Method for evaluating wear of a quiet roller shutter door assembly based on vibration signal analysis

Abstract

The present application relates to the technical field of roller shutter door assembly wear evaluation, and particularly relates to a silent roller shutter door assembly wear evaluation method based on vibration signal analysis, which comprises the following steps: collecting a roller shutter door vibration signal, performing wavelet packet decomposition, reconstructing a denoising signal according to energy entropy and kurtosis product, and generating a window adjustment coefficient; calculating a slope ratio to construct a double-threshold band scaled by the adjustment coefficient to determine candidate pulse points; dividing pulse clusters according to time interval mutations, extracting amplitude Gini coefficient, interval discrete entropy and center frequency offset to fuse into a distortion factor; constructing a wear characteristic vector from the distortion factor cumulative slope and peak interval ratio, determining a wear grade through Mahalanobis distance comparison; recursively updating a covariance matrix based on a forgetting factor, and performing a remaining service interval penalty compression measurement. The present application overcomes the strong noise masking problem and guarantees the reliability of the prediction result.

Description

Technical Field

[0001] This invention relates to the field of wear assessment technology for roller shutter door components, and more particularly to a method for assessing wear of silent roller shutter door components based on vibration signal analysis. Background Technology

[0002] Roller shutters are widely used in industrial, commercial, and residential facilities, serving as crucial equipment for spatial isolation, security, and access control. They typically withstand high-frequency opening and closing operations over extended periods. During continuous operation, core components such as guide rails, slats, rollers, bearings, and transmission mechanisms are subjected to the combined effects of mechanical friction, load fluctuations, impact vibrations, and external environmental disturbances. This gradually leads to increased clearance, lubrication degradation, localized jamming, and cumulative wear. Relying solely on traditional manual inspections or fixed-cycle component replacements makes it difficult to accurately assess the true degradation status of components. This can result in premature replacement leading to over-maintenance, or undetected latent wear causing sudden malfunctions, easily impacting the operational safety and reliability of the roller shutter.

[0003] To achieve real-time perception of the status of door control systems, existing technologies have begun to explore the use of sensors to collect dynamic signals during equipment operation for fault assessment. Referring to Chinese patent application CN116002474A, a fault prediction method for elevator door systems is disclosed. This method involves installing acceleration sensors on the two car doors to collect vertical and horizontal acceleration during the opening and closing process. When the door system is in operation, the real-time detected acceleration curve is compared with a baseline curve during normal operation. If the curve does not match the normal curve, the fault is categorized based on the fault characteristic curve, and the specific fault point is output. This provides a technical path for anomaly detection of door components based on physical signal monitoring.

[0004] However, the vibration signals generated during the opening and closing of roller shutter doors exhibit significant non-stationarity and are often highly coupled with multi-source mechanical noise generated by motor rotation, guide rail friction, and external wind load disturbances. Under continuous operation and variable load conditions, the baseline noise of the gate control system will dynamically drift. If traditional envelope spectrum analysis or full-band RMS monitoring methods are used, the static threshold is easily invalidated when faced with the drift of the baseline noise, making it difficult to effectively isolate weak wear pulses under strong background noise masking. This can easily lead to a large number of false pulses due to baseline failure. At the same time, since the weak impact pulses in the early stage of gate component deterioration are often submerged by transient mechanical noise, the clustering, density abruptness, and cluster structure changes of the impact points on the time axis as mechanical wear intensifies are not fully considered. It is difficult to characterize the early wear distortion state of the components from multiple dimensions such as amplitude imbalance, interval dispersion, and frequency shift.

[0005] In subsequent condition assessment and life prediction, static models and fixed benchmark comparison methods are difficult to adapt to the data distribution drift throughout the equipment's entire life cycle. Because these methods usually use static logic that compares with the initial normal curve, they lack an adaptive sample update mechanism for historical features and an adjustment mechanism for prediction errors, resulting in insufficient accuracy in wear classification and the output remaining life prediction results are often biased towards idealization. Summary of the Invention

[0006] To address the challenges of extracting early, weak wear pulses and achieving high false alarm rates in silent roller shutters under complex variable load conditions due to dynamic drift of background noise, and the difficulty of adapting static state assessment models to the full life-cycle degradation evolution, this invention provides a wear assessment method for silent roller shutter components based on vibration signal analysis.

