level generates a low-frequency sub-band signal and a high-frequency sub-band signal, namely an approximation coefficient sub-band signal and a detail coefficient sub-band signal; the high-frequency sub-band signals from the 3rd to the 5th levels are selected for analysis, corresponding to different time-frequency resolutions, which can capture various abnormal patterns from rapid impact to slow drift; For each selected high-frequency subband signal, the time-domain integral of the squared modulus of its wavelet coefficients is calculated as a feature of abrupt change energy. The squared modulus is the sum of the squared real part and the squared imaginary part of the wavelet coefficients. The integration is performed over the entire time window corresponding to the locking process, and is approximately the sum of the squares of the discrete coefficient sequence multiplied by the sampling interval. The larger the integral value, the more severe the instantaneous abrupt change or vibration in the locking process at that scale, which may be related to transient events such as slippage or collision. For low-frequency subband signals, empirical mode decomposition (EMD) is performed to further separate trends and fluctuations. EMD, through an iterative sieving process, decomposes the low-frequency subband signal into a series of intrinsic mode functions (EMFs) and a residual trend term. The purpose of this decomposition is to extract the low-frequency fluctuation components reflecting the final steady-state characteristics of the latch-up signal. The first EMF contains relatively high-frequency fluctuation information from the low-frequency subband signal, while the residual trend term represents the ultra-low frequency drift or long-term trend of the low-frequency subband signal. After removing the residual trend term, the first EMF is selected. The intrinsic mode function is calculated; the mean of the intrinsic mode function is calculated over the entire time window; the final value of the function at the last 10 sampling points at the end of the window is also calculated, i.e., the average value; these two values ​​are subtracted from the theoretical locking depth to obtain the mean deviation and the final deviation; the mean deviation and the final deviation are weighted and summed to obtain the steady-state offset feature: steady-state offset feature = 0.7 × final deviation + 0.3 × mean deviation. This weighting method emphasizes the degree of deviation of the final locking result, and this feature directly reflects the overall trend of buoyancy or sinking.

