A remote operation and maintenance and predictive maintenance method for a bamboo chopping board production line

Abstract

The application provides a bamboo chopping board production line remote operation and maintenance and predictive maintenance method. The method comprises: acquiring working condition parameters and multi-modal sensor signals of each process equipment; performing working condition normalization processing based on the working condition parameters to obtain a degradation component signal; extracting a degradation sensitive feature; weighting and fusing the degradation sensitive feature to obtain an equipment health index and performing cross-process degradation conduction correction. Further, the degradation component signal can also be subjected to frequency domain analysis to identify the degradation source and locate the degradation source process, and the maintenance requirement is redirected to the degradation source process based on the degradation source determination result. The application separates the degradation component through working condition normalization, eliminates the false influence of upstream degradation through cross-process correction, and realizes degradation source identification and accurate maintenance positioning without cross-process communication by utilizing the frequency characteristics of the physical byproducts on the workpiece surface caused by upstream degradation.

Description

Technical Field

[0001] This application relates to the field of industrial equipment health management technology, and in particular to a method for remote operation and predictive maintenance of a bamboo cutting board production line. Background Technology

[0002] Bamboo cutting board production lines typically involve multiple processing steps, including slicing, hot pressing and gluing, milling, sanding, and finishing. Over long-term operation, equipment in each step inevitably experiences degradation, such as tool wear, seal aging, and sanding belt wear. Current maintenance methods primarily rely on periodic preventative maintenance, replacing vulnerable parts at fixed intervals, supplemented by manual inspections to assess equipment condition. However, bamboo, as a heterogeneous natural material, has uneven density distribution both radially and longitudinally. This results in significant load fluctuations when processing different bamboo areas on the same equipment. Alarm methods based on fixed thresholds struggle to distinguish between signal fluctuations caused by changes in operating conditions and signal drift caused by equipment degradation, leading to a high false alarm rate. Furthermore, the degradation status of each process is monitored independently, failing to identify the coupling effect of upstream equipment degradation on downstream equipment, thus affecting the accuracy of degradation assessment.

[0003] In addition, existing cross-process degradation status acquisition schemes rely on dedicated sensor networks and communication protocols to explicitly transmit degradation information between devices. This results in high communication infrastructure costs, cross-process perception interruptions when the network fails, and the inability to determine whether the current process signal anomaly originates from its own degradation or is transmitted from upstream degradation, leading to maintenance decisions pointing to the wrong devices. Summary of the Invention

[0004] To address the shortcomings of existing technologies in bamboo processing equipment, such as high false alarm rates in degradation monitoring due to material heterogeneity, inability to identify cross-process degradation transmission effects in multi-process equipment, and inability to distinguish degradation sources leading to incorrect maintenance decisions, this application provides a remote operation and maintenance and predictive maintenance method for bamboo cutting board production lines. This method separates degradation components by performing operating condition normalization processing on multimodal sensor signals, introduces cross-process degradation transmission correction to eliminate interference from upstream equipment degradation on downstream assessment, and utilizes the frequency characteristics of physical byproducts left on the workpiece surface by upstream degradation to achieve degradation source identification and collaborative maintenance decisions without cross-process communication. This enables reliable degradation trend perception and accurate maintenance location even under non-stationary operating conditions. Specifically, this application provides the following technical solutions: This application provides a remote operation and maintenance and predictive maintenance method for a bamboo cutting board production line, applied to a production line monitoring system including multimodal sensors and edge computing devices. The method includes: Acquire operating parameters and multimodal sensor signals of equipment in each process of the production line; Based on the operating parameters, the multimodal sensor signal is subjected to operating condition normalization processing to obtain the degradation component signal; Extract degradation-sensitive features from the degradation component signal; The degradation-sensitive features are weighted and fused to obtain the equipment health index. In multi-process scenarios, the health index of the current process is corrected for cross-process degradation propagation based on the health index of the upstream process equipment.

[0005] Optionally, the operating condition normalization process includes: determining the expected signal value under the current operating condition based on a pre-established operating condition-signal baseline mapping model, and using the difference between the multimodal sensor signal and the expected signal value as the residual signal.

[0006] Optionally, the method further includes performing an orthogonalization transformation on the residual signal to eliminate the correlation between the signal channels, thereby obtaining the degraded component signal.

[0007] Optionally, the cross-process degradation transmission correction includes: positively compensating the health index of the current process based on the degree of degradation of the upstream process equipment, so as to eliminate the false decrease in the health index caused by the degradation of the upstream process equipment transmitted to the current process.

[0008] Optionally, the method further includes: performing frequency domain analysis on the degradation component signal, determining the characteristic frequency based on the ratio of the spatial spacing parameter of each upstream process processing component to the current process feed speed, extracting the narrowband energy in the frequency band corresponding to the characteristic frequency as the upstream degradation by-product characteristic, determining the degradation source and locating the degradation source process based on the degradation source ratio of the upstream degradation by-product characteristic to its own degradation characteristic, and correcting the maintenance decision based on the determination result of the degradation source, and when the degradation source is dominated by upstream transmission, redirecting the maintenance requirement to the located degradation source process.

[0009] Optionally, the fusion weights of the weighted fusion are updated online based on maintenance event feedback data.

[0010] Optionally, the degradation-sensitive features include at least two of the following: time-domain statistical features, frequency-domain energy features, and time-series trend features.

[0011] Optionally, it further includes: predicting the remaining usable life of the equipment based on the time-series trajectory of the equipment health index, and jointly determining the degradation level based on the equipment health index and the remaining usable life to generate a maintenance work order.

[0012] Optionally, when the operating parameters exceed the preset calibration range, the degradation assessment of the corresponding equipment is suspended.

[0013] The beneficial effects of this application are as follows: By introducing a working condition normalization layer into the preprocessing link of multimodal sensor signals, the working condition component and degradation component in the original signal are separated, so that the subsequent degradation assessment model only processes the signal components related to equipment degradation, thereby avoiding the misjudgment of working condition fluctuations caused by the non-uniformity of bamboo density as equipment degradation. Through a cross-process degradation transmission correction mechanism, the degradation status information of the upstream process is used to compensate for the health assessment of the downstream process, eliminating false degradation judgments caused by the transmission of upstream equipment degradation to the downstream process, and realizing the degradation perception capability of multi-process collaboration; furthermore, by using the characteristic frequency corresponding to the physical by-products left on the workpiece surface by the degradation of the upstream process to identify the degradation source, the source of the current process signal anomaly can be determined and the degradation source process can be accurately located without relying on cross-process communication infrastructure. Based on this, maintenance decisions are redirected to the equipment that truly needs maintenance, avoiding mis-maintenance of non-degraded equipment, reducing maintenance costs and improving maintenance efficiency. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall process of the predictive maintenance method provided in the embodiments of this application.

[0015] Figure 2 This is a schematic diagram of the working condition normalization and orthogonal decomposition subprocess provided in the embodiments of this application.

[0016] Figure 3 This is a schematic diagram of the degradation feature extraction and fingerprint construction sub-process provided in an embodiment of this application.

[0017] Figure 4 This is a schematic diagram of the multimodal fusion and cross-process correction sub-process provided in the embodiments of this application.

[0018] Figure 5 This is a schematic diagram of the remaining useful life prediction and maintenance decision-making sub-process provided in an embodiment of this application.

[0019] Figure 6 This is a schematic diagram comparing signals before and after normalization of operating conditions, provided in an embodiment of this application.

[0020] Figure 7 This is a schematic diagram of the health index degradation curve and remaining life prediction provided in the embodiments of this application.

[0021] Figure 8 This is a schematic diagram showing the comparison of health indices before and after cross-process correction, provided in an embodiment of this application.

[0022] Figure 9 This is a schematic diagram of the predictive maintenance system module architecture provided in an embodiment of this application.

[0023] Figure 10 This is a schematic diagram of the edge-cloud collaborative deployment architecture provided in the embodiments of this application.

