A data analysis method of an intelligent monitoring device
By aligning saw blade wear status data through corner domain resampling and asymmetric feature extractors, the identification problem of low sampling rate sensors under variable speed conditions is solved, providing reliable wear status predictions and confidence intervals, and improving the accuracy and reliability of predictive maintenance.
Patent Information
- Application Number
- CN202611254032.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies using low-cost, low-sampling-rate microelectromechanical system vibration sensors on large processing equipment such as band saws are insufficient to achieve the same level of saw band wear condition identification as high-precision piezoelectric sensors in the laboratory. Furthermore, wear characteristics drift under varying speed and load conditions, leading to identification failures and a lack of reliable predictive maintenance decisions.
By acquiring data from a laboratory high-frequency piezoelectric accelerometer and a field low-cost microelectromechanical system vibration sensor, an asymmetric dual-stream feature extractor is designed using rotational speed-correlated signals for angular domain resampling. This is combined with latent space distribution difference loss and gradient inversion layer adversarial loss to achieve feature distribution alignment and output the confidence interval of the wear state.
It achieves stable wear characteristic representation under varying speed and load conditions, avoids the cost of manual calibration on site, provides reliable wear state prediction results and quantifiable confidence intervals, and reduces the risk of misjudgment.
Smart Images

Figure CN123115779A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of condition monitoring and fault diagnosis technology for intelligent manufacturing equipment, and in particular to a data analysis method for an intelligent monitoring device applied to large processing equipment such as band saws. Background Technology
[0002] With the continuous upgrading of the intelligent manufacturing equipment industry, real-time online monitoring of the wear status of cutting tools or saw bands on large processing equipment such as CNC band saws has become a key link in improving production efficiency, reducing scrap rate, and supporting the construction of predictive maintenance systems. Currently, a large number of theoretical studies and laboratory models are based on high-frequency, high-precision piezoelectric accelerometers to acquire cutting vibration signals, and then combined with deep learning methods to establish wear state prediction models. One-dimensional convolutional neural networks, residual networks, and other structures are often used for automatic feature extraction of vibration signals. Order analysis, as a technique for resampling and spectral analysis of vibration signals based on spindle angle rather than time, has a relatively mature mathematical foundation in the field of rotating machinery fault diagnosis. It is usually used to estimate the rotation speed signal by using spindle encoders or motor current to complete angular domain resampling. In addition, the distribution alignment method based on the maximum mean difference and the adversarial domain adaptation method based on gradient inversion layer in domain adaptation theory have also been verified in image recognition and some mechanical fault diagnosis transfer scenarios.
[0003] The closest existing technical solution to this invention is a scheme that trains a deep learning wear prediction model based on high-precision sensor data from the laboratory and directly deploys it to the industrial site. Its typical implementation steps are as follows: Under controlled laboratory conditions, a high-frequency piezoelectric accelerometer is used to collect the vibration time-domain signal of the tool or saw blade throughout its entire life cycle and divides it according to a fixed sampling frequency. The segmented signal is subjected to short-time Fourier transform or wavelet transform to extract time-frequency domain features. A supervised training sample set is constructed by combining manually labeled wear level labels. A convolutional neural network or recurrent neural network is used to fit the mapping relationship between time-frequency features and wear labels to obtain a wear state prediction model. The trained model is directly deployed to the monitoring device in the industrial site. The vibration signal collected on site is subjected to the same time-frequency transformation and then input into the model to output the wear level prediction result.