[0020] It should be further explained that, in the specific implementation process, the locking stage is identified based on the torque-angle curve of the electric screwdriver, and the sudden energy characteristics and steady-state offset characteristics are fused according to the locking stage to generate a fused defect feature vector, including: Calculate the first and second derivative sequences of the torque-angle curve; From the first derivative sequence, the point where its value first exceeds the preset first slope threshold is identified as the starting point of the locking process from the no-load to the pre-tightening stage. The first slope threshold is an empirical value determined based on the standard deviation of torque fluctuation during the no-load operation of the electric screwdriver, used to characterize the start of significant torque growth. The point corresponding to the global maximum value in the first derivative sequence is identified as the starting point of the locking process from the pre-tightening stage to the main locking stage. From the second derivative sequence, find the zero-crossing point where its value changes from positive to negative, and after the zero-crossing point, identify the point where the value of the first derivative sequence first falls below the preset second slope threshold as the starting point of the locking process from the main locking stage to the final tightening stage. The second slope threshold is 1 / 5 to 1 / 2 of the first slope threshold, which is used to characterize the slowdown of torque growth. The hysteresis loop area sequence of torque as a function of angle is calculated synchronously. If the increment of the hysteresis loop area sequence in an adjacent calculation window exceeds a preset area mutation threshold, the identification position of the main locking stage start point is corrected forward by a preset time compensation amount. The area mutation threshold is determined based on the average fluctuation range of the hysteresis loop area in the preceding continuous normal locking process and is used to identify frictional force state mutations. The time compensation amount is an integer multiple of the duration of a single data acquisition cycle and is used to compensate for system response lag. Establish a three-row, two-column weight allocation matrix. The three rows of the matrix correspond to the pre-tightening stage, the master locking stage, and the final tightening stage, respectively, and the two columns correspond to the mutation energy characteristics and the steady-state offset characteristics, respectively. The matrix elements are assigned values ​​according to the preset rules. The weight of the mutation energy feature is higher in the main locking stage than in the pre-tightening stage and the final tightening stage. The weight of the steady-state offset feature is higher in the final tightening stage than in the pre-tightening stage and the main locking stage. The mutation energy feature and steady-state shift feature corresponding to each stage are multiplied by the corresponding weights in the matrix and summed to obtain the fusion intermediate value of that locking stage. The fusion intermediate values ​​of the pre-tightening stage, the main locking stage, and the final tightening stage, as well as the proportion of the duration of each locking stage to the total locking time, are sequentially concatenated into a feature array. The feature array is subjected to max-min normalization, and the normalized feature array is output as the fusion defect feature vector. Specifically, identifying the start point of the preload stage based on the first derivative sequence includes: calculating the standard deviation σF of the torque fluctuation during the no-load operation phase of the electric screwdriver, where the no-load operation phase can be the first 50ms before rotation begins; setting the first slope threshold to 3σF; scanning backward from the start point of the first derivative sequence, marking the point where the value of the first first derivative sequence exceeds the first slope threshold as the start point of the preload stage, which indicates that the torque begins to increase from the background noise and the screw head begins to contact the workpiece; The starting point of the main locking stage is identified by finding the global maximum point of the first derivative sequence. This point usually corresponds to the moment when the torque growth rate is the fastest, indicating that the screw thread has been fully engaged and the main locking stage of linear tightening has begun. The method of identifying the starting point of the final tightening stage by combining the first derivative sequence and the second derivative sequence includes: finding the first zero-crossing point in the second derivative sequence where the value changes from positive to negative, indicating that the torque growth acceleration begins to decrease; after this zero-crossing point, continuing to look for the point in the first derivative sequence where the value first falls below the second slope threshold; the second slope threshold is set to 1 / 3 of the first slope threshold, and the second slope threshold can be adjusted between 1 / 5 and 1 / 2 to characterize the torque growth slowing down significantly and entering the final tightening stage of plastic deformation or bonding; The hysteresis loop area sequence is obtained by calculating and accumulating the small trapezoidal area enclosed by the torque-angle curves in two adjacent sampling periods; the average hysteresis loop area A and its standard deviation σA near the main locking stage of the first 10 consecutive normal locking processes are recorded; the area mutation threshold is set to 2σA; in the current locking process, the difference between the hysteresis loop area sequence values ​​of adjacent time windows near the starting point of the main locking stage is calculated, i.e., the hysteresis loop area increment; if the hysteresis loop area increment exceeds the area mutation threshold, it indicates that the friction force state has changed abruptly. If a foreign object is encountered, the identification position of the starting point of the main locking stage is corrected forward by a time compensation amount. This time compensation amount is set to 2 data acquisition periods, such as 2ms, to compensate for the hysteresis of the sensor or system response, so that the stage division is more in line with the physical reality; A weighting matrix is ​​constructed, which is a 3x2 matrix. The three rows correspond to pre-tightening, main locking, and final tightening, respectively, and the two columns correspond to sudden energy characteristics and steady-state offset characteristics, respectively. Weights are assigned according to pre-defined rules based on process knowledge: during the main locking stage, the screw experiences complex stress, and the sudden energy characteristic has the highest weight, set to 0.6; during the pre-tightening and final tightening stages, the weight is set to 0.2; the steady-state offset characteristic best reflects the final quality during the final tightening stage, so its weight is set to 0.7; during the main locking stage, it is set to 0.2; and during the pre-tightening stage, it is set to 0.1. An example of the weighting matrix is ​​shown below. Pre-tightening phase: ; Master lock phase line: ; Final tightening phase: ; In the formula, The weights for the steady-state offset characteristics during the pre-tightening stage. The weights for the steady-state offset characteristics during the pre-tightening stage. The weights of the steady-state offset characteristics during the master locking phase. The weights of the steady-state offset characteristics during the master locking phase. The weights for the steady-state offset characteristics in the final compaction stage. The weights for the steady-state offset characteristics in the final compaction stage; For a specific locking process, extract the mutation energy eigenvalues ​​and steady-state offset eigenvalues ​​corresponding to each of the three locking stages; let Fe and Fo be the mutation energy eigenvalues ​​and steady-state offset eigenvalues ​​of any locking stage, respectively, then the fusion intermediate value of that locking stage = W[row,1]×Fe+W[row,2]×Fo; where W[row,1] represents the weight of the mutation energy eigenvalue in the first column of the row of the weight allocation matrix, and W[row,2] represents the weight of the steady-state offset eigenvalue in the second column of the row of the weight allocation matrix; thus, three fusion intermediate values ​​are calculated, which are denoted as M1 for the pre-tightening stage, M2 for the main locking stage, and M3 for the final tightening stage; The duration of each locking stage is obtained from the locking stage division results, namely the pre-tightening stage duration ST1, the main locking stage duration ST2, and the final tightening stage duration ST3, and the total duration To is calculated: To = ST1 + ST2 + ST3; the proportion of each stage duration to the total duration is calculated: R1 = ST1 / To, R2 = ST2 / To, R3 = ST3 / To, where R1 is the proportion of the pre-tightening stage, R2 is the proportion of the main locking stage, and R3 is the proportion of the final tightening stage; the proportions of the three locking stage durations are used as the temporal structure features; finally, the six values ​​are concatenated in order to form the initial feature array: [M1, M2, M3, R1, R2, R3]; The initial feature array is subjected to min-max normalization, and the normalized 6-dimensional feature array is output, which is the final fusion defect feature vector. The fusion defect feature vector encodes the spatial intensity information of multi-scale features and the temporal structure information of the locking process.