[0024] Figure 11 This is a schematic diagram of the degradation source identification and collaborative maintenance decision-making sub-process provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0027] The industrial processing of bamboo cutting boards involves multiple continuous processes, including slicing, hot pressing and gluing, milling, sanding, and finishing. The core equipment in each process experiences varying degrees of degradation over long-term operation, manifesting as progressive deterioration such as increased tool wear, loss of seal elasticity, and shedding of abrasive grains from sanding belts. For such multi-process continuous production lines, the engineering significance of equipment degradation monitoring lies in shifting maintenance from passive post-failure repair to proactive intervention based on degradation status. This allows for scheduling maintenance windows before degradation leads to processing quality deterioration or sudden shutdowns, thereby reducing unplanned downtime and extending the effective service life of critical components.

[0028] Existing equipment condition monitoring solutions mostly rely on single-sensor signal amplitude thresholds for alarm judgment. This type of solution is well-suited for processing homogeneous materials with relatively stable operating conditions. However, bamboo, as a natural heterogeneous material, exhibits a density that decreases radially from the outer bark to the inner layer of the bamboo wall and longitudinally from the base to the tip of the culm. The mechanical properties of different batches of blanks can vary by tens of percentage points. When processing frequently involves blanks of different ages and moisture contents, the normal fluctuations in equipment load are often on the same order of magnitude as the signal shifts caused by early degradation. Criteria based on fixed thresholds cannot distinguish between the two, leading to frequent false alarms or missed alarms. Therefore, this application provides a predictive maintenance method for a bamboo cutting board production line. Through operating condition normalization processing, the operating condition component and degradation component in the multimodal sensor signal are orthogonally separated, allowing the degradation assessment model to process only the signal components related to changes in equipment condition.

[0029] The predictive maintenance method provided in this application can be applied to bamboo processing equipment groups such as automated production lines for bamboo cutting boards, fine processing lines for bamboo cutting boards, and processing lines for bamboo engineered wood products. It can also be extended to the health management of equipment groups in non-homogeneous material processing scenarios such as CNC processing lines for wood products and continuous press lines for engineered wood products. The execution carriers of the method include, but are not limited to, industrial edge computing gateways, embedded industrial control computers, workshop-level industrial servers, and combinations of programmable logic controllers with multi-channel analog signal acquisition capabilities and host industrial control computers. In a typical deployment, edge acquisition nodes are deployed on the equipment side of each process to handle signal preprocessing, workshop-level edge gateways are responsible for feature fusion and health assessment calculations, and cloud servers are responsible for model training and global scheduling.

[0030] The production line monitoring system described in this application includes a multimodal sensor array distributed across various processes and at least one edge computing device. The multimodal sensor array includes at least several types of sensors, such as current sensors, power sensors, vibration acceleration sensors, and acoustic emission sensors. Each sensor establishes a data link with the edge computing device via an industrial Ethernet network. The processor of the edge computing device loads a pre-calibrated baseline parameter file, which contains the operating condition-signal mapping coefficients and health status covariance matrix for each sensor channel in each process.

[0031] like Figure 1 As shown, the predictive maintenance method for a bamboo cutting board production line provided in this embodiment includes the following steps: S000 acquires the operating parameters and multimodal sensor signals of equipment in each process of the production line, and establishes an operating condition-signal baseline mapping model based on the data collected by the healthy equipment under controlled operating condition traversal conditions.

[0032] like Figure 2 As shown, the process of establishing the baseline mapping model is as follows.

[0033] S010, Perform health status verification on key equipment for all processes in the production line.

[0034] The edge computing device sends a calibration start request to the process management system, which then schedules each process's equipment to enter calibration mode. In calibration mode, key components of each process are in a confirmed health state: cutting tools are brand new or have just been sharpened; the hydraulic system has passed a leak rate test to confirm its seal is intact; the sanding belt is unused and brand new; and the heating element's power output has been tested to confirm its deviation is within the rated allowable range. The edge computing device receives health confirmation signals from each device. When all health confirmation signals for key devices are passed, the system enters the operational condition traversal and acquisition phase.

[0035] S020, under controlled operating condition traversal conditions, synchronously collects multiple original sensor signals and corresponding operating condition parameters of equipment in each process.

[0036] The operating parameters include the bamboo density, moisture content, and feed speed of the currently processed blank. The operating condition traversal covers typical combinations of operating conditions that may occur during actual production line operation, including density gradients corresponding to different bamboo ages, moisture content gradients corresponding to different degrees of dryness, and feed speed gradients corresponding to different processing cycles. For example, when bamboo density is set to 3 levels (low density approximately 0.45 g / cm³, medium density approximately 0.65 g / cm³, high density approximately 0.85 g / cm³), moisture content is set to 2 levels (dry and semi-dry), and feed speed is set to 3 levels (low speed, medium speed, high speed), there are 18 possible operating condition combinations. Multiple blanks are processed under each combination to collect a sufficient sample size.

[0037] During the described process cycle, edge acquisition nodes at each process stage synchronously acquire multiple sensor signals via an industrial Ethernet interface using a unified time base. These multimodal sensor signals include current signals from the spindle or drive motor of each process stage, power signals from the feed drive, vibration acceleration signals from key structural components, and acoustic emission signals from the processing area. The sampling rate of each signal channel is set according to the frequency characteristics of the degradation phenomenon of interest; higher sampling rates are used for vibration and acoustic emission channels to capture impact-related degradation characteristics, while lower sampling rates are used for current and power channels to match their gradual variation characteristics. Time alignment between channels is ensured through a time protocol to maintain synchronization accuracy.

[0038] S030, Based on the acquired data, establish the operating condition-signal baseline mapping model for each sensor channel using a parameterized mapping method.

[0039] A parameterized mapping model is independently established for each sensor channel in each process step. This parameterized mapping model uses the operating condition parameters as independent variables and the expected value of the sensor signal from the health device under that operating condition as the dependent variable, establishing a functional relationship between the two. The parameterized mapping model takes the following form: in For the first The channel mapping coefficient vector, This is the extended vector of the operating parameters after feature mapping. The feature mapping includes a constant term, a first-order term for each operating parameter, and an interaction term with physical coupling significance. For example, taking the spindle current channel in a milling process as an example, the operating parameter is bamboo density. Moisture content and feed rate The feature mapping is Interactive items The coupling effect of density and feed rate on cutting force is captured. The mapping coefficients are solved from the calibration data using the least squares method. ,in This is the feature matrix for all sampling points.

[0040] All The mapping coefficients of each sensor channel are arranged in matrix form to obtain the baseline mapping coefficient matrix. ,in This represents the total number of sensor channels. This represents the dimension after feature mapping. For example, when the production line includes 5 processes with a total of 14 sensor channels and the feature mapping dimension is 6, for The matrix. In typical engineering implementations, the total number of channels... For example, it can be 8, 14, 20, etc., depending on the actual number of sensors deployed. The feature mapping dimension... Typically, the value is in the range of 4 to 10; in this embodiment, it is 6.

[0041] In another embodiment, the parameterized mapping model can also be implemented using Gaussian process regression. Gaussian process regression can capture the nonlinear mapping relationship between operating parameters and signals, and is suitable for scenarios where the operating condition-signal relationship exhibits significant nonlinear characteristics. In yet another embodiment, the parameterized mapping model can also be implemented using piecewise linear mapping. Multiple sub-regions are divided in the operating condition parameter space, and linear mappings are established in each sub-region. This is suitable for scenarios where the operating condition-signal relationship exhibits obvious piecewise characteristics.

[0042] S040, estimate the joint covariance matrix of multiple signals under healthy conditions.

[0043] Collect the normalized residuals of each channel in the data from all operating conditions, and calculate the sample covariance matrix. The normalized residual is the difference between the actual signal and the baseline mapping model prediction. The covariance matrix characterizes the inherent correlation structure between signal channels of a healthy device after normalization of operating conditions. For example, in a milling process, spindle current and feed power are positively correlated under healthy conditions (a larger cutting force inevitably leads to a larger feed power consumption), and vibration acceleration and acoustic emission signals are also positively correlated (structural vibration and elastic wave emission share a common excitation source). The estimation of the covariance matrix requires the calibration sample size to meet certain conditions. Empirical conditions are used to ensure the statistical stability of the estimate. When the condition number of the covariance matrix exceeds 1000, it indicates the existence of highly redundant sensor channels, requiring inspection of the sensor configuration or implementation of channel filtering.