[0007] This invention provides a wear assessment method for silent roller shutter door components based on vibration signal analysis, comprising: acquiring the opening and closing vibration signal of the roller shutter door and performing wavelet packet decomposition; filtering and retaining the preceding nodes to reconstruct the denoised signal based on the energy entropy and kurtosis product of each sub-band node, and extracting the energy mean of the discarded nodes to generate a window adjustment coefficient; calculating the ratio of the mutation rate of the first half of the denoised signal envelope within the sliding window to the mutation rate of the second half to generate a slope ratio sequence, constructing a double threshold band based on the slope ratio sequence and logarithmically scaling the boundary of the window adjustment coefficient, and determining candidate pulse points exceeding the double threshold band; dividing the pulse clusters according to the mutation position of the time interval of the candidate pulse points, and extracting the amplitude within the pulse clusters. The Gini coefficient, interval discrete entropy, and center frequency offset are weighted and fused into a distortion factor based on the fusion weight. A cumulative evolution curve is constructed based on the distortion factor, and the cumulative slope is extracted. The cumulative slope is divided by the average time interval between the peaks of the distortion factor to obtain the peak interval ratio. A wear feature vector is constructed from the cumulative slope and peak interval ratio of the distortion factor. The wear level is determined by comparing the wear feature vector with the Mahalanobis distance of the sample library. A forgetting factor is generated based on the cumulative slope and the peak interval ratio. The covariance matrix corresponding to the wear level is updated recursively based on the forgetting factor. The remaining service interval is penalized and compressed based on the window adjustment coefficient.

[0008] This invention utilizes wavelet packet decomposition combined with energy entropy and kurtosis product to screen nodes, effectively removing broadband background noise. The energy mean of discarded nodes is mapped to a window adjustment coefficient, which is then used to logarithmically scale the boundaries of the dual threshold band. When background noise intensifies, the dual threshold band automatically expands to reduce the risk of false triggering. Pulse clusters are divided using the reciprocal of the time interval, and multi-dimensional features are weighted and fused to capture subtle anomalies in the early stages of gate component degradation. A wear feature vector is constructed, and a forgetting factor is generated. The forgetting factor is used to recursively update the covariance matrix, accelerating the model's adaptation to abnormal wear samples. Penalty compression calculations are performed based on the window adjustment coefficient, conservatively reducing the lower bound of the remaining service range under conditions of surging noise.

[0009] Preferably, the reconstructed denoised signal includes: truncating the opening and closing vibration signal of the roller shutter door into data slices of equal length, performing wavelet packet decomposition on the data slices to obtain the bottom sub-band nodes, keeping the wavelet packet coefficients of the preceding nodes unchanged, setting the wavelet packet coefficients of the discarded nodes to zero, and reconstructing the denoised signal by inverse transformation through a reconstruction filter bank.

[0010] This invention preserves the pulse component carrying the impact characteristics, thus avoiding signal distortion caused by the reconstruction process.

[0011] Preferably, the generation window adjustment coefficient includes: calculating the energy of discarded nodes below a set threshold, summing the energy and dividing by the total number of discarded nodes to obtain the average energy; obtaining the static baseline energy obtained from the benchmark test under ideal wear-free conditions as the denominator reference value, dividing the average energy by the reference value to obtain the proportion value, and adding 1 to the proportion value as the window adjustment coefficient.

[0012] This invention provides an accurate data benchmark for subsequent dynamic boundary determination by evaluating the background noise intensity of the current system.

[0013] Preferably, constructing the dual threshold band includes: extracting the natural logarithm of the window adjustment coefficient as a boundary expansion factor; multiplying the preset upper bound fixed scaling parameter and lower bound fixed scaling parameter by the sum of the boundary expansion factor and 1, respectively, to generate an upper bound adjustment amount and a lower bound adjustment amount; calculating the sliding median of the slope ratio sequence within the sliding window as the sliding median, and calculating the absolute median difference of the slope ratio sequence within the sliding window as the baseline variation; adding the product of the upper bound adjustment amount and the baseline variation to the sliding median to form the upper bound of the dual threshold band, and subtracting the product of the lower bound adjustment amount and the baseline variation from the sliding median to form the lower bound of the dual threshold band.

[0014] This invention enables the discrimination boundary to adaptively expand with the noise level, thereby reducing the misjudgment rate in high-noise environments.