[0021] It should be further explained that, in the specific implementation process, the fused defect feature vector is input into a pre-trained locking defect classification model to identify at least one locking defect type among floating, sinking, slipping, and missing lock, including: The locking defect classification model consists of a sequentially connected one-dimensional convolutional layer, pooling layer, gated recurrent unit layer, and fully connected classification layer; the fused defect feature vector is reshaped into a two-dimensional tensor with a time step, input into the one-dimensional convolutional layer for convolution operation, and dynamic pattern features of the multi-scale locking process are extracted. The dynamic pattern features of the multi-scale locking process are input into the gated recurrent unit layer to extract the evolution correlation features of its locking stage. Before the fully connected classification layer, an attention mechanism layer is introduced to assign dynamic weights to different time steps of the lock-attachment stage evolution associated features, generating weighted lock-attachment stage attention features. The weighted attention features of the locking stage are input into the fully connected classification layer, and the output is the probability distribution representing four types of locking defects: floating, sinking, slipping, and missing lock. Specifically, the locking defect classification model used is a hybrid neural network; the 6-dimensional fusion defect feature vector is reshaped into a 6×1 two-dimensional tensor, which can be regarded as a sequence with 6 time steps, where each time step contains 1 feature scalar, which is the value of the fusion defect feature vector at that index position; for example, the first time step corresponds to the fusion intermediate value M1 in the pre-tightening stage, the second time step corresponds to the fusion intermediate value M2 in the main locking stage, the third time step corresponds to the fusion intermediate value M2 in the final tightening stage, the fourth time step corresponds to the pre-tightening stage ratio R1, and so on. The tensor is input to a one-dimensional convolutional layer, which uses 32 convolutional kernels with a width of 3 and a stride of 1. Edge padding is performed in the same manner to keep the sequence length unchanged. The convolution operation is to slide each convolutional kernel along 6 time steps, perform multiplication and addition operations, and extract the local dependencies between features of adjacent time steps (i.e., M1 and M2, M2 and M3, etc.). The one-dimensional convolutional layer outputs a 6×32 feature map, which is the dynamic pattern feature of the multi-scale locking process. The feature map is processed through a max pooling layer with a pooling size of 2 and a stride of 2; the dimension reduction results in a 3×32 feature map, which improves computational efficiency and introduces a certain degree of translation invariance while preserving the main information. The feature map is input into a gated recurrent unit layer containing 16 hidden units; the gated recurrent unit layer processes the input of 3 time steps in sequence, and through its update gate and reset gate mechanism, it remembers or forgets the information of the previous steps, thereby learning the evolutionary correlation features of the locking and attachment process from pre-tightening to master lock and then to final tightening, where the evolutionary correlation features of the locking and attachment stage are 16-dimensional vectors. An attention mechanism layer is introduced, which takes the lock attachment stage evolution correlation feature vector output by the gated recurrent unit layer as input, and calculates a 3-dimensional attention weight vector through a trainable fully connected layer. The three attention weight vectors correspond to the three time steps after the original pooling, namely the information representation of the three lock attachment stages: pre-tightening, main locking, and final tightening. The calculation process is as follows: the output vector of the gated recurrent unit layer is linearly transformed and then followed by a Softmax function to ensure that the sum of the three attention weight vectors is 1; the attention weight vectors are combined with the hidden states (3×16 matrices) generated by the gated recurrent unit layer at each time step to generate a weighted attention feature vector for the lock attachment stage, so that the lock attachment defect classification model can focus attention on the most critical lock attachment stage for current defect identification. The attention feature vector from the locking stage is fed into a fully connected classification layer, which is a Softmax layer with four output nodes, corresponding to four defect types. In the fully connected classification layer, a linear transformation is performed on the attention feature vector from the locking stage. This is done by multiplying the attention feature vector from the locking stage with a 16×4 weight matrix and adding a bias vector of dimension 4 to complete the linear combination operation. The result of the linear combination is normalized by the Softmax function, outputting a 4-dimensional probability distribution vector, namely [P1, P2, P3, P4], which corresponds to the locking defect types of floating, sinking, slipping, and missing lock, respectively. The category with the highest probability is identified as the current locking defect type.