[0044] The edge computing device maps the baseline coefficient matrix. Health status covariance matrix and the vector of health baseline mean Write the data into non-volatile memory as fixed parameters for subsequent online operating condition normalization processing.

[0045] S100, based on the operating condition parameters, perform operating condition normalization processing on the multimodal sensor signal to obtain the degradation component signal.

[0046] like Figure 2 As shown, the working condition normalization process performs online transformation on the multimodal sensor signals at each sampling moment during the real-time operation of the production line, specifically including the following sub-steps.

[0047] S110: Obtain the raw multimodal sensor signals and operating parameters at the current moment.

[0048] Each process edge acquisition node continuously acquires the raw signals from each sensor channel at a preset sampling rate. Simultaneously, current operating parameters are obtained from the process management system. The operating parameters include the estimated bamboo density, detected moisture content, and current feed rate setting of the currently processed blank. These parameters are updated once each new blank enters the workstation and remain constant throughout the same blank processing cycle. The acquisition clock for each sensor channel is guaranteed by a precise time protocol uniformly distributed by the edge computing device, with a time alignment error between channels of less than 100 microseconds.

[0049] S120, perform validity verification on the operating condition parameters to determine whether the current operating condition is within the calibration coverage of the baseline mapping model.

[0050] The edge computing device compares each component of the current operating condition parameters with a preset calibration range. For example, when the calibration range for bamboo density is 0.40 g / cm³ to 0.90 g / cm³, the calibration range for moisture content is 6% to 16%, and the calibration range for feed speed is 0.4 to 1.1 times the rated value, if any component of the current operating condition parameter exceeds the corresponding range, the edge computing device marks the current data segment as out-of-range, suspends the degradation assessment of the corresponding device, does not generate a degradation judgment conclusion for that period, and continues to cache the original data for traceability processing after the operating condition returns to normal. This mechanism, from an architectural perspective, avoids false degradation alarms caused by decreased accuracy in the extrapolation region of the baseline mapping model.

[0051] S130, calculate the expected value of the signal under the current operating condition based on the operating condition-signal baseline mapping model, and use the difference between the actual signal and the expected value as the residual signal.

[0052] The edge computing device utilizes a stored baseline mapping coefficient matrix. After performing feature mapping on the operating parameters at the current moment, the expected value of each channel signal is calculated, and the normalized residual signal is obtained by subtracting the expected value from the actual signal. The residual signal Signal fluctuations caused by changes in operating conditions are eliminated, retaining only the signal offset components caused by changes in equipment status and random noise. When the equipment is in a healthy state, the residual signal is a zero-mean random process; when the equipment degrades, the residual signal exhibits a systematic offset, the direction and magnitude of which reflect the mode and degree of degradation.

[0053] like Figure 6 As shown, exemplarily, when milling 3-year-old bamboo (density 0.65 g / cm³, moisture content 10%, feed rate 75% of rated value), the baseline mapping model predicts an expected spindle current of approximately 10.5 A. If the actual measured value is 11.35 A (due to increased cutting force caused by a tool flank wear band width reaching 0.15 mm), then the normalized residual for this channel is... A. This positive offset is attributed to the systematic increase in cutting resistance caused by tool degradation. If a fixed threshold is set directly without performing condition normalization, when the operating condition changes from 3-year-old bamboo to 5-year-old bamboo, the normal increase in current from the health device due to the increased bamboo density will exceed the threshold set based on the low-density operating condition, generating a false degradation alarm. After condition normalization, the signal fluctuations caused by the change in operating condition are accurately predicted and eliminated by the baseline mapping model, and only the true offset caused by degradation is retained.

[0054] In another embodiment, the operating condition normalization process can also be implemented through an adaptive filter. The operating condition parameter signal is used as a reference input, and the operating condition component is estimated and eliminated in real time through a recursive least squares or least mean square adaptive algorithm. This is suitable for scenarios where the operating condition-signal mapping relationship drifts slowly as the equipment ages.

[0055] S140, an orthogonalization transformation is performed on the residual signal to eliminate the correlation between each signal channel, thereby obtaining the degraded component signal.

[0056] The edge computing device utilizes a stored health status covariance matrix. Perform a whitening transformation on the residual signal: in Let be the inverse square root matrix of the covariance matrix. The inverse square root matrix is ​​obtained by... Perform eigenvalue decomposition Take later Obtain, among which The matrix consists of orthogonal eigenvectors. This is an eigenvalue diagonal matrix. After the whitening transformation, the degraded component signal... In a healthy state, each component satisfies the statistical properties of zero mean, unit variance, and mutual independence. When the device degrades, the mean of the degraded component signal deviates from zero, and the direction of deviation corresponds to the projection direction of the degradation onto each sensor channel. The magnitude of the deviation is measured in units of the standard deviation of the healthy state to indicate the severity of the degradation.

[0057] The whitening transformation matrix remains fixed during system operation and is not updated as degraded data accumulates. This design ensures that signal shifts caused by degradation are not absorbed into the transformation matrix, thus preventing a loss of detection sensitivity. When the minimum eigenvalue of the covariance matrix... When the value is too small, perform regularization on the small eigenvalue. ,in This is to avoid excessive amplification of noise by whitening transformation.

[0058] For example, when the normalized residual of the milling operation channel 4 is (Units are A, W, m / s², V, respectively). When the standard deviations of the health status of each channel are 0.5 A, 0.4 W, 0.3 m / s², and 0.15 V, respectively, the components of the whitened degradation component signal are approximately... All values ​​exceeded one standard deviation and were consistently positive, reflecting a consistent degradation pattern where tool wear caused a systematic increase in the signal intensity of each channel. The spindle current channel exhibited the largest whitened amplitude (1.70), indicating that this channel has the highest sensitivity to tool wear degradation.

[0059] In another embodiment, the orthogonalization transformation can also be achieved through independent component analysis (ICA). ICA can separate statistically independent signal source components, and is suitable for scenarios where the degraded excitation source and the operating condition excitation source statistically satisfy the independence assumption. In yet another embodiment, the orthogonalization transformation can also be achieved through the Gram-Schmidt orthogonalization process, which eliminates the linear correlation between channels through channel-by-channel iterative orthogonalization.

[0060] The degraded component signal The data is written into a circular buffer of the edge computing device. The capacity of the circular buffer is not less than twice the length of the sliding window required for subsequent feature extraction steps, so as to ensure the continuous availability of data during the feature extraction process.

[0061] S200, extract degradation-sensitive features from the degradation component signal.

[0062] like Figure 3As shown, the process of extracting degradation-sensitive features involves segmenting the degradation component signal into segments according to a sliding time window, extracting multiple types of degradation-sensitive features within each window and concatenating them into a degradation fingerprint vector. Specifically, it includes the following sub-steps.

[0063] S210, the degraded component signal is segmented according to a sliding time window to obtain the signal data segment within the current evaluation window.

[0064] The edge computing device extracts segments of degenerate component signals from a circular buffer according to preset sliding window parameters. The length of the sliding window... It must cover at least 5 complete processing cycles to ensure that the window contains sufficient processing statistics. The step size of the sliding window... Set to half the window length, i.e. Windows overlap by 50% to avoid missing degradation changes at window boundaries. The data segment for each evaluation window is represented as follows: .

[0065] For example, taking a milling process as an example, the milling cycle of a single anvil blank is approximately 30 seconds, and the window length... Typically, the window duration is between 100 and 200 seconds; in this embodiment, it is 150 seconds. The window step size is 75 seconds, meaning a new degradation evaluation result is generated every 75 seconds. When the sampling rate is 1 kHz, each window contains 150,000 sampling points.

[0066] S220, extract time-domain statistical features for each signal channel within the evaluation window.

[0067] The time-domain statistical features include the root mean square value, kurtosis, and skewness of each channel, reflecting the changes in signal energy, impulse characteristics, and distribution symmetry caused by degradation, respectively. (Regarding the window...) Each channel Calculate the following statistics independently.