[0004] The aforementioned existing technical solutions have the following shortcomings: First, the solution assumes that the field-deployed sensors and laboratory sensors have similar sampling rates, bandwidths, and signal-to-noise ratios. However, in industrial settings, low-sampling-rate microelectromechanical system (MEMS) vibration sensors are commonly used to control costs. Their sampling rates, bandwidths, and signal-to-noise ratios differ significantly from those of laboratory high-frequency piezoelectric sensors. This leads to a severe mismatch between the directly deployed model input feature distribution and the training distribution, making it difficult to directly apply the high-precision prediction model, which requires significant computing power, to industrial settings. Second, the time-frequency feature extraction in this solution is based on the assumption that the cutting parameters are in a steady state. However, in the actual band sawing process, the rotational speed and load have high time-varying characteristics, causing the vibration frequency components corresponding to the same wear stage to drift at different rotational speeds. The frequency features of different wear stages overlap, making it difficult to achieve stable wear stage segmentation under unsupervised conditions. Third, the prediction model in this solution directly outputs deterministic results of the wear level without quantifying the reliability of the prediction results. This makes it difficult for field maintenance personnel to determine the reliable boundaries of the prediction results when the target domain data is unlabeled and the model's generalization ability is unknown, increasing the risk of misjudgment in predictive maintenance decisions. Summary of the Invention
[0005] The technical problem this invention aims to solve is: in the process of intelligent transformation of large processing equipment such as band saws, how to utilize the unlabeled data collected by low-cost, low-sampling-rate microelectromechanical system vibration sensors deployed on-site to achieve the same accuracy in identifying saw band wear as the model trained by high-precision piezoelectric sensors in the laboratory, while overcoming the identification failure problem caused by wear feature drift under unsteady working conditions such as variable speed and variable load, and providing a reliable decision-making basis with confidence for predictive maintenance.
[0006] To address the aforementioned technical problems, this invention provides a data analysis method for an intelligent monitoring device, comprising the following technical solutions: acquiring source domain vibration time-domain signals collected by a laboratory high-frequency piezoelectric accelerometer and target domain vibration time-domain signals collected by a low-cost microelectromechanical system vibration sensor deployed in an industrial field, as well as rotational speed correlation signals related to the rotation of the equipment spindle. The source domain vibration time-domain signals are labeled with wear level, while the target domain vibration time-domain signals are unlabeled. Instantaneous angular velocity is estimated based on the rotational speed correlation signals. The instantaneous angular velocity is integrated over time to obtain the cumulative rotation angle. A sequence of equal-angle sampling points is generated at a preset angle sampling interval. The source domain and target domain vibration time-domain signals are resampled using angular domain interpolation based on the mapping relationship between the cumulative rotation angle and the acquisition time. Finally, the resampled signals are analyzed... The scattered Fourier transform yields source domain order features and target domain order features independent of rotational speed. The source domain order features are input into a source domain feature extractor with a different structure than the target domain feature extractor to extract latent source domain features. The target domain order features are input into a lightweight target domain feature extractor to extract latent target domain features. The source domain latent features and their labels are used to calculate the source domain classification loss. The latent space distribution difference loss is calculated using the source domain latent features and the target domain latent features. After passing both through a gradient inversion layer, they are input into a domain discriminator to calculate the domain adversarial loss. A weighted joint loss is constructed and backpropagated to update the network parameters until convergence. The converged target domain latent features are input into the wear classifier. During the inference phase, the classifier is kept in a randomly deactivated state and kept in an active state for multiple independent forward propagations. The predicted mean and variance are calculated, and the confidence interval for the wear state prediction is constructed and output.
[0007] The beneficial technical effects that this invention can achieve include:
[0008] 1. This invention converts the non-steady-state time-domain vibration signal into a stationary order-domain feature sequence independent of the rotational speed by angular domain resampling based on instantaneous rotational speed. This fundamentally eliminates the fundamental frequency drift interference caused by variable rotational speed and variable load conditions of the band saw, so that the same wear stage presents a stable and consistent feature representation under different working conditions.
[0009] 2. This invention achieves feature distribution alignment in the latent space between high-sampling-rate data in the source domain and low-sampling-rate data in the target domain by designing asymmetric dual-stream feature extractors with different structures for the source and target domains respectively, and by combining latent space distribution difference loss and gradient reversal layer adversarial loss for joint optimization. This enables the identification of wear status of unlabeled data in the field using only laboratory historical labeled data, avoiding the cost of manual labeling during on-site shutdown.