[0022] It should be further explained that, in the specific implementation process, if a floating or sinking defect is identified, a reverse compensation strategy is generated based on the steady-state offset characteristics to control the electric screwdriver to perform reverse rotation and re-forward locking actions, including: Based on the sign of the steady-state offset feature, determine whether the type of locking defect is floating or sinking; if it is identified as a floating defect, the value of the steady-state offset feature is positive; if it is identified as a sinking defect, the value of the steady-state offset feature is negative. Based on the absolute value of the steady-state offset characteristic, query the preset deviation-compensation angle mapping table to obtain the basic reverse compensation angle. The reverse compensation strategy consists of the superposition of feedforward control and feedback control: the feedforward control is determined by the basic reverse compensation angle; the feedback control angle is obtained by performing proportional-integral calculation on the difference between the real-time acquired locking depth timing signal and the preset target locking depth value during the compensation process, and is used as the feedback control. The total compensation angle command is determined based on the sum of the feedforward control quantity and the feedback control quantity. The electric screwdriver is controlled to rotate in the opposite direction according to the total compensation angle command, and then performs a re-forward locking action according to the preset forward locking parameters; wherein, the forward locking parameters include locking speed, target torque and final depth; Specifically, when the type of locking defect is identified as floating or sinking, a reverse compensation strategy is activated: the determination is based on the sign of the steady-state offset characteristic: if the steady-state offset characteristic value is positive, it means that the measured locking depth is greater than the theoretical locking depth, and it is judged as floating; if the steady-state offset characteristic value is negative, it is judged as sinking. Maintain an offset-compensation angle mapping table, which defines the correspondence between offset value ranges and angles. This table is obtained through experimental calibration during the debugging phase. Based on the absolute value of the identified steady-state offset characteristic value Fo, query the preset offset-compensation angle mapping table to obtain the basic reverse compensation angle θ1. The basic reverse compensation angle is a constant that is determined and kept constant before the compensation action is implemented. For example, when |Fo|∈[0.1,0.2), θ1=15°; when |Fo|∈[0.2,0.3), θ1=25°; and so on. The same applies to negative value ranges. Obtain the basic reverse compensation angle θ1 by looking up the table. For sinking defects (characteristic value is negative), their absolute values ​​are also applicable to this table, but the compensation direction is opposite to the buoyancy. The reverse compensation strategy employs a feedforward-feedback composite control. The feedforward control quantity is directly determined by the basic reverse compensation angle θ1. The feedback control quantity is calculated in real time by a proportional-integral controller. During the compensation execution process, the locking depth time-series signal d(t) is continuously acquired at a frequency of 1kHz, where t is a continuous time variable, and its instantaneous error e(t) with the theoretical locking depth DT is calculated: e(t) = d(t) - DT. The feedback control angle θ2(t) = Kc × [Kp × e(t) + Ki × ∫e(t)dt], where Kc is the angular displacement conversion coefficient, and Kp and Ki are the pre-tuned dimensionless proportional gain and integral gain, respectively, used to adjust the dynamic response of the controller. The integral term ∫e(t)dt is accumulated from the start of the compensation action. It can be understood that the controlled variable of the proportional-integral controller is the locking depth, and the control output is the additional compensation angle of the electric screwdriver. The total compensation angle is θ(t) = θ1 + θ2(t); a total compensation angle command is generated based on the total compensation angle; this command is sent to the electric screwdriver servo driver to control the electric screwdriver to perform reverse rotation; after the reverse rotation is completed, it is paused for 100ms to allow the internal stress to relax; a re-forward locking action is performed according to the preset forward locking parameters; in the torque-angle hybrid control mode, the electric screwdriver rotates until the target torque is reached and the depth falls within the tolerance range to ensure that the locking is firm and in place.