[0015] Preferably, the step of dividing the pulse clusters based on the abrupt change positions of the candidate pulse point time intervals includes: recording the timestamp sequence of all the candidate pulse points, calculating the time interval between adjacent candidate pulse points, taking the reciprocal of the time interval to obtain a density index sequence, calculating the first-order difference value between adjacent points on the density index sequence, marking the corresponding time point as a cluster splitting point when the absolute value of the first-order difference value exceeds the deviation judgment threshold, and dividing the candidate pulse points into independent pulse clusters using all the cluster splitting points.

[0016] This invention achieves accurate segmentation of independent impact events by relying on the location of abrupt changes in density indices, thereby improving the accuracy of pulse cluster extraction.

[0017] Preferably, the fusion weights are determined by the following process: obtaining a training sample set containing historical real wear labels and corresponding pulse characteristic parameters; establishing a linear regression prediction model with amplitude Gini coefficient, interval discrete entropy, and center frequency offset as independent variables; iteratively optimizing the independent variable coefficients with the goal of minimizing the sum of squared residuals; and after the model converges, normalizing the absolute values ​​of each independent variable coefficient at convergence as the fusion weights.

[0018] Preferably, generating the forgetting factor includes: normalizing the cumulative slope of the current period based on historical extreme values ​​to obtain a first scaling factor, normalizing the peak interval ratio of the current period to obtain a second scaling factor, multiplying the first scaling factor and the second scaling factor to obtain a product value, and subtracting the product value from 1 to obtain the forgetting factor.

[0019] This invention uses a forgetting factor to smooth out the rate of component state deterioration, providing a scientific decay weight for online recursion of the covariance matrix.

[0020] Preferably, the step of recursively updating the covariance matrix corresponding to the wear level based on the forgetting factor includes: performing weighted decay on the historical covariance matrix corresponding to the wear level using the forgetting factor, and completing the online recursive update of the covariance matrix by combining the outer product matrix of the eigenvector residuals of the current period.

[0021] Preferably, the penalty compression calculation includes: obtaining the estimated number of days of the lower bound of the remaining service interval corresponding to the wear level; performing a square root operation on the window adjustment coefficient to obtain the penalty compression denominator; dividing the estimated number of days by the penalty compression denominator to obtain a quotient value, and performing a floor operation on the quotient value to obtain a conservative lower bound of the remaining service days.

[0022] This invention introduces an environmental degradation penalty term and outputs a conservative lower bound for lifetime that includes a safety margin, thus avoiding premature warnings.

[0023] Preferably, the extraction of the amplitude Gini coefficient, interval discrete entropy, and center frequency offset within the pulse cluster includes: extracting the envelope amplitude of candidate pulses and calculating the amplitude Gini coefficient using the Lorentz curve integral method; dividing the pulse time interval into discrete states based on the first quantile and the second quantile, obtaining a state transition probability matrix by counting the number of transitions between adjacent discrete states, and calculating the interval discrete entropy based on the state transition probability matrix; obtaining the actual center frequency by power weighting, and calculating the absolute value of the difference between the actual center frequency and the healthy reference center frequency as the center frequency offset.

[0024] The technical solution of the present invention has the following beneficial technical effects: This invention relies on wavelet packet decomposition combined with energy entropy and kurtosis product to perform refined denoising. It extracts and discards the average energy of nodes to construct a window adjustment coefficient to dynamically scale the boundary of the dual threshold band, effectively removing broadband mechanical background noise interference and reducing the risk of false pulse triggering during the friction noise rise stage. It divides pulse clusters with density index and weighted fuses multi-dimensional features, covering amplitude non-uniformity, temporal disorder and frequency offset. Multi-level and multi-dimensional feature extraction enhances the sensitivity to the initial slight wear state of silent roller shutter door components.

[0025] Furthermore, by constructing component wear feature vectors using cumulative slope and peak interval ratio, and by recursively updating the covariance matrix online based on a specific forgetting factor and comparing with Mahalanobis distance, the algorithm model's tracking speed for abnormal degradation states is improved. By using window adjustment coefficient to perform conservative square root compression and rounding operation on the lower bound of the remaining service range, the potential damage risk of a surge in operating environment noise is fully considered. The compression operation establishes the engineering safety margin bottom line and ensures the reliability of life prediction results under complex scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for assessing wear of silent roller shutter door components based on vibration signal analysis; Figure 2 This is a schematic diagram of dual threshold bands and candidate pulse detection; Figure 3 This is a schematic diagram showing the ablation study and performance comparison. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] Figure 1 This is a flowchart of a wear assessment method for silent roller shutter door components based on vibration signal analysis according to an embodiment of the present invention, with reference to... Figure 1This includes steps S1-S3.