[0068] The root mean square value Calculated using the following formula: in This represents the number of sampling points within the window. For the first The channel is within the window. The degraded component signal values ​​at each sampling point. Since each channel of the degraded component signal has a unit variance in a healthy state after whitening transformation, the root mean square value is approximately 1.0 in a healthy state. When device degradation leads to an increase in signal energy, the root mean square value continues to rise.

[0069] The kurtosis Excess kurtosis is defined using the following formula: in For the first The central moment of the first order, The mean value within the window is used. The excess kurtosis is defined such that the kurtosis value is approximately 0 in a healthy state (approximately Gaussian distribution), and increases to a positive value when degradation causes intermittent impact pulses. The kurtosis is extremely sensitive to early local degradation; incidental impacts caused by localized wear surfaces or microcracks can be captured by the kurtosis even before the root mean square value changes significantly. Before calculating the kurtosis, values ​​exceeding the mean within the window are considered. Outliers are removed to prevent transient electromagnetic interference from causing false spikes in kurtosis.

[0070] The skewness Calculated using the following formula: The skewness is approximately 0 in a healthy state (symmetrically distributed). When degradation causes the signal system to shift systematically in a certain direction, the skewness deviates from zero. The sign of the skewness reflects the physical direction of the degradation; positive skewness corresponds to degradation with monotonically increasing cutting force, and negative skewness corresponds to degradation with increasing clearance.

[0071] For example, when the spindle current channel of the milling operation has been running for 200 hours (the width of the tool flank wear band is approximately 0.15 mm), the statistical characteristics within the window are: window mean Second-order central moments Third central moment Fourth-order central moments Based on this calculation, , , The root mean square value of 2.03 is much higher than the healthy reference value of 1.0, indicating that degradation has led to a significant increase in signal energy; the excess kurtosis of 0.65 is positive, indicating the presence of occasional impulse pulses; and the skewness of 0.23 is positive, indicating that the signal system is systematically shifted in the positive direction.

[0072] S230, extract frequency domain energy features for each signal channel within the evaluation window.

[0073] The frequency domain energy characteristics include the energy proportion of characteristic frequencies in each channel and the spectral centroid shift, reflecting the energy changes at characteristic frequencies and the overall spectral shape changes caused by degradation, respectively. For signals within the window... Perform a short-time Fourier transform to obtain the spectrum Based on this, the following frequency domain characteristics are calculated.

[0074] The characteristic frequency energy ratio Calculated using the following formula: in For the first The set of characteristic frequency bands of the corresponding device for each channel. The number of Fourier transform points. The characteristic frequency band is determined by the mechanical structure parameters of the equipment, including the rotational frequency of the rotating parts and a narrow frequency band near their integer multiples of harmonic frequencies. The energy ratio eliminates the influence of changes in the total signal power, focusing only on selective changes in the frequency structure. For example, when the milling spindle speed is 6000 r / min, the tool rotation frequency is 100 Hz, and the characteristic frequency band is set to the range of the fundamental frequency and the second and third harmonics, each plus or minus 5 Hz. Hz. Take At a sampling rate of 1 kHz, the frequency resolution is approximately 0.24 Hz. After 200 hours of device operation, the energy percentage of the characteristic frequency is approximately 0.147, which is significantly higher than the healthy reference value of 0.08, reflecting the enhanced frequency harmonic energy caused by degradation.

[0075] The spectral centroid shift Calculated using the following formula: in The spectral centroid reference value for the healthy state determined in phase S000. Before inputting the degenerate fingerprint vector, the spectral centroid offset is normalized by dividing by the healthy reference value: Degradation typically leads to an increase in high-frequency components (impact noise, friction noise), causing the spectral centroid to shift towards higher frequencies. For example, the spectral centroid of a milling spindle current channel is 285 Hz after 200 hours of machine operation, compared to a healthy reference value of 250 Hz, with a normalized offset of [missing value]. .

[0076] In another embodiment, the frequency domain energy features can also be extracted by continuous wavelet transform. Continuous wavelet transform has more flexible resolution characteristics in the time-frequency plane and can simultaneously capture the time location and frequency components of transient impact events, which is suitable for scenarios where the degradation sensitivity features exhibit transient non-stationary characteristics.

[0077] S240, extract timing trend features for each signal channel within the evaluation window.

[0078] The temporal trend feature is the linear slope of each channel within the evaluation window, reflecting the rate of change of the degraded component signal within the window timescale. The linear slope... Calculated using least squares linear fitting: in This is the center value of the time index. The linear slope is approximately zero in a healthy device state (the degradation component signal is approximately stationary within the window scale), and exhibits a continuous positive value (the signal monotonically increases) as degradation progresses, with its absolute value reflecting the instantaneous rate of degradation development. For example, the linear slope of the milling spindle current channel within the window of the slow degradation phase is approximately... At each sampling point, a positive but small slope indicates that degradation is progressing slowly but has not yet accelerated. When the slope continues to increase, it indicates that degradation has entered an accelerated phase.

[0079] The linear fitting is not applicable when the signal within the window exhibits a step change rather than a gradual trend (such as returning to normal after a sudden tool change). This can be addressed by checking the goodness-of-fit. Make a judgment: if The slope feature of the window is marked as unreliable, and its weight is reduced in subsequent fusion steps.

[0080] S250 concatenates all features from all channels into a degenerate fingerprint vector.

[0081] The edge computing device concatenates the time-domain statistical features (root mean square value, kurtosis, and skewness, a total of 3), frequency-domain energy features (characteristic frequency energy ratio and normalized spectrum centroid shift, a total of 2), and time-series trend features (linear slope, 1) extracted from each channel into a single degenerate fingerprint vector in channel order: The dimension of the degraded fingerprint vector is ,in 6 represents the total number of sensor channels, and 6 represents the number of features extracted per channel. For example, when all 14 channels participate in feature extraction, the dimension of the degraded fingerprint vector is... The dimensions of the degraded fingerprint vector are all dimensionless (temporal features are naturally dimensionless after whitening, and spectral centroid shift is dimensionless after normalization), ensuring the comparability of features in subsequent fusion steps.

[0082] For example, the degenerate fingerprint vectors of four channels (spindle current, feed power, vibration acceleration, acoustic emission) for a milling operation after 200 hours of machine operation: Compared to the degraded fingerprint vector of healthy devices (RMS value of each channel approximately 1.0, kurtosis approximately 0, skewness approximately 0, energy percentage approximately 0.08, centroid offset approximately 0, slope approximately 0), the fingerprint of degraded devices deviates from the healthy baseline in all dimensions. Among them, the vibration acceleration channel has the highest kurtosis (0.88), indicating that this channel is most sensitive to shock pulses caused by degradation.

[0083] When the degenerate fingerprint vector dimension When the dimensions exceed 50, principal component analysis (PCA) can be optionally performed on the degraded fingerprint vector for dimensionality reduction. The transformation matrix of the PCA is determined and fixed by the health status data in the S000 phase and is not updated with the degraded data during operation. The top-performing data with a cumulative variance explanation rate of 95% are retained. One principal component is used to compress the fingerprint dimension to the range of 20 to 50 dimensions. The degenerate fingerprint vector after dimensionality reduction is: ,in Principal component projection matrix, This is the mean vector of healthy fingerprints.

[0084] In another embodiment, the degradation-sensitive features can also be extracted jointly through wavelet packet energy decomposition and empirical mode decomposition. Wavelet packet energy decomposition uses the energy distribution of the signal across multiple frequency bands as the degradation-sensitive features, while empirical mode decomposition decomposes the signal into multiple intrinsic mode functions and extracts the energy and frequency features of each mode. This approach is suitable for scenarios where the degradation-sensitive features exhibit multi-scale time-frequency coupling characteristics. In yet another embodiment, the degradation-sensitive features can also be extracted through the latent variable representation of an autoencoder. The bottleneck layer output of the autoencoder trained on healthy data is used as the degradation fingerprint. This method can automatically learn the feature combination most sensitive to degradation.