[0023] It should be further explained that, in the specific implementation process, if a stripped thread defect is identified, a screw retraction and refeeding strategy is generated based on the mutation energy characteristics. This strategy controls the electric screwdriver to retract the screw and triggers the feeder to refeed the screw, including: The phase space of the numerical sequence of mutation energy characteristics during the master-locking phase is reconstructed, and its maximum Lyapunov exponent is calculated. Simultaneously, the centroid frequency of the power spectrum of this numerical sequence is calculated; The maximum Lyapunov exponent and the power spectrum centroid frequency are input into a preset linear discriminant function; the coefficients of the linear discriminant function are obtained by logistic regression training on historical slipped tooth sample data; the linear discriminant function outputs the corresponding screw removal mode command, which includes high-speed pulse screw removal mode, medium-speed oscillation screw removal mode and low-speed constant torque screw removal mode. The electric screwdriver is controlled to perform a reverse rotation action according to the screw retraction mode command. During the screw retraction process, a fast Fourier transform is performed on the real-time monitored output torque signal of the electric screwdriver to calculate the proportion of signal energy in the characteristic frequency band related to the screw thread meshing vibration to the total energy. When this proportion is continuously lower than the preset energy proportion threshold, it is determined that the screw has come loose, the screw retraction is stopped, and the feeder is triggered to perform screw feeding and air blowing cleaning operations. The energy proportion threshold is a preset empirical value, which is the theoretical lower limit of the energy proportion in the characteristic frequency band when the screw thread and the workpiece screw hole maintain effective contact. Specifically, the numerical sequence En of mutation energy characteristics in the main locking phase is analyzed in depth, where n=1,2,…,N, n is the sampling point index in the main locking phase, and N is the number of sampling points in the main locking phase; The analysis of numerical sequences through phase space reconstruction includes: selecting a delay time τ based on the delay corresponding to the first zero crossing of the autocorrelation function of the numerical sequence; determining the embedding dimension m using the spurious nearest neighbor method, which gradually increases the embedding dimension to determine whether the spurious proximity phenomenon of adjacent points on the phase space trajectory disappears in a higher-dimensional space, thereby finding the minimum dimension that can fully unfold the dynamic trajectory; and reconstructing a series of phase space state points Z(v) based on the determined delay time τ and embedding dimension m: Z(v) = [E(v), E(v+τ), ..., E(v+(m-1)τ)]; where v represents the index of the reconstructed phase space state point, E(v) represents the value of the v-th sampling point in the numerical sequence En, and E(v+τ) represents the value of the point after the v-th sampling point in the numerical sequence En. The value of time step E(v+(m-1)τ) represents the value of the numerical sequence En after the v-th sampling point at time step (m-1)τ. In phase space reconstruction, the phase space state points constitute the evolution trajectory in the phase space. The expression is used to convert the one-dimensional numerical sequence En into trajectory points in the high-dimensional phase space to analyze its dynamic characteristics. The maximum Lyapunov exponent λ of the numerical sequence is calculated using the small data method: a reference point is selected in the phase space, its nearest neighbor is found, and the initial distance between the two is calculated. As time evolves, the orbital distance between these two points is tracked, and the average growth rate of their pair over time is calculated. The average growth rate is the maximum Lyapunov exponent. If the maximum Lyapunov exponent λ is positive, it indicates that the numerical sequence has chaotic characteristics, and the slippage process may be accompanied by complex nonlinear vibrations. Simultaneously, calculating the power spectrum centroid frequency Fc of the numerical sequence En includes: performing a fast Fourier transform on the numerical sequence En to obtain the power spectrum Q(f), and further calculating the power spectrum centroid frequency Fc = Σ(f×Q(f)) / Σ(Q(f)), where f is the frequency, Σ(f×Q(f)) is the frequency-weighted power sum, and Σ(Q(f)) is the total power; the power spectrum centroid frequency reflects the main frequency band of the anomalous energy. The maximum Lyapunov exponent λ and the power spectrum centroid frequency Fc are input into the linear discriminant function: D = w1 × λ + w2 × Fc + b; where the weights w1, w2 and the bias b are obtained by logistic regression training on historical slipped tooth sample data, where the historical slipped tooth sample data are labeled with their optimal screw removal mode; the function output value D is mapped to three preset screw removal intervals, corresponding to three screw removal modes: if D is greater than Th1, it corresponds to the high-speed pulse screw removal mode, which uses high speed and intermittent forward and reverse rotation, and is used for severe slipped teeth; if Th2 is less than D and D is less than or equal to Th1, it corresponds to the medium-speed oscillation screw removal mode; if D is less than or equal to Th2, it corresponds to the low-speed constant torque screw removal mode, which uses constant small torque to slowly remove the screw, and is used for mild slipped teeth; where Th1 and Th2 are classification thresholds determined by analyzing the distribution of the function output value D of the training samples, and Th1 > Th2; Controlling the electric screwdriver to perform reverse rotation in a specified mode includes: monitoring the output torque signal of the electric screwdriver at a sampling rate of 5kHz during the screw retraction process; performing a fast Fourier transform on the output torque signal of the electric screwdriver every 100ms time window to analyze the signal energy EB and total energy ET within the characteristic frequency band, where the characteristic frequency band is the structural resonance frequency generated by the engagement of the screw thread and the screw hole, for example, 100-500 Hz; calculating the proportion of signal energy in the characteristic frequency band to the total energy RE=EB / ET; setting an energy proportion threshold RE0, such as 0.05, which is obtained by analyzing the distribution of the proportion RE during a large number of normal screw retraction processes; when the proportion RE is lower than the energy proportion threshold RE0 for three consecutive time windows, it is determined that the screw thread has completely loosened and has no effective contact with the screw hole; the screw retraction action is then stopped, and the feeder is triggered to perform a screw feeding operation, while the air blowing solenoid valve is controlled to open for 50ms to clean the metal debris in the screw hole and prepare for re-locking.

[0024] It should be further clarified that, in the specific implementation process, if the fastening defect classification model identifies the current fastening process as a missed fastening defect, it indicates that the screw has not been effectively fastened or the fastening depth is severely insufficient. Missed fastening defects originate from fundamental faults such as screw feeding failure, the bit failing to pick up the screw, or severe product misalignment, rather than process deviations that can be corrected by adjusting fastening parameters online. Therefore, this type of defect is handled using a different branch than defects such as floating, sinking, and stripping. This locking operation did not meet the preset compliance conditions, the reset verification step was skipped, and the failure count for this point was directly accumulated; at the same time, the highest priority alarm was immediately triggered to ensure that for fundamental faults that cannot be automatically repaired, the process is immediately interrupted and operators are alerted to intervene, thereby avoiding invalid automatic reset attempts and possible secondary damage, and ensuring production safety and efficiency.