[0010] 3. This invention approximates Bayesian inference by maintaining random deactivation and activation during the inference phase and performing multiple independent forward propagations. It outputs the cognitive uncertainty confidence interval of the wear state prediction result, providing a quantifiable and verifiable basis for predictive maintenance decisions and avoiding the risk of misjudgment caused by a single deterministic prediction result. Attached Figure Description
[0011] Figure 1 This is a flowchart of the non-steady-state operating condition order feature resampling and reconstruction process of the present invention;
[0012] Figure 2 This is a flowchart of the cross-sampling rate asymmetric feature alignment network of the present invention;
[0013] Figure 3 This is a flowchart of the unsupervised adaptive evaluation process for wear state with confidence level according to the present invention. Detailed Implementation
[0014] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0015] This invention provides a data analysis method for an intelligent monitoring device, applicable to large processing equipment such as band saws. It consists of three sequentially connected functional modules: an order feature resampling and reconstruction module for unsteady-state conditions, an asymmetric feature alignment network module across sampling rates, and an unsupervised adaptive evaluation module for wear state with confidence. The output of each module serves as the input to the next, collectively forming a complete technical link from the original vibration signal to a high-confidence wear state output. The intelligent monitoring device involved in this invention includes a source domain data acquisition unit, a target domain data acquisition unit, an order resampling unit, an asymmetric feature extraction unit, a latent space alignment unit, a Bayesian uncertainty evaluation unit, and a predictive maintenance decision output unit. The source domain data acquisition unit is a high-frequency piezoelectric accelerometer deployed in a laboratory environment, used to collect vibration data of the tool or saw band throughout its entire lifespan under controlled conditions. The system acquires time-domain vibration signals from the source domain and pre-labels them with wear level tags. The target domain data acquisition unit is a low-cost microelectromechanical system vibration sensor installed near the saw wheel of a band saw. It is used to acquire vibration time-domain signals under actual operating conditions in the industrial field. These signals do not have wear level tags. The order resampling unit includes two resampling sub-units corresponding to the source domain data acquisition unit and the target domain data acquisition unit, respectively. It is used to convert the time-domain vibration signal into order-domain features independent of rotational speed. The asymmetric feature extraction unit includes a source domain feature extractor and a target domain feature extractor. The two have different network structures. The latent space alignment unit is used to calculate the joint loss and drive network training. The Bayesian uncertainty assessment unit is used to output the wear state prediction results with confidence intervals. The predictive maintenance decision output unit can be an edge computing gateway or a programmable logic controller. It is used to receive the wear level prediction results and confidence intervals and generate maintenance decision prompts.
[0016] I. Order Feature Resampling Reconstruction under Unsteady-State Conditions
[0017] This module abandons the traditional time-frequency analysis method based on fixed time intervals and adopts angular domain resampling technology based on instantaneous rotational speed to convert the non-steady-state time-domain vibration signal collected by the band saw under variable speed and load conditions into a stationary order domain feature sequence independent of rotational speed. The specific implementation steps are as follows.
[0018] The first step is to collect vibration and rotational speed-related signals from the band saw. These vibration signals are acquired at fixed time intervals using a piezoelectric accelerometer or a microelectromechanical system (MEMS) vibration sensor, and are denoted as time-domain signals. The speed-related signal is indirectly estimated through the spindle encoder pulse signal or the servo motor current signal. The two can be used alone or in combination to improve the speed estimation accuracy.
[0019] The second step is to estimate the instantaneous angular velocity. This involves performing spectral peak tracking processing on the encoder pulse count or current signal to obtain an estimated value of the instantaneous angular velocity of the band saw spindle or saw wheel at any given moment, denoted as […]. .
[0020] The third step is to calculate the cumulative angle by integrating the instantaneous angular velocity over time to obtain the cumulative angle function:
[0021] ;
[0022] in, Indicates the time from the start of data collection to time... The cumulative angle of rotation, in radians. express The instantaneous angular velocity at time t. It is an integral variable, and its value ranges from zero to... The closed interval between.
[0023] The fourth step is to determine the sequence of equal-angle resampling points, set a fixed angle sampling interval, and generate equal-angle sampling points:
[0024] ;
[0025] in, Indicates the first The cumulative rotation angle corresponding to each equal angle sampling point Indicates the preset angle sampling interval. The sampling point number, This represents the total number of points in the signal sequence after resampling.
[0026] The fifth step involves resampling the original time-domain signal using angular domain interpolation. Utilizing the mapping relationship between time and rotation angle obtained in the third step, cubic spline interpolation is used to convert the signal, originally sampled at equal time intervals, into a signal sampled at equal rotation angle intervals.