[0085] The degraded fingerprint vector The timestamp of the corresponding window is written to the fingerprint time series database for subsequent fusion steps to read.

[0086] S300, the degradation-sensitive features are weighted and fused to obtain the equipment health index. In the case of multi-process scenarios, the health index of the current process is corrected for cross-process degradation propagation based on the health index of the upstream process equipment.

[0087] like Figure 4 As shown, the weighted fusion and cross-process correction process compresses the multidimensional degradation fingerprint vector into a scalar equipment health index and eliminates the interference of upstream process degradation on downstream evaluation. Specifically, it includes the following sub-steps.

[0088] S310 projects the degraded fingerprint vector into the original device health index based on the fusion weight vector.

[0089] The edge computing device reads the latest degraded fingerprint vector from the fingerprint time-series database. Calculate its mean fingerprint compared to the healthy baseline. offset vector and along the fusion weight vector Calculate the degradation distance in the direction:

[0090] The degradation distance This reflects the component of the current degraded fingerprint deviating from the healthy baseline in the main direction of degradation, under the device's healthy state. Approximately zero, the more severe the degradation. The larger the value, the better. Mapping the degradation distance to the interval between 0 and 1 yields the original device health index:

[0091] in This is a mapping reference value, corresponding to the expected degradation distance when the device reaches the failure judgment criteria. A device health index value of 1 indicates that the device is in a fully healthy state, while a value of 0 indicates that the device has reached or exceeded the preset failure threshold.

[0092] The fusion weight vector For a normalized vector, satisfying The relative magnitudes of each component determine the contribution weight of each degradation-sensitive feature to the health index. The initial fusion weight vector is determined by the direction of the first principal component of the covariance matrix of the S000 stage health data, which corresponds to the direction with the largest variance in the degradation fingerprint space. For example, when the sum of the dimensions of the offset vector of the 24-dimensional degradation fingerprint across the 4 channels of the milling process is 6.56, under uniform weighting... Lower degeneration distance .set up The original health index .

[0093] The mapping reference value The initial setting is estimated based on the multiple by which each feature dimension deviates from the health standard deviation, typically in the range of 12 to 25; in this embodiment, 4.0 is used. The mapping reference value is corrected by the subsequent S600 stage based on actual maintenance event data.

[0094] In another embodiment, the weighted fusion can also be implemented using Fisher discriminant analysis. Fisher discriminant analysis finds the projection direction that maximizes the separability between the healthy and degraded classes as the fusion weights, which is suitable for scenarios with accumulated labeled degradation data. In yet another embodiment, the weighted fusion can also be implemented using a weighted average method, where domain experts assign empirical weights to each feature dimension. In yet another embodiment, the weighted fusion can also be implemented using an attention mechanism, dynamically adjusting the weight allocation based on the real-time correlation between each feature dimension and the degradation trend.

[0095] S320 performs cross-process degradation propagation correction on the health index of the current process based on the health index of the upstream process equipment.

[0096] like Figure 8As shown, in multi-process scenarios, the degradation of upstream process equipment leads to a decrease in the quality of its output workpieces, which in turn alters the processing load characteristics of downstream process equipment, causing a shift in the degradation fingerprint of downstream equipment. This shift is not caused by the degradation of the downstream equipment itself; if left uncorrected, it will result in a false decrease in the health index of the downstream equipment. The cross-process degradation transmission correction utilizes the health index of the upstream process to quantify the degree of upstream degradation and accordingly provides positive compensation to the downstream health index:

[0097] in For process The upstream process set, For process To process The degradation transmission coefficient, For upstream processes The degree of degradation. When the degree of degradation of upstream equipment increases ( When the health index declines, the correction formula applies positive compensation to the downstream health index, eliminating the spurious decrease caused by upstream degradation. The corrected health index is truncated to no more than 1.

[0098] The degradation transmission coefficient The coefficients are obtained during the calibration phase of the aforementioned operating conditions or from the historical operating data of the production line. In the process transmission relationship of the bamboo cutting board production line, the transmission coefficient from the shaving process to the milling process is typically in the range of 0.10 to 0.25 (wear of the shaving saw teeth leads to increased surface roughness of the board blank, and fluctuations in the milling allowance increase the milling force), the transmission coefficient from the hot pressing process to the milling process is typically in the range of 0.08 to 0.20 (decreased gluing flatness leads to fluctuations in milling force), and the transmission coefficient from the milling process to the sanding process is typically in the range of 0.15 to 0.30 (increased roughness of the milled surface increases the sanding belt load). These transmission coefficients satisfy the single-process constraints. and multiple upstream cumulative constraints To prevent overcorrection.

[0099] The health index of each process is calculated sequentially according to the process flow (slitting, hot pressing and gluing, milling and shaping, sanding, coating) to ensure that the health index of the upstream process has been calculated when the downstream process performs correction.

[0100] For example, when the original health index of the milling process Upstream slicing process (conduction coefficient) ), upstream hot pressing process (conduction coefficient) When ), the correction calculation is as follows: The correction value of 0.054 indicates that approximately 5% of the decrease in the milling process health index is attributed to changes in workpiece quality caused by degradation in upstream flaring and hot pressing processes, which are then transmitted to the milling load, rather than to degradation of the milling cutter itself. The corrected value of 0.72 more accurately reflects the degradation state of the milling cutter itself.

[0101] In another embodiment, the cross-process degradation transmission correction can also be achieved through nonlinear mapping, establishing a nonlinear regression model between the upstream degradation degree and the downstream health index offset, suitable for scenarios where the transmission effect exhibits saturation or threshold characteristics. In yet another embodiment, the cross-process degradation transmission correction can also be achieved through Bayesian network inference, modeling the degradation state of each process as nodes in a Bayesian network, and using probabilistic inference to jointly estimate the true degradation state of each node.

[0102] S330, Perform time-series smoothing processing on the corrected health index.

[0103] The edge computing device performs exponentially weighted moving average smoothing on the corrected device health index sequence to suppress transient fluctuations in the health index caused by random fluctuations in statistics within a single evaluation window.

[0104] in As a smoothing factor, it controls the response speed to new data. This is a smoothed result from the previous evaluation window. The smoothing factor is typically in the range of 0.1 to 0.3; in this embodiment, it is set to 0.2. A larger smoothing factor results in a more sensitive response to the latest evaluation result but a weaker smoothing effect; a smaller smoothing factor results in a stronger smoothing effect but a slower response to degradation changes. The initial smoothing value is set to 1.0, assuming the device is in a brand-new, healthy state when the system starts.

[0105] For example, when the current adjusted health index previous smoothing value Smoothing factor At that time, the smoothed health index The smoothed value decreased slightly from 0.74 to 0.74 (the change was suppressed by smoothing), demonstrating the ability of the exponentially weighted moving average to filter out transient fluctuations.

[0106] The edge computing device calculates the linear trend slope of the most recent evaluation windows on the smoothed health index sequence to determine the direction of the degradation trend. When the absolute value of the trend slope exceeds a preset trend threshold, it is determined that degradation is in progress (slope is negative) or abnormal recovery (slope is positive); otherwise, it is determined to be in a stable state. The smoothed health index, along with the trend direction label, is output to subsequent steps.

[0107] S400: Based on the time-series trajectory of the equipment health index, predict the remaining usable life of the equipment, and jointly determine the degradation level based on the equipment health index and the remaining usable life to generate a maintenance work order.

[0108] like Figure 5 As shown, the remaining life prediction and maintenance decision-making process fits a degradation model based on the historical trajectory of the equipment health index and extrapolates it to the failure threshold. It then triggers maintenance actions in conjunction with degradation level determination rules. Specifically, it includes the following sub-steps.

[0109] S410, Fit a degradation model to the time-series trajectory of the device health index to estimate the degradation rate parameter.

[0110] The edge computing device or cloud server obtains a smoothed sequence of device health indices from the target device over several recent evaluation windows, and fits the sequence into an exponential degradation model:

[0111] in To fit the initial health index values, For degradation rate parameters, The relative running time is used. The exponential model is transformed into a linear regression form by taking the natural logarithm of both sides. The degradation rate parameter is estimated from the time series data of the health index using the least squares method. and its standard error .