[0025] It should be further explained that, in the specific implementation process, the locking depth timing signal and the electric screwdriver torque-angle curve are verified in real time during the automatic reset process. If they meet the preset compliance conditions, the reset is considered successful; otherwise, the number of failures is accumulated and a manual intervention mechanism is triggered, including: After the reverse rotation and re-forward locking action or the pin ejection and refeeding operation is completed, the verification time window is activated. Within this verification time window, the locking depth timing signal and the electric screwdriver torque-angle curve are simultaneously acquired; after applying a moving average filter to the locking depth timing signal, the steady-state data is extracted and calculated to obtain the reset verification depth value; For the torque-angle curve of the electric screwdriver, extract the torque data segment after entering the final tightening stage, calculate the dynamic time warping distance between this data segment and its ideal torque profile curve, and use it as the torque curve matching index. Set a dynamic qualified depth range and a consistency threshold; wherein, the center value of the dynamic qualified depth range is the theoretical depth, and its upper and lower boundary values ​​are adaptively adjusted based on the variance of the actual depth data measured by the displacement sensor based on recent consecutive successful locking; the consistency threshold is determined based on the percentile of the dynamic time-normalized distance statistical distribution of historical qualified locking samples. Compare the reset verification depth value with the qualified depth range; compare the torque curve matching index with the preset matching threshold. The reset operation is considered successful if and only if the reset verification depth value falls within the dynamic qualified depth range and the torque curve matching index is less than the matching threshold. If the reset operation is determined to have failed, the type of locking defect identified in this operation is obtained. Based on the probability distribution output by the locking defect classification model, a confidence-weighted entropy value is calculated to characterize the severity of the defect. The cumulative number of failures at the current locking point is increased by an increment positively correlated with the confidence-weighted entropy value. It is then determined whether the cumulative number of failures exceeds the set repair threshold. If it exceeds the set repair threshold, the intervention process is executed. Specifically, a verification time window is set, the length of which is usually 1.5 times the normal duration of the locking action, such as 1.5 seconds, to ensure that the complete post-reset state can be captured; Based on the locking depth time-series signal acquired within the verification time window, a moving average filter is applied to the original locking depth time-series signal with a window width of 11 sampling points to smooth noise. The steady-state stage of the locking depth time-series signal is identified: the standard deviation of the locking depth time-series signal in the second half of the data within the entire time window is calculated. If it is lower than a preset stability threshold, such as 0.01mm, the data segment is considered to have entered a steady state. The second half of the steady-state data is extracted, and the arithmetic mean of the second half of the steady-state data is calculated as the reset verification depth value. Based on the torque-angle curve of the electric screwdriver collected within the verification time window, the torque data segment identified as entering the final tightening stage within the verification time window is extracted, i.e., the torque curve after reset; at the same time, the ideal torque profile curve stored in the process library for the locking point is called, where the ideal torque profile curve is a standard curve obtained by statistical averaging under calibrated qualified conditions. To quantify the difference between the reset torque curve and the ideal torque profile curve, the dynamic time warping distance is calculated. The steps are as follows: a) The two sequences of reset torque curve and ideal torque profile curve are Z-score normalized in terms of amplitude; b) An X×Y cost matrix is ​​constructed, where X is the length of the reset torque curve sequence and Y is the length of the ideal torque profile curve sequence; the matrix elements are the Euclidean distance between the reset torque curve and the ideal torque profile curve; c) A path from (1,1) to (X,Y) is found in the cost matrix such that the sum of the elements on the path is minimized. This minimum sum is the torque curve fit index. The smaller the torque curve fit index value, the more similar the shapes of the two curves are. The severity of the locking defect and the certainty of the locking defect classification model's judgment are quantified by calculating the confidence-weighted entropy value HW. Let the probability distribution vector output by the locking defect classification model be [P1, P2, P3, P4], corresponding to locking defect types such as floating, sinking, slipping, and missing lock. The Shannon entropy of this probability distribution vector is calculated. In the formula, Indicates the first The probability of a lock-on defect. The index represents the type of locking defect, with values ​​from 1 to 4 corresponding to the four types of locking defects: floating, sinking, slipping, and missing lock, respectively. A higher entropy value indicates that the locking defect classification model is less certain. The probability Pk of the main locking defect type k that has been identified is obtained. For example, in the probability distribution vector [P1, P2, P3, P4] output by the locking defect classification model, assuming that P3 = 0.85 is the maximum value, then the currently identified locking defect type is slipping, and its probability Pk = 0.85. The confidence-weighted entropy value HW = H × (1-Pk); the larger the factor (1-Pk), the lower the confidence of the locking defect classification model in the main judgment, which may indicate that the locking defect is more complex or more serious. The cumulative number of failures HN at the locking point is increased by an increment ΔNW0 that is positively correlated with the confidence-weighted entropy value HW; the increment ΔNW0 is set as 1 + α × HW, where α is an adjustment coefficient, such as 0.5; this means that a severe failure with low confidence will lead to a greater increase in the number of failures, accelerating the triggering of manual intervention; A preset repair threshold HN0 is set, for example, 3 failures; if the updated cumulative failure count HN is greater than or equal to the repair threshold HN0, then the intervention process is executed: 1) Set the status flag of this point to pause automatic on the software interface; automatically filter out all locking tasks containing this point from the current and subsequent production task queues to avoid continuous failures; 2) Automatically start the background optimization routine; collect the most recent cumulative number of failures and failure processes at this point, as well as all relevant data from previous successful processes, such as the locking depth timing signal, the torque-angle curve of the electric screwdriver, and process parameters; define the process parameter search space, including locking speed (range ±20%), depth compensation (range ±0.1mm), torque upper limit (range ±10%), etc.; use a Bayesian optimization algorithm, with the reset success rate as the optimization objective, to iteratively optimize within the process parameter search space; this algorithm utilizes a Gaussian process model, based on existing historical process data. The parameter combinations and corresponding results, i.e., success / failure labels, are used to fit the relationship between the process parameter search space and the success rate. An acquisition function is set, which is used to balance the exploration of unknown areas and the utilization of predicted high success rate areas to select the next set of process parameter combinations to be tested, i.e., the process parameter values ​​most likely to improve the success rate. After optimization for a preset number of generations, such as 50 generations, the best process parameter combination found is updated to the process formula library for that point, for engineers to review or for subsequent automatic calling, thereby attempting to fundamentally solve the repetitive defect problem at that point.