[0029] S1. Candidate pulse point detection based on adaptive dual threshold band.

[0030] The vibration signal of the roller shutter door opening and closing is collected and wavelet packet decomposition is performed; the preceding nodes are selected and retained to reconstruct the denoised signal based on the energy entropy and kurtosis product of each bottom sub-band node, and the average energy of the discarded nodes is extracted to generate the window adjustment coefficient; the steep rise and fall slope ratio of the denoised signal envelope within the sliding window is calculated, and a double threshold band is constructed based on the slope ratio sequence and the boundary is logarithmically scaled by the window adjustment coefficient, and candidate pulse points exceeding the double threshold band are determined.

[0031] In one embodiment, an analog signal is acquired using a piezoelectric accelerometer and converted into a discrete digital sequence. A sampling frequency of 20,000 Hz is set, and a rectangular time window of 0.5 seconds is used to divide the continuous opening and closing vibration signal of the roller shutter door into equal-length, non-overlapping data slices. Each data slice contains 10,000 sampling points. A 4-level orthogonal decomposition is performed on each data slice using a db4 wavelet basis, uniformly dividing the frequency band into 16 bottom-level sub-band nodes. The energy distribution probability of each bottom-level sub-band node is calculated, the energy entropy is calculated, the kurtosis characterizing the signal distribution pattern is obtained, and the energy... The entropy and kurtosis products are multiplied and sorted in descending order. According to the preferred retention ratio of 25% to 40%, the top 4 nodes are selected as the preceding nodes. The wavelet packet coefficient vectors in the preceding nodes are extracted and kept unchanged. Background noise nodes that are stripped due to the energy entropy and kurtosis product being lower than a set threshold are judged as discard nodes. Then, all elements of the wavelet packet coefficient matrix of the bottom 12 discarded nodes are set to 0. The processed coefficients are input into the corresponding conjugate orthogonal reconstruction filter bank to perform wavelet packet inverse transform reconstruction, thereby obtaining a single-slice denoised signal with a length of 10,000 sampling points.

[0032] The average energy is obtained by summing the squares of the wavelet packet coefficients of all discarded nodes and dividing the sum by the total number of discarded nodes. The average energy is typically between 0.025 and 0.085. The static baseline energy, determined through more than 10 benchmark tests under ideal wear-free conditions, is extracted as the denominator reference value and set to 0.015. To construct dynamic adjustment parameters based on the system's inherent noise level, a relationship is established between the window adjustment coefficient, the average energy, and the static baseline energy. The specific relationship is as follows:

[0033] in, Indicates the window adjustment coefficient. The average energy is obtained by summing the energies of all discarded nodes and dividing by the total energy. This represents the static baseline energy, obtained through benchmark testing under ideal, wear-free conditions.

[0034] An analytical transformation is performed on the denoised signal, and the absolute value is taken to obtain the envelope. A sliding window with a fixed step size is constructed, with a length of 100 sampling points. To eliminate ambiguity in the definition of slope extraction, the sliding window is divided into a first half containing sampling points 1 to 50 and a second half containing sampling points 51 to 100. To adapt to the computing power constraints of different levels of controllers, the rate of change is extracted for the envelope amplitude in the first and second half of the data interval. Specifically, for edge nodes with limited computing power, the difference between the first and last endpoints, the difference in extreme values, or the absolute value of the first derivative between the first and second half of the data interval can be used as the rate of change. Under a collaborative architecture with sufficient computing power, a linear regression can also be performed separately, and a linear function can be obtained by fitting it using the least squares method. The absolute value of the slope of the regression line in the first half of the data interval is extracted as the steep rate of change, and the absolute value of the slope of the regression line in the second half of the data interval is extracted as the falling rate of change. Regardless of whether the local trend is upward or downward, the absolute value of the change rate is used to characterize the mutation rate of the data, and the ratio of the mutation rate of the first half to the mutation rate of the second half is used to generate a slope ratio sequence.