[0112] The degradation rate parameter The physical meaning of is the exponential decay constant of the equipment health index; a larger value indicates faster degradation. The exponential degradation model assumes that the degradation rate is proportional to the current degree of degradation, an assumption consistent with the physical laws of mechanical wear: an increase in the wear surface leads to an increase in the contact area, thus accelerating subsequent wear. For example, when a milling cutter has been running for 850 hours in a milling operation, health index data from the most recent 200 evaluation windows are fitted, and the degradation rate is obtained through log-linear regression. / h, corresponding to a health index half-life of approximately Hour.

[0113] When the coefficient of determination of the fit A value below 0.7 indicates that the exponential degradation model does not fit the actual degradation trajectory well (the device may be in a stable period or experiencing a step event), at which point the linear extrapolation mode should be switched. When the estimated degradation rate... When the value is negative (health index is on the rise), the remaining lifespan is set to infinite, and the degradation level is forcibly judged as normal.

[0114] S420, extrapolate the remaining usable life of the device and its confidence interval based on the degradation rate parameter.

[0115] The edge computing device calculates when the device health index drops to a preset shutdown threshold based on the current device health index value and the estimated degradation rate parameter. The required remaining time is used as the predicted remaining usable lifetime:

[0116] in This is the current smoothed health index value. This is the shutdown threshold. The shutdown threshold is usually set in the range of 0.2 to 0.3. In this embodiment, it is set to 0.3. When the equipment degrades to this health index level, continued operation will significantly affect the processing quality or pose a risk of sudden shutdown.

[0117] Using the standard error of the degradation rate parameter Estimate the confidence interval for the predicted remaining useful life. The half-width of the confidence interval at a 95% confidence level. Calculated using the error propagation formula. For example, when the current health index is 0.67, the shutdown threshold is 0.3, and the degradation rate... / h, remaining usable lifetime prediction Hours, approximately 24 days.

[0118] S430, the degradation level is determined by combining the equipment health index and the remaining usable life.

[0119] The edge computing device or cloud server determines the degradation level of the current device status according to preset maintenance decision rules. These maintenance decision rules are based on a combined condition of a device health index threshold and a remaining usable life threshold, classifying the device status into four levels: normal, warning, alarm, and shutdown. The determination rule is an OR logic: if either the health index condition or the remaining usable life condition is met, the degradation level is upgraded.

[0120] The health index thresholds include early warning thresholds. Alarm threshold and shutdown threshold These are typically set to 0.7, 0.5, and 0.3, respectively. The remaining usable lifetime threshold includes the planned maintenance window. and emergency maintenance window Typically, these are set to 168 hours (one week) and 48 hours (two days), respectively. A warning level is established when the equipment health index falls below the warning threshold or the remaining usable lifespan falls below the planned maintenance window; an alarm level is established when the equipment health index falls below the alarm threshold or the remaining usable lifespan falls below the emergency maintenance window; and a shutdown level is established when the equipment health index falls below the shutdown threshold. All of these threshold parameters can be adjusted by remote maintenance personnel via command issuance.

[0121] For example, when the current health index of the milling process Below the warning threshold of 0.7, while the remaining usable lifespan... When the number of hours exceeds the planned maintenance window of 168 hours, the warning level will be determined based on the health index conditions.

[0122] S440, Generate a maintenance work order based on the degradation level.

[0123] When the degradation level reaches a warning level or above, the system automatically generates a maintenance work order. The maintenance work order includes equipment identification, process name, degradation level, current equipment health index value, predicted remaining useful life and its confidence interval, recommended maintenance time window, and recommended replacement parts information. When multiple devices simultaneously trigger maintenance requests, the system prioritizes the maintenance work orders based on a priority score. This priority score comprehensively considers three dimensions: current degradation level, remaining useful life urgency, and degradation acceleration. The degradation acceleration is approximated by the second difference of the equipment health index.

[0124] The maintenance work order is pushed to the remote operation and maintenance platform via a communication link for maintenance management personnel to review and schedule. After maintenance is completed, maintenance feedback information (maintenance time, actual health index at the time of maintenance, and information on replaced parts) is sent back to the system for Bayesian updates of the degradation rate parameters, thereby gradually improving the accuracy of subsequent remaining usable life predictions.

[0125] like Figure 7 As shown, the time-series trajectory of the equipment health index shows a monotonically decreasing trend starting from 1.0. The degradation model fitting curve matches the actual trajectory. The extrapolated curve intersects with the downtime threshold at the remaining usable life prediction time. The confidence interval gradually widens as the prediction time increases.

[0126] S500: Edge-cloud collaboration and remote operation and maintenance interaction.

[0127] like Figure 10As shown, the data transmission and remote operation and maintenance interaction of the predictive maintenance method are implemented based on an edge-cloud collaborative architecture. Edge acquisition nodes at each process stage perform real-time calculations for signal preprocessing, feature extraction, and health assessment. Diagnostic conclusions are transmitted to the cloud-based remote operation and maintenance platform via an incremental data upload protocol. The cloud aggregates data from multiple devices for horizontal analysis and work order management. Remote operation and maintenance personnel issue parameter adjustment and model update commands to edge devices through the platform.

[0128] The edge computing device maintains the last uploaded value of the health index for each device, and after each health index update, it determines whether the change between the current value and the last uploaded value exceeds a preset incremental threshold. When the change exceeds the incremental threshold, an upload is triggered, encapsulating the current device health index, trend direction, predicted remaining lifetime, and confidence interval into a status message and uploading it to the cloud. If the change does not exceed the incremental threshold, the upload is skipped. This mechanism significantly reduces the upload frequency during stable device operation (only heartbeat messages maintain device online confirmation), and automatically increases the upload frequency during accelerated degradation due to faster changes in the health index. The incremental threshold is typically in the range of 0.01 to 0.05; in this embodiment, it is set to 0.02. The heartbeat interval is 3600 seconds to ensure that the cloud can still confirm the device's online status even when the device health index remains unchanged.

[0129] When a sudden increase in degradation acceleration or a jump in degradation level is detected, the edge computing device automatically triggers an abnormal event snapshot upload. The snapshot contains multi-channel raw sensor signal data from the 300 seconds prior to the trigger, which is extracted back from the circular buffer, losslessly compressed, and transmitted to the cloud for remote experts to perform in-depth root cause analysis.

[0130] The cloud-based remote operation and maintenance platform receives diagnostic messages from all devices, performs horizontal comparisons of degradation rates for devices of the same model, visualizes historical degradation curves for individual devices, and provides an overall health status overview. Remote operation and maintenance personnel can use the platform to issue maintenance decision threshold adjustment commands, fusion weight update commands, and emergency shutdown commands to edge devices. All remote commands are executed with a dual confirmation mechanism: the operation and maintenance personnel input confirmation on the platform, and the edge device receives and sends a confirmation receipt, which is then recorded in its local security log. Data transmission is protected by transport layer encryption and application layer message signatures, and remote commands are managed using role-based access control.

[0131] S600, online model update and adaptive correction.

[0132] The fusion weights of the weighted fusion are updated online based on maintenance event feedback data. When maintenance personnel complete equipment maintenance and enter feedback information (including maintenance time, actual wear measurement value, and human degradation level assessment), the system uses the maintenance feedback data to update the model parameters online.

[0133] The online update process converts the actual wear and tear provided by maintenance feedback into a true degradation distance value compatible with the device health index scale, which serves as the update target. Using the data pair consisting of the degradation fingerprint offset vector before maintenance and the true degradation distance value, the fused weight vector is updated online via a recursive least squares method. The recursive least squares method maintains the gain matrix and forgetting factor. Incremental updates of weights are implemented, with weight adjustments completed with low computational cost each time new maintenance feedback data is obtained, eliminating the need to store all historical data. The forgetting factor is typically in the range of 0.9 to 1.0; this embodiment uses 0.95, allowing the algorithm to assign higher weights to recent maintenance events to adapt to the slow changes in degradation patterns.