[0027] ;
[0028] in, Indicates the angle after resampling The signal amplitude at that location, This represents the interpolation function, whose inputs are the original signal, the mapping relationship between time and rotation angle, and the target angle point.
[0029] Step 6: Perform an order spectrum transform on the equal-angle domain signal, and then perform a discrete Fourier transform on the resampled signal sequence to obtain an order spectrum independent of rotational speed.
[0030] ;
[0031] in, This represents the magnitude of the order spectrum at the order value. This represents the harmonic order relative to one revolution of the spindle; it is a dimensionless quantity and is independent of rotational speed. It is the imaginary unit.
[0032] The seventh step involves using the aforementioned order spectrum as a stationary feature representation of the signal segment and outputting it to the subsequent asymmetric feature alignment network module to complete feature reconstruction from the non-steady-state time domain to the stationary order domain. In the above resampling process, a speed change judgment step is also set: continuously monitor the change amplitude of instantaneous angular velocity, and when the change amplitude exceeds a preset threshold, it is determined that a speed change has occurred. Taking the time of the change as the segment node, the angular domain interpolation resampling and order spectrum transformation of the signals before and after the change are re-performed in steps three to six, thereby avoiding angular domain distortion in a single segment due to drastic changes in speed. In a specific embodiment, the sampling rate of the piezoelectric accelerometer is 25.6 kHz, the sampling rate of the microelectromechanical system vibration sensor is 2 kHz, and the angle sampling interval is 0.1 degree. The above parameter configuration can take into account the difference in sampling capabilities between the two types of sensors while ensuring order resolution.
[0033] II. Asymmetric Feature Alignment Network Across Sampling Rates
[0034] This module addresses the heterogeneous characteristics of the source domain data, which has a high sampling rate, large data volume, and accurate labels, and the target domain data, which has a low sampling rate, limited data volume, and no labels. It designs an asymmetric two-stream feature extractor to achieve cross-domain feature distribution alignment in the latent space. The specific implementation steps are as follows.
[0035] The first step is to construct an asymmetric dual-stream feature extractor. The source domain branch adopts a deep one-dimensional residual network, denoted as the source domain feature extractor, which is composed of multiple stacked one-dimensional residual blocks. Its convolutional kernel receptive field is designed to be relatively large, which can make full use of the high-frequency detail information contained in the high-sampling rate data of the source domain. The target domain branch adopts a lightweight network, denoted as the target domain feature extractor. It adopts a depthwise separable one-dimensional convolutional structure to reduce the number of parameters and computational cost, adapting to the characteristics of low-sampling rate data in the target domain and the computing power constraints of the on-site edge computing hardware. The source domain feature extractor and the target domain feature extractor are different in terms of network depth, convolutional kernel size and receptive field size, which is different from the technical route of existing domain adaptation methods that use shared or structurally identical feature extractors for the source and target domains.
[0036] The second step is to extract the latent space features of the source and target domains. The source domain order features are input into the source domain feature extractor to obtain the source domain latent features, denoted as... The target domain order features are input into the target domain feature extractor to obtain the target domain latent features, denoted as... .
[0037] The third step is to calculate the supervised classification loss of the source domain. The latent features of the source domain are input into the wear classifier, and the cross-entropy classification loss is calculated using the existing accurate labels in the source domain.
[0038] ;
[0039] in, Represents the source domain classification loss. Indicates the total number of samples in the source domain. Indicates the total number of wear level categories. Indicates the first Each source domain sample in category The actual label on it takes a value of zero or one. Indicates the wear classifier for the first... Each source domain sample is predicted as a category. The probability of.
[0040] The fourth step is to calculate the distribution difference loss of latent features between the source and target domains. This difference is measured in the Hilbert space of the reproducing kernel, and the maximum mean difference method is used for calculation.
[0041] ;
[0042] in, Indicates the loss due to distributional differences. Indicates the total number of samples in the target domain. This represents mapping features to a reproducing kernel Hilbert space. kernel mapping function, , They represent the first The source domain sample and the first The latent space features of a target domain sample, where the norm represented by the symbol is the norm in the Hilbert space.