[0035] The natural logarithm of the window adjustment coefficient is extracted as the boundary expansion factor. The length of the local statistical window for the slope ratio sequence is set to 100 sampling points. The sliding median of the slope ratio sequence within the sliding window is calculated as the sliding median, and the absolute median difference of the slope ratio sequence within the sliding window is calculated as the benchmark variation for identifying the degree of local fluctuation. To achieve adaptive expansion of the discrimination boundary with noise intensity, the relationship between the upper and lower bounds of the dual threshold band, the sliding median, the boundary expansion factor, and the absolute median difference is constructed as follows:

[0036] in, Indicates the upper bound of the double threshold band. Indicates a lower bound for the double threshold band. This represents the sliding median, obtained by taking the midpoint of the sequence within a sliding window. This represents a fixed scaling parameter at the upper bound, obtained through a pre-set empirical value; in this embodiment, it is set to 2. This indicates a fixed scaling parameter at the lower bound, obtained through a pre-set empirical value; in this embodiment, it is set to 1.5. This represents the window adjustment coefficient, which is calculated using the aforementioned average energy ratio. This represents the absolute median difference, which is obtained through statistical variation calculation within a sliding window.

[0037] Traversing the slope ratio sequence, when a local slope ratio is greater than the upper bound of the dual threshold band or less than the lower bound, the corresponding time position is determined as a candidate pulse point. The effect of dual threshold band and candidate pulse detection is as follows: Figure 2 As shown.

[0038] For example, setting the static baseline energy to 0.015 and the currently measured average energy to 0.045, the calculated window adjustment coefficient is 4. The natural logarithm of the window adjustment coefficient, 1.386, is extracted as the boundary expansion factor, resulting in an upper adjustment limit of 4.772 and a lower adjustment limit of 3.579. If the moving median is 1.2 and the absolute median difference is 0.1, the calculated upper limit of the double threshold band is 1.6772, and the lower limit is 0.8421. If the slope ratio at a certain point is 1.8, since 1.8 is greater than the upper limit of the double threshold band (1.6772), the corresponding time position is determined as a candidate pulse point.

[0039] Thus, wave packet decomposition combined with energy entropy and kurtosis product for reconstruction effectively removes broadband mechanical background noise. The window adjustment coefficient is used to perform nonlinear logarithmic expansion of the dual threshold band, reducing the risk of misjudging weak wear pulses under strong friction noise.

[0040] S2. Distortion factor generation based on multidimensional feature weighted fusion.

[0041] Pulse clusters are divided based on the abrupt changes in the time interval of candidate pulse points; the amplitude Gini coefficient, interval discrete entropy, and center frequency offset within the pulse clusters are extracted, and the aforementioned items are weighted and fused into a distortion factor.

[0042] Record the timestamp sequence of all candidate pulse points, calculate the time interval between adjacent candidate pulse points, and take the reciprocal of the time interval to obtain the density index sequence in Hertz. Calculate the first-order difference between adjacent points in the density index sequence. Set the deviation judgment threshold to three times the standard deviation of the density sequence in the healthy baseline sample. When the absolute value of the first-order difference exceeds the deviation judgment threshold, it is determined that a sudden change in density has occurred. Mark the corresponding time position as a cluster splitting point. Use all cluster splitting points to divide the candidate pulse points into independent pulse clusters. A single cluster typically contains 5 to 20 associated pulse points.

[0043] The envelope amplitudes of candidate pulses within a pulse cluster are extracted and sorted in ascending order. The Gini coefficient of the amplitude is calculated using the Lorentz curve integral method. To calculate the randomness and disorder of pulse occurrence, quantile thresholds are first determined. Time interval data of all pulses under the healthy baseline sample are extracted, and the 33.3% and 66.7% quantiles of the time interval data are calculated as quantile thresholds. According to the quantile thresholds, the pulse time intervals are divided into short, medium, and long discrete states. Time intervals less than or equal to the 33.3% quantile are marked as short intervals; time intervals greater than the 33.3% quantile and less than or equal to the 66.7% quantile are marked as medium intervals; and time intervals greater than the 66.7% quantile are marked as long intervals. The quantile threshold division method ensures the balanced distribution of each state under the baseline state. The continuous time interval sequence within the pulse cluster is mapped to a discrete state sequence. The number of transitions between adjacent discrete states is counted, and the state transition probability matrix is ​​obtained by row normalization. The discrete entropy of the interval is calculated based on the state transition probability matrix.