[0134] The magnitude of a single update to the fusion weights is limited by security constraints: when the rate of change of the fusion weight vector before and after the update exceeds 10%, the update amount is scaled up to this upper limit to prevent abnormal maintenance events from causing drastic changes in model parameters. After the update is completed, the system substitutes the updated parameters back into the most recent historical predictions for backtesting verification. If the backtesting prediction error is larger than before the update, the update is rolled back, the original parameters are retained, and the reason for the rollback is recorded in the system log. After successful verification, the updated fusion weights are sent to the corresponding edge computing devices via the communication link to take effect.

[0135] The degradation rate parameter The update is achieved through Bayesian prior update: each maintenance event is used as new observation data to update the posterior distribution parameters of the degradation rate, so that the subsequent remaining usable lifetime prediction becomes more accurate with the accumulation of maintenance events.

[0136] S700, perform frequency domain analysis on the degraded component signal to identify the source of degradation, and correct the maintenance decision based on the determination result of the source of degradation.

[0137] like Figure 11 As shown, the degradation source identification and collaborative maintenance decision-making process utilizes the frequency characteristics of physical byproducts left on the workpiece surface by upstream process degradation to determine the source of the current process signal anomaly and locate the degradation source process without the need for cross-process communication. Specifically, it includes the following sub-steps.

[0138] S710, Establish a reference table of characteristic frequencies of by-products.

[0139] During the system initialization phase (executed concurrently with S000), the edge computing device acquires the spatial spacing parameters of the processing components in each upstream process. These spatial spacing parameters include the tooth spacing of the spar saw, the tooth spacing of the milling cutter (determined by the number of blades and the cutter diameter), and the physical distribution spacing of the heating elements on the hot press plate. These spatial spacing parameters are obtained from the equipment's mechanical drawings or through actual measurements, with an accuracy requirement better than 5%.

[0140] For each pair of upstream processes, the characteristic frequency is determined based on the space-time frequency transformation relationship:

[0141] in For upstream processes Degradation byproducts in downstream processes Characteristic frequencies in sensor signals, For downstream processes feed rate, For upstream processes Spatial spacing parameters of the processed components. The physical meaning of the transformation relationship is: the degradation of the upstream process causes its processed components to leave marks on the workpiece surface. As a physical byproduct of spatial cycles (such as periodic scratches caused by saw tooth wear, or periodic tool marks caused by tool wear), when the workpiece moves at a speed During downstream processes, the frequency of the downstream sensor signals is as follows: The narrowband component.

[0142] Considering that the feed rate may fluctuate during actual operation, the edge computing device establishes a frequency band tolerance range for each characteristic frequency:

[0143] in The relative fluctuation coefficient of the feed rate is set to 0.05 in this embodiment. The characteristic frequency bands corresponding to each upstream component must not overlap, i.e. ( This ensures the unique identifiability of the degradation source. When frequency bands overlap, separation is achieved by adjusting the feed rate setting or selecting a higher harmonic frequency band. The characteristic frequency must fall within the effective bandwidth of the downstream sensor and not exceed 40% of the sampling rate.

[0144] For example, when the tooth spacing of the slitting saw... Milling process feed rate At that time, the characteristic frequency of the degenerative byproducts of the saw blade in the sensor signal of the milling process is: The frequency band is .

[0145] During S000 health calibration, the edge computing device records the background energy level in each frequency band. The byproduct characteristic frequency reference table includes the center frequency, frequency band range, and health status background energy of each upstream-downstream process pair.

[0146] The S720 separates the degradation component signal into two channels according to its frequency structure.

[0147] During the real-time operation of the production line, the edge computing device processes the degradation component signals obtained in step S100. Calculate power spectral density Dual-channel separation is performed in each target frequency band calibrated by S710. The dual-channel separation divides the frequency domain energy of the degradation component signal into two positive channels: channel A contains only broadband or trend features caused by its own degradation, and channel B contains only narrowband characteristic frequency components caused by upstream degradation byproducts.

[0148] Power spectral density of the degraded component signal Integrating within each target frequency band calibrated by S710, the energy of each upstream component is obtained:

[0149] The total energy of type B is obtained by summing the energies of each upstream component. Type A energy is obtained through orthogonal separation. .

[0150] The edge computing device updates the frequency band center position in real time based on the current feed speed. When fluctuations in the feed speed cause the frequency band to exceed the preset tolerance range, type B feature extraction is paused and the device is marked as unavailable. The frequency band width must not exceed 50% of the interval between adjacent frequency bands to prevent energy leakage.

[0151] To determine whether the energy in each frequency band is a genuine degradation byproduct signal rather than background noise, the edge computing device calculates the signal-to-noise ratio of each upstream component relative to a healthy baseline:

[0152] Further calculate the energy increment beyond the healthy baseline. , ,in , The baseline energy for the health status calibrated for the S710. It is truncated to zero when the increment is negative.

[0153] S730, a process for determining and locating degradation sources based on the dual-channel energy ratio.

[0154] The edge computing device calculates the degradation source ratio based on the dual-channel energy increment separated from the S720:

[0155] The degradation source ratio The value range is from 0 to 1. When A value close to 0 indicates that the signal abnormality mainly stems from its own degradation. A value close to 1 indicates that the signal abnormality mainly originates from upstream degradation conduction.

[0156] To avoid misjudgments caused by transient fluctuations, the edge computing device... Median filtering is applied to the time series data, and the median of the most recent evaluation windows is used as a stable estimate. Degradation source determination requires that the number of evaluation windows continuously meeting the threshold condition reaches a preset number of confirmation windows. In this embodiment .

[0157] The degradation source determination rule is as follows: when the ratio of filtered degradation sources is lower than a low threshold... When it is determined to be dominated by self-degeneration; when it exceeds a high threshold The condition is determined to be upstream conduction dominant when the threshold is reached; a mixed mode is determined when the threshold is between two thresholds. In this embodiment... , .

[0158] When upstream conduction is determined to be dominant, the edge computing device further performs degradation source localization: the process corresponding to the maximum signal-to-noise ratio of each upstream component is taken as the degradation source localization result. When all All are below the minimum signal-to-noise ratio threshold Even at that time If the threshold is exceeded, the positioning operation will not be performed to avoid false positioning caused by frequency leakage noise. In this embodiment... .

[0159] The confidence level of the localization result is based on the signal-to-noise ratio ratio of the largest energy component to the second largest energy component. Determined: When the maximum value exceeds 1.5 times the second largest value, it is marked as high confidence; otherwise, it is marked as low confidence or multi-source inseparable.

[0160] For example, when the milling process detects , hour, When the threshold is reached, the upstream transmission dominance determination is triggered. At this time... , The maximum value corresponds to the slitting process and exceeds the second largest value by 3.8 times. The location result is the slitting process, and the confidence level is high.

[0161] S740, adjust maintenance decisions based on degradation source determination results.

[0162] The system will output the degradation source determination triplet (degradation source ratio) from S730. Degradation source process positioning The maintenance decisions generated by S400 are corrected using location reliability, and maintenance redirection logic is introduced.

[0163] when Exceeding the high threshold When the location confidence level is high, the system forcibly marks the maintenance work order for the current equipment as being in an upstream transmission suspended state, and simultaneously generates a maintenance redirection work order to direct the maintenance request to the located degradation source process equipment. The current equipment enters a wait-and-see mode, waiting for the upstream equipment to complete maintenance and then observing whether its health index has recovered.

[0164] The security constraint for maintenance redirection is: even if it is determined that upstream transmission is dominant, if the current equipment health index is already below the shutdown threshold... The shutdown protection action will still be executed, prioritizing safety over diagnostic conclusions. The maintenance redirection work order data source is indicated as byproduct characteristic analysis, used for cross-verification between maintenance personnel and the S300 based on communication-based correction results.

[0165] The degradation source ratio It is also used for fine-tuning the cross-process degradation conduction correction of S320. When When the threshold is exceeded, it indicates that a significant proportion of the decrease in the original health index of the current process is attributable to upstream transmission rather than its own degradation. The corrected health index is then adjusted in the following ways:

[0166] in The redirection compensation coefficient is set to 0.4 in this embodiment. The correction formula further refines the compensation by utilizing the degradation source ratio based on the cross-process transmission correction of S320. When the upstream health index obtained by S300 based on communication is inconsistent with the degradation source determination conclusion of S730 based on by-product frequency identification, the correction result that is more beneficial to the current equipment (higher health index) shall be used to ensure that the correction does not underestimate the true health status of the equipment.