[0026] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. The focus of each embodiment is on its differences from other embodiments. In particular, the apparatus embodiments are described simply because they are fundamentally based on the method embodiments; relevant details can be found in the descriptions of the method embodiments.

[0027] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0028] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0029] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0030] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0031] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0032] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0033] In conclusion, the above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent detection and automatic reset control of locking defects based on displacement sensing, characterized in that, The method includes: S1. The locking depth time-series signal during the locking process is acquired in real time by a displacement sensor, and the torque-angle curve of the electric screwdriver is acquired simultaneously. The locking depth time-series signal is decomposed into multiple scales to extract its abrupt energy characteristics and steady-state offset characteristics. S2. Identify the locking stage based on the torque-angle curve of the electric screwdriver, and fuse the sudden energy characteristics and steady-state offset characteristics according to the locking stage to generate a fused defect feature vector. S3. Input the fused defect feature vector into the pre-trained locking defect classification model to identify at least one type of locking defect among floating, sinking, slipping, and missing lock. S4. If a floating or sinking defect is identified, a reverse compensation strategy is generated based on the steady-state offset characteristics to control the electric screwdriver to perform reverse rotation and re-forward locking actions. S5. If a stripped tooth defect is identified, a screw removal and refeeding strategy is generated based on the mutation energy characteristics. The electric screwdriver is controlled to remove the screw and trigger the feeder to refeed the screw. S6. During the automatic reset process, the locking depth timing signal and the electric screwdriver torque-angle curve are verified in real time. If they meet the preset compliance conditions, the reset is determined to be successful. Otherwise, the number of failures is accumulated and the manual intervention mechanism is triggered.

2. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 1, characterized in that, S1 includes: A dual-tree complex wavelet transform is performed on the locked depth time series signal to decompose the locked depth time series signal into multiple sub-band signals of different scales, wherein the sub-band signal includes at least one high-frequency sub-band signal and one low-frequency sub-band signal; Calculate the integral of the squared modulus of the wavelet coefficients of each high-frequency sub-band signal in the time domain, and use the integral value as the abrupt energy characteristic. Empirical mode decomposition is performed on the low-frequency subband signal, and the first intrinsic mode function is obtained by selection. Calculate the mean of the first intrinsic mode function over the entire time window and the final value at the end of the time window corresponding to the locking depth timing signal. Subtract the mean and the final value from the theoretical locking depth to obtain the corresponding deviations, namely the mean deviation and the final value deviation. The two deviations are combined by weighted summation, and the combined result is used as the steady-state offset feature.

3. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 2, characterized in that, S2 includes: Calculate the first and second derivative sequences of the torque-angle curve; From the first derivative sequence, the point where its value first exceeds the preset first slope threshold is identified as the starting point of the locking process from the unloaded to the pre-tightening stage; the point corresponding to the global maximum value in the first derivative sequence is identified as the starting point of the locking process from the pre-tightening stage to the main locking stage. From the second derivative sequence, find the zero-crossing point where its value changes from positive to negative, and after the zero-crossing point, identify the point in the first derivative sequence where the value first falls below the preset second slope threshold as the starting point of the locking process from the main locking stage to the final tightening stage. The hysteresis loop area sequence of torque as a function of angle is calculated synchronously. If the increment of the hysteresis loop area sequence in the adjacent calculation window exceeds the preset area mutation threshold, the identification position of the starting point of the main lock stage is corrected forward by a preset time compensation amount.

4. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 2, characterized in that, S2 further includes: Establish a three-row, two-column weight allocation matrix. The three rows of the matrix correspond to the pre-tightening stage, the master locking stage, and the final tightening stage, respectively, and the two columns correspond to the mutation energy characteristics and the steady-state offset characteristics, respectively. The matrix elements are assigned values ​​according to the preset rules. The weight of the mutation energy feature is higher in the main locking stage than in the pre-tightening stage and the final tightening stage. The weight of the steady-state offset feature is higher in the final tightening stage than in the pre-tightening stage and the main locking stage. The mutation energy feature and steady-state shift feature corresponding to each stage are multiplied by the corresponding weights in the matrix and summed to obtain the fusion intermediate value of that locking stage. The fusion intermediate values ​​of the pre-tightening stage, the main locking stage, and the final tightening stage, as well as the proportion of the duration of each locking stage to the total locking time, are sequentially concatenated into a feature array. The feature array is subjected to max-min normalization, and the normalized feature array is output as the fusion defect feature vector.

5. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 4, characterized in that, S3 includes: The locking defect classification model consists of a sequentially connected one-dimensional convolutional layer, pooling layer, gated recurrent unit layer, and fully connected classification layer; the fused defect feature vector is reshaped into a two-dimensional tensor with a time step, input into the one-dimensional convolutional layer for convolution operation, and dynamic pattern features of the multi-scale locking process are extracted. The dynamic pattern features of the multi-scale locking process are input into the gated recurrent unit layer to extract the evolution correlation features of its locking stage. Before the fully connected classification layer, an attention mechanism layer is introduced to assign dynamic weights to different time steps of the lock-attachment stage evolution associated features, generating weighted lock-attachment stage attention features. The weighted attention features of the locking stage are input into the fully connected classification layer, and the output is the probability distribution representing four locking defect types: floating, sinking, slipping, and missing lock.

6. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 5, characterized in that, S4 includes: Based on the sign of the steady-state offset feature, determine whether the type of locking defect is floating or sinking; if it is identified as a floating defect, the value of the steady-state offset feature is positive; if it is identified as a sinking defect, the value of the steady-state offset feature is negative. Based on the absolute value of the steady-state offset characteristic, query the preset deviation-compensation angle mapping table to obtain the basic reverse compensation angle. The reverse compensation strategy consists of the superposition of feedforward control and feedback control: the feedforward control is determined by the basic reverse compensation angle; the feedback control angle is obtained by performing proportional-integral calculation on the difference between the real-time acquired locking depth timing signal and the preset target locking depth value during the compensation process, and is used as the feedback control. The total compensation angle command is determined based on the sum of the feedforward control quantity and the feedback control quantity. The electric screwdriver is controlled to rotate in the reverse direction according to the total compensation angle command, and then performs a re-forward locking action according to the preset forward locking parameters.

7. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 1, characterized in that, S5 includes: The phase space of the numerical sequence of mutation energy characteristics during the master-locking phase is reconstructed, and its maximum Lyapunov exponent is calculated. Simultaneously, the centroid frequency of the power spectrum of this numerical sequence is calculated; The maximum Lyapunov exponent and the power spectrum centroid frequency are input into a preset linear discriminant function; the coefficients of the linear discriminant function are obtained by logistic regression training on historical slipped tooth sample data; the linear discriminant function outputs the corresponding screw removal mode command, which includes high-speed pulse screw removal mode, medium-speed oscillation screw removal mode and low-speed constant torque screw removal mode. The electric screwdriver is controlled to perform a reverse rotation action according to the screw ejection mode command. During the screw ejection process, a fast Fourier transform is performed on the real-time monitored output torque signal of the electric screwdriver to calculate the proportion of signal energy in the characteristic frequency band related to the screw thread meshing vibration to the total energy. When this proportion is continuously lower than the preset energy proportion threshold, it is determined that the screw has come loose, the screw ejection is stopped, and the feeder is triggered to perform screw feeding and air blowing cleaning operations.

8. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 7, characterized in that, S6 includes: After the reverse rotation and re-forward locking action or the pin ejection and refeeding operation is completed, the verification time window is activated. Within this verification time window, the locking depth timing signal and the electric screwdriver torque-angle curve are simultaneously acquired; after applying a moving average filter to the locking depth timing signal, the steady-state data is extracted and calculated to obtain the reset verification depth value; For the torque-angle curve of the electric screwdriver, extract the torque data segment after entering the final tightening stage, and calculate the dynamic time warping distance between this data segment and its ideal torque profile curve, which serves as the torque curve matching index.

9. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 8, characterized in that, S6 further includes: Set the dynamic acceptable depth range and the consistency threshold; Compare the reset verification depth value with the qualified depth range; compare the torque curve matching index with the preset matching threshold. The reset operation is considered successful if and only if the reset verification depth value falls within the dynamic qualified depth range and the torque curve matching index is less than the matching threshold.

10. The intelligent detection and automatic reset control method for locking defects based on displacement sensing according to claim 1, characterized in that, S6 further includes: If the reset operation is deemed a failure, the type of locking defect identified in this operation is obtained. Based on the probability distribution output by the locking defect classification model, a confidence-weighted entropy value is calculated to characterize the severity of the defect. The cumulative number of failures at the current locking point is increased by an increment positively correlated with the confidence-weighted entropy value. It is then determined whether the cumulative number of failures exceeds the set repair threshold. If it exceeds the set repair threshold, the intervention process is executed.

Images