[0044] Calculate the power spectral density of the pulse cluster, obtain the average frequency by power weighting as the actual center frequency, calculate the absolute value of the difference between the actual center frequency and the healthy reference center frequency as the center frequency offset, and uniformly linearly map the amplitude Gini coefficient, the interval discrete entropy, and the center frequency offset to the interval of 0 to 1 to complete the dimensionless processing.

[0045] Obtain a training sample set containing historical wear labels and corresponding feature parameters, comprising 500 to 2000 records, and construct the following linear regression prediction model:

[0046] in, The wear label represents the actual wear and tear, obtained through physical measurement. The magnitude Gini coefficient, after dimensionless processing, is obtained through linear mapping. The dimensionless discrete entropy is obtained through a linear mapping. This represents the center frequency offset after dimensionless processing, obtained through linear mapping. Indicates the coefficient of the first independent variable. Indicates the coefficient of the second independent variable. This represents the coefficient of the third independent variable. Indicates the bias term. , , and All were obtained through iterative optimization of the model.

[0047] The batch gradient descent algorithm with a learning rate of 0.01 to 0.05 is used to optimize the independent variable coefficients. The independent variable coefficient vector is updated after 1000 to 5000 iterations. When the rate of change of loss between two adjacent iterations is less than 1000, the coefficients are considered to be optimized. When the linear regression prediction model is determined to be fully converged, the absolute values ​​of the coefficients of each independent variable at the time of convergence are extracted and normalized, and the resulting decimals are used as the fusion weights.

[0048] The relationship between the distortion factor and various dimensionless processing features and fusion weights is constructed as follows:

[0049] in, Indicates the distortion factor. This represents the amplitude Gini coefficient after dimensionless processing. This represents the discrete entropy of the interval after dimensionless processing. This represents the center frequency offset after dimensionless processing. This represents the fusion weight corresponding to the amplitude Gini coefficient. This represents the fusion weight corresponding to the discrete entropy of the interval. This represents the fusion weight corresponding to the center frequency offset. , and All were obtained through normalization operations.

[0050] For example, if the dimensionless amplitude Gini coefficient is 0.6, the dimensionless interval discrete entropy is 0.4, and the dimensionless center frequency offset is 0.8, and the fusion weights are 0.2, 0.3, and 0.5 respectively, the distortion factor calculated by substituting them into the relational formula is 0.64.

[0051] In this way, pulse clusters are divided using the reciprocal of the time interval as a density index, and the amplitude Gini coefficient, interval discrete entropy, and center frequency offset are comprehensively extracted. The pulse state is scientifically defined by the quantile threshold, which improves the sensitivity of sensing the initial wear state of the component.

[0052] S3. Wear assessment based on covariance update and penalty compression.

[0053] Wear feature vectors are constructed from the cumulative slope of the distortion factor and the peak interval ratio. Wear level is determined by Mahalanobis distance comparison and covariance matrix is ​​updated. Penalty compression calculation of the remaining service range is performed based on window adjustment coefficient.

[0054] Specifically, the distortion factors generated during a single opening and closing cycle of the roller shutter are accumulated to construct a cumulative evolution curve. The overall upward slope is calculated to obtain the cumulative slope. The maximum points are extracted to form a peak sequence. The average time interval between adjacent peaks is calculated. The cumulative slope is divided by the average time interval to obtain the peak interval ratio. The cumulative slope and the peak interval ratio are used to construct a two-dimensional wear feature vector.

[0055] The Mahalanobis distance between the wear feature vector and the centroid feature vector of each wear level in the cloud sample library is calculated respectively. The covariance matrix of each wear level at the initial time is obtained by initializing the covariance matrix of the corresponding wear level sample in the training sample set.

[0056] The cumulative slope of the current period is normalized based on historical extreme values. The normalization result is truncated to an upper limit of 0.95, thus outputting a first scaling factor between 0 and 0.95. The same operation is performed on the peak interval ratio to obtain a second scaling factor. The forgetting factor is calculated as follows:

[0057] in, Indicates the forgetting factor, This represents the first scaling factor, obtained by truncating the cumulative slope normalization. This represents the second scaling factor, obtained by normalizing the peak interval ratio.