[0043] The fifth step involves constructing a domain adversarial loss through a gradient inversion layer. The latent features of both the source and target domains are passed through this layer and then input into the domain discriminator to calculate the domain discriminant adversarial loss.
[0044] ;
[0045] in, Indicates domain adversarial loss, This represents the latent features of samples in the source or target domain. If the index does not exceed the total number of samples in the source domain, it is taken from the source domain; otherwise, it is taken from the target domain. This represents the field label of the sample, with the source field having a value of one and the target field having a value of zero. This indicates a gradient reversal layer, which performs an identity mapping during forward propagation and multiplies the gradient by a negative coefficient during backward propagation. The domain discriminator outputs the probability that a sample belongs to the source domain. The introduction of the gradient reversal layer creates an adversarial training relationship between the domain discriminator and the two feature extractors: the domain discriminator attempts to accurately distinguish whether a sample comes from the source domain or the target domain, while the two feature extractors are driven by the reversed gradient to generate feature representations that are difficult for the domain discriminator to distinguish, thereby achieving distribution alignment in the latent space.
[0046] Step 6: Construct a joint loss function and backpropagate to update the network parameters, summing the three losses in a weighted manner:
[0047] ;
[0048] in, Indicates joint training losses, , These represent the weighting coefficients for the distributional difference loss and the adversarial loss, respectively, used to balance the contributions of the three losses. In a specific embodiment, the weighting coefficients... The value is 0.5, and the weighting coefficient is... The value is set to 0.3. By minimizing the joint training loss, the backpropagation algorithm is used to alternately update the parameters of the source domain feature extractor, the target domain feature extractor, the wear classifier, and the domain discriminator until the network converges, thereby aligning the feature distributions of the source domain and the target domain in the latent space.
[0049] III. Unsupervised Adaptive Evaluation of Wear State with Confidence Level
[0050] After the model completes heterogeneous feature alignment in the latent space, it uses Monte Carlo random deactivation approximate Bayesian inference to output wear prediction results and their cognitive uncertainty confidence intervals. The specific implementation steps are as follows.
[0051] The first step is to retain a random deactivated layer in the fully connected layer of the trained wear classifier and keep the random deactivated layer active during the model inference stage, but turn it off during non-conventional inference, thereby approximating the posterior sampling of parameters in the Bayesian neural network.
[0052] The second step involves performing multiple independent forward propagations on the latent features of the target domain after manifold alignment. Each forward propagation, due to the random nulling effect of random inactivation, is equivalent to sampling a different set of network parameters from the approximate posterior distribution, resulting in multiple sets of predicted outputs, denoted as... to .
[0053] The third step is to calculate the point estimate, i.e., the mean, of the wear condition prediction:
[0054] ;
[0055] in, The final predicted output represents the wear condition. This indicates the total number of forward propagations in Monte Carlo. Indicates the first The forward propagation corresponds to the... Predicted output of group of randomly sampled parameters.
[0056] The fourth step is to calculate the cognitive uncertainty of the prediction results, i.e., the variance:
[0057] ;
[0058] in, This indicates the model's cognitive uncertainty regarding the prediction result for that sample; the larger the value, the lower the reliability of the model's prediction result.
[0059] Fifth step: Construct confidence intervals for the prediction results:
[0060] ;
[0061] in, This represents the confidence interval for predicting wear condition. This represents the standard normal distribution quantile corresponding to the confidence level. This represents the square root of the variance obtained in step four, i.e., the standard deviation of cognitive uncertainty. In a specific embodiment, the number of Monte Carlo forward propagations... The value is taken as fifty times, and the confidence level is taken as 95%.
[0062] The sixth step is to output the wear level point estimate and confidence interval together to the predictive maintenance decision output unit, and determine whether the width of the confidence interval exceeds the preset threshold. If it exceeds the threshold, a low confidence warning sign is triggered, prompting maintenance personnel to further verify the prediction result in combination with human experience, thereby achieving a reliable adaptive assessment of wear status under unsupervised conditions.