[0167] In another embodiment, the degradation source identification can also be achieved through time-domain cross-correlation analysis, using the peak value of the delay cross-correlation function between the downstream process sensor signal and the upstream process motion characteristics as the criterion for degradation propagation. This is suitable for scenarios where the feed rate fluctuation range is too large, making the frequency band method unsuitable. In yet another embodiment, the degradation source identification can also be achieved through causal analysis based on the transfer function, using the Granger causality test to determine whether the upstream process signal change constitutes a statistical causal precursor to the downstream process signal change.

[0168] Based on the same inventive concept, embodiments of this application also provide a predictive maintenance system.

[0169] like Figure 9 As shown, the predictive maintenance system includes a signal acquisition module, a working condition normalization module, a feature extraction module, and a fusion correction module.

[0170] The signal acquisition module is configured to acquire operating parameters of equipment in each process of the production line and multimodal sensor signals. The signal acquisition module includes a multimodal sensor array distributed across each process and edge acquisition nodes. The sensor array includes at least several types of sensors, such as current sensors, power sensors, vibration acceleration sensors, and acoustic emission sensors. The edge acquisition nodes acquire signals from each channel via an industrial Ethernet interface using a unified time reference guaranteed by a precision time protocol, and obtain current operating parameters from the process management system.

[0171] The operating condition normalization module is configured to perform operating condition normalization processing on the multimodal sensor signals based on the operating condition parameters to obtain the degradation component signal. The operating condition normalization module loads a pre-calibrated baseline mapping coefficient matrix and health state covariance matrix, calculates the expected value under the current operating condition for the multimodal sensor signal at each sampling time, obtains the residual signal, eliminates inter-channel correlation through whitening transformation, and outputs the orthogonalized degradation component signal to a circular buffer. The operating condition normalization module also includes an effectiveness gating unit, which suspends the degradation evaluation of the corresponding device when the operating condition parameters exceed a preset calibration range.

[0172] The feature extraction module is configured to extract degradation-sensitive features from the degradation component signal. The feature extraction module reads degradation component signal data segments from the circular buffer using a sliding window, extracts time-domain statistical features, frequency-domain energy features, and time-series trend features for each channel within each evaluation window, and concatenates all features from all channels into a degradation fingerprint vector, which is then output to the fingerprint time-series database.

[0173] The fusion correction module is configured to perform weighted fusion of the degradation-sensitive features to obtain an equipment health index, and for multi-process scenarios, perform cross-process degradation propagation correction on the health index of the current process based on the health index of the upstream process equipment. The fusion correction module reads the degradation fingerprint vector from the fingerprint time-series database, projects it along the fusion weight direction and maps it to a scalar value of the equipment health index, performs positive compensation correction using the upstream process health index and propagation coefficient, and outputs the corrected equipment health index and trend direction after time-series smoothing through an index-weighted moving average.

[0174] The predictive maintenance system also includes a degradation source identification module, configured to perform frequency domain analysis on the degradation component signal to identify degradation sources and correct maintenance decisions. Based on a pre-calibrated reference table of byproduct characteristic frequencies, the degradation source identification module extracts narrowband energy within the corresponding frequency band of each upstream process from the power spectral density of the degradation component signal. It separates the frequency domain energy into two positively correlated channels: its own degradation characteristics and upstream degradation byproduct characteristics. Based on the dual-channel energy ratio, it calculates the degradation source ratio and determines the degradation source. When upstream conduction is determined to be dominant, it matches and locates the degradation source process based on the energy distribution within the corresponding frequency band of each upstream process and redirects the maintenance requirement to the located degradation source process.

[0175] The predictive maintenance system further includes a lifespan prediction module and a communication and coordination module. The lifespan prediction module is configured to fit a degradation model based on the time-series trajectory of the equipment health index and extrapolate to calculate the remaining usable lifespan. It then jointly determines the degradation level based on the equipment health index and the remaining usable lifespan to generate a maintenance work order. The communication and coordination module is configured to transmit diagnostic conclusions to a cloud-based remote operation and maintenance platform via an incremental data upload protocol, receive remote operation and maintenance instructions, and distribute them to the corresponding functional modules for execution.

[0176] The signal acquisition module, operating condition normalization module, and feature extraction module are deployed on edge computing devices in each process step. The fusion correction module is deployed on a workshop-level edge gateway. The lifespan prediction module is deployed on a cloud server or a workshop-level edge gateway. The communication and collaboration module spans the edge layer and the cloud layer. Each module implements its corresponding function by executing computer programs stored in memory through a processor.

[0177] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

[0178] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for remote operation and predictive maintenance of a bamboo cutting board production line, applied to a production line monitoring system including multimodal sensors and edge computing devices, characterized in that, The method includes: Acquire operating parameters and multimodal sensor signals of equipment in each process of the production line; Based on the operating parameters, the multimodal sensor signal is subjected to operating condition normalization processing to obtain the degradation component signal; Extract degradation-sensitive features from the degradation component signal; The degradation-sensitive features are weighted and fused to obtain the equipment health index; In multi-process scenarios, the health index of the current process is corrected for cross-process degradation based on the health index of the upstream process equipment. Frequency domain analysis is performed on the degraded component signal, and the characteristic frequency is determined based on the ratio of the spatial spacing parameter of each upstream process component to the current process feed speed. The narrowband energy within the frequency band corresponding to the characteristic frequency is extracted as the upstream degradation by-product feature, and the remaining frequency domain energy in the degradation component signal after excluding the upstream degradation by-product feature is taken as its own degradation feature. The degradation source is determined and the degradation source process is located based on the degradation source ratio of the upstream degradation by-product characteristics to the degradation source of the degradation characteristics. Based on the determination of the degradation source, the maintenance decision is revised. When the degradation source is dominated by upstream transmission, the maintenance requirement is redirected to the identified degradation source process.

2. The method according to claim 1, characterized in that, The operating condition normalization process includes: determining the expected signal value under the current operating condition based on a pre-established operating condition-signal baseline mapping model, and using the difference between the multimodal sensor signal and the expected signal value as the residual signal.

3. The method according to claim 2, characterized in that, It also includes performing an orthogonalization transformation on the residual signal to eliminate the correlation between the signal channels, thereby obtaining the degraded component signal.

4. The method according to claim 1, characterized in that, The cross-process degradation conduction correction includes: The health index of the current process is positively compensated based on the degree of degradation of upstream process equipment in order to eliminate the false decrease in health index caused by the transmission of degradation of upstream process equipment to the current process.

5. The method according to claim 1, characterized in that, The sources of degradation are identified as including: The frequency domain energy of the degradation component signal is separated into two positive conduction channels: the energy of its own degradation characteristics and the energy of the upstream degradation by-products. The degradation source ratio is determined based on the ratio of the upstream degradation by-products' energy to the total abnormal energy. When the degradation source ratio exceeds a preset threshold for multiple consecutive evaluation windows, it is determined that the upstream conduction is dominant.

6. The method according to claim 5, characterized in that, The determination of the source of degradation also includes: when it is determined that the degradation source is dominated by upstream conduction, matching and locating the degradation source process according to the energy distribution in the corresponding frequency band of each upstream process, wherein the confidence level of the location result is determined based on the ratio of the largest energy component to the second largest energy component.

7. The method according to claim 1, characterized in that, The fusion weights of the weighted fusion are updated online based on maintenance event feedback data.

8. The method according to claim 1, characterized in that, The degradation sensitivity features include at least two of the following: time-domain statistical features, frequency-domain energy features, and time-series trend features.

9. The method according to claim 1, characterized in that, Also includes: The remaining usable life of the equipment is predicted based on the time-series trajectory of the equipment health index, and the degradation level is determined by combining the equipment health index and the remaining usable life to generate a maintenance work order.

Images