[0058] The covariance matrix corresponding to the candidate wear levels is updated using the exponentially weighted moving average filtering algorithm as follows:

[0059] in, Let represent the updated covariance matrix. The historical covariance matrix is ​​obtained by extracting data from the previous time step. Indicates the forgetting factor, The wear feature vector is obtained by fusing the cumulative slope and the peak interval ratio. The centroid vector representing the wear level is obtained by extracting the mean of the wear level features, with superscript indicating the centroid. Represents the transpose of a vector or matrix.

[0060] Based on historical aging test data, a health degradation curve is established, and the estimated number of days for the lower bound of the remaining service interval is found. The square root of the window adjustment coefficient is performed to obtain the penalized compression denominator, and the conservative lower bound relationship for the remaining service days is constructed as follows:

[0061] in, This represents the lower bound of the conservative remaining service days. This represents the estimated lower limit of the remaining service range in days, obtained by looking up the health degradation curve. This indicates that the penalty is applied to compress the denominator, which is obtained by taking the square root of the coefficients through window adjustment.

[0062] For example, if the estimated lower bound of the remaining service interval is 150 days, the window adjustment coefficient is 4, the penalty compression denominator is 2, and the final output conservative lower bound of the remaining service interval is 75 days.

[0063] Real vibration signals from a certain type of industrial roller shutter door were collected during 5000 consecutive opening and closing aging tests as the test set for ablation experiments. Four control and verification models were set up for the experiment. The baseline group used traditional denoising, a fixed threshold, and static Mahalanobis distance; variant group one used wavelet packet denoising based on the product of energy entropy and kurtosis; variant group two added a dual-threshold adjustment mechanism driven by a window adjustment coefficient; the complete scheme group deployed the entire process including wavelet packet denoising, dual-threshold adjustment mechanism, forgetting factor update, and penalty compression mechanism. The sampling frequency was set to 20000 Hz, and the slice length was 10000 points.

[0064] Under continuous aging tests, the baseline group achieved an average signal-to-noise ratio (SNR) of 12.4 dB, a false alarm rate of 19.3%, and a lifetime prediction error of 23 days. Variant group one, after reconstruction, saw its average SNR increase to 18.5 dB, with the false alarm rate dropping to 11.6%. Variant group two reduced the false alarm rate to 4.5% and the prediction error to 12 days. The complete solution group maintained a stable average SNR of 18.7 dB, with a false alarm rate of only 1.8% and a low average prediction error of 3 days, achieving zero late warnings throughout the entire lifecycle. Figure 3 The results show the comparison between the complete scheme group and the control validation model.

[0065] Thus, the Mahalanobis distance comparison based on the forgetting factor and the lifetime downsampling operation based on the penalty denominator improve the algorithm model's sensitivity to tracking sudden wear. This invention effectively establishes a safety margin baseline for engineering, ensuring a high fault tolerance rate for the monitoring system.

[0066] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for assessing wear of silent roller shutter door components based on vibration signal analysis, characterized in that, include: S1. Collect the vibration signal of the roller shutter door opening and closing and perform wavelet packet decomposition; based on the energy entropy and kurtosis product of each sub-band node, screen and retain the preceding node to reconstruct the denoised signal, extract the energy mean of the discarded node to generate the window adjustment coefficient; calculate the ratio of the mutation rate of the first half of the denoised signal envelope in the sliding window to the mutation rate of the second half to generate the slope ratio sequence, construct a double threshold band based on the slope ratio sequence and logarithmically scale the boundary of the window adjustment coefficient, and determine the candidate pulse points that exceed the double threshold band; S2. Divide the pulse clusters according to the abrupt change position of the candidate pulse point time interval, extract the amplitude Gini coefficient, interval discrete entropy and center frequency offset in the pulse clusters, and fuse them into a distortion factor based on the fusion weight. S3. Construct a cumulative evolution curve based on the distortion factor and extract the cumulative slope, and divide the cumulative slope by the average time interval between the peaks of the distortion factor to obtain the peak interval ratio; construct a wear feature vector based on the cumulative slope and peak interval ratio of the distortion factor, and determine the wear level by comparing the wear feature vector with the Mahalanobis distance of the sample library; generate a forgetting factor based on the cumulative slope and peak interval ratio, recursively update the covariance matrix corresponding to the wear level based on the forgetting factor, and perform a penalty compression calculation of the remaining service interval based on the window adjustment coefficient.

2. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The reconstructed and denoised signal includes: The opening and closing vibration signal of the roller shutter door is truncated into data slices of equal length. Wavelet packet decomposition is performed on the data slices to obtain the bottom sub-band nodes. The wavelet packet coefficients of the preceding nodes are kept unchanged, and the wavelet packet coefficients of the discarded nodes are set to zero. The denoised signal is obtained by inverse transformation reconstruction through the reconstruction filter bank.

3. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The generation window adjustment coefficient includes: Calculate the energy of discarded nodes that are below a set threshold, and sum the energy and divide by the total number of discarded nodes to obtain the average energy; The static baseline energy obtained from the benchmark test under ideal wear-free conditions is used as the denominator reference value. The average energy is divided by the reference value to obtain the proportion value. The proportion value is added to 1 as the window adjustment coefficient.

4. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, Constructing the dual threshold bands includes: Extract the natural logarithm of the window adjustment coefficient as the boundary expansion factor; The upper and lower bound fixed scaling parameters are multiplied by the sum of the boundary expansion factor and 1 to generate the upper and lower bound adjustment values, respectively. The sliding median of the slope ratio sequence within the sliding window is calculated as the sliding median, and the absolute median difference of the slope ratio sequence within the sliding window is calculated as the baseline variation. The upper bound of the dual threshold band is formed by adding the product of the upper bound adjustment and the reference variation to the sliding median, and the lower bound of the dual threshold band is formed by subtracting the product of the lower bound adjustment and the reference variation from the sliding median.

5. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The step of dividing the pulse clusters based on the abrupt change positions of the candidate pulse point time intervals includes: Record the timestamp sequence of all candidate pulse points, calculate the time interval between adjacent candidate pulse points, take the reciprocal of the time interval to obtain the density index sequence, calculate the first-order difference between adjacent points on the density index sequence, and mark the corresponding time point as the cluster splitting point when the absolute value of the first-order difference exceeds the deviation judgment threshold. Use all the cluster splitting points to divide the candidate pulse points into independent pulse clusters.

6. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The fusion weights are determined by the following process: Obtain a training sample set containing historical wear labels and corresponding pulse feature parameters; A linear regression prediction model was established using the amplitude Gini coefficient, the interval discrete entropy, and the center frequency offset as independent variables. The coefficients of the independent variables are iteratively optimized with the goal of minimizing the sum of squared residuals. After the model converges, the absolute values ​​of the coefficients of each independent variable at the time of convergence are normalized and used as the fusion weights.

7. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The generated forgetting factor includes: The cumulative slope of the current period is normalized based on historical extreme values ​​to obtain a first scaling factor, and the peak interval ratio of the current period is normalized to obtain a second scaling factor. The first scaling factor and the second scaling factor are multiplied to obtain a product value, and the forgetting factor is obtained by subtracting the product value from 1.

8. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The step of recursively updating the covariance matrix corresponding to the wear level based on the forgetting factor includes: The historical covariance matrix corresponding to the wear level is weighted and decayed using the forgetting factor, and the covariance matrix is ​​updated online by combining the outer product matrix of the eigenvector residuals of the current period.

9. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The penalty compression calculation includes: Obtain the estimated number of days of the lower bound of the remaining service interval corresponding to the wear level; Perform a square root operation on the window adjustment coefficient to obtain a penalized compressed denominator; Divide the estimated number of days by the penalty compression denominator to obtain the quotient, and perform a floor operation on the quotient to obtain the lower bound of the conservative remaining service days.

10. The wear assessment method for silent roller shutter door components based on vibration signal analysis according to claim 1, characterized in that, The extraction of the amplitude Gini coefficient, interval discrete entropy, and center frequency offset within the pulse cluster includes: Extract the envelope amplitude of the candidate pulses and calculate the amplitude Gini coefficient using the Lorentz curve integral method; The pulse time interval is divided into discrete states based on the first quantile and the second quantile. The number of transitions between adjacent discrete states is counted to obtain the state transition probability matrix. The interval discrete entropy is calculated based on the state transition probability matrix. The actual center frequency is obtained by weighting by power, and the absolute value of the difference between the actual center frequency and the healthy reference center frequency is calculated as the center frequency offset.

Images