[0063] IV. Device Examples
[0064] This invention also provides an intelligent monitoring device for implementing the above-mentioned data analysis method, comprising a source domain data acquisition unit, a target domain data acquisition unit, an order resampling unit, an asymmetric feature extraction unit, a latent space alignment unit, a Bayesian uncertainty assessment unit, and a predictive maintenance decision output unit. The source domain data acquisition unit is connected to the order resampling unit, and the target domain data acquisition unit is connected to the order resampling unit. The output of the order resampling unit is respectively connected to the source domain feature extractor and the target domain feature extractor with different structures in the asymmetric feature extraction unit. The outputs of the source domain feature extractor and the target domain feature extractor are jointly connected to the latent space alignment unit. The latent space alignment unit includes a classification loss calculation subunit, a distribution difference measurement subunit, and a domain decision unit connected via a gradient inversion layer. The sub-unit is used to construct the joint loss during the training phase and feed back to update the network parameters of the asymmetric feature extraction unit. After training, the latent features of the target domain output by the target domain feature extractor are input into the Bayesian uncertainty assessment unit. The Bayesian uncertainty assessment unit is connected to the predictive maintenance decision output unit and transmits the wear level prediction results and confidence intervals to the predictive maintenance decision output unit. The predictive maintenance decision output unit generates maintenance decision prompts. After the network converges, the target domain feature extractor and wear classifier can be deployed to the edge computing gateway installed at the processing equipment site. The edge computing gateway performs angular domain interpolation resampling, feature extraction, and confidence assessment on the target domain vibration signals collected in real time at the site, realizing lightweight deployment and online inference at the site.
[0065] V. Other Implementation Methods
[0066] In another embodiment of the present invention, the source domain feature extractor in the asymmetric feature extraction unit can also adopt a recurrent neural network structure such as a long short-term memory network or a gated recurrent unit to perform time-series modeling of the order domain feature sequence. It can also extract the feature information contained in the high-frequency data of the source domain and connect to the latent space alignment unit together with the target domain lightweight feature extractor to achieve cross-domain feature alignment.
[0067] In another embodiment of the present invention, the distribution difference measurement subunit in the latent space alignment unit can also use the correlation alignment method instead of the maximum mean difference method. Distribution alignment is achieved by minimizing the difference between the second-order statistics of the features of the source domain and the target domain. This method can also be combined with the gradient reversal layer adversarial loss to form the total loss function, thereby achieving feature alignment of heterogeneous data across sampling rates.
[0068] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.
Claims
1. A data analysis method for an intelligent monitoring device, characterized in that, Includes the following steps: The source domain vibration time-domain signal and the target domain vibration time-domain signal, as well as the rotation speed correlation signal related to the spindle rotation of the processing equipment, are acquired respectively. The source domain vibration time-domain signal is collected by a laboratory high-frequency piezoelectric accelerometer and has a wear level label. The target domain vibration time-domain signal is collected by a microelectromechanical system vibration sensor deployed in the industrial field and does not have a wear level label. The instantaneous angular velocity is estimated based on the rotational speed-related signal. The instantaneous angular velocity is integrated over time to obtain the cumulative rotation angle. An equal-angle sampling point sequence is generated according to a preset angle sampling interval. Using the mapping relationship between the cumulative rotation angle and the acquisition time, the source domain vibration time-domain signal and the target domain vibration time-domain signal are respectively subjected to angular domain interpolation and resampling to obtain the source domain equal-angle domain signal and the target domain equal-angle domain signal. Then, the source domain equal-angle domain signal and the target domain equal-angle domain signal are respectively subjected to discrete Fourier transform to obtain the source domain order features and target domain order features that are independent of rotational speed. The source domain order features are input into the source domain feature extractor for feature extraction to obtain the source domain latent features. The source domain feature extractor is a deep one-dimensional residual network containing multiple stacked one-dimensional residual blocks. The target domain order features are input into the target domain feature extractor for feature extraction to obtain the target domain latent features. The target domain feature extractor is a lightweight network using a depth-separable one-dimensional convolutional structure. The network structures of the source domain feature extractor and the target domain feature extractor are different. The source domain classification loss is calculated using the latent features of the source domain and the corresponding wear level label. The latent space distribution difference loss is calculated using the latent features of the source domain and the latent features of the target domain. The latent features of the source domain and the latent features of the target domain are then passed through a gradient inversion layer and input into a domain discriminator to calculate the domain adversarial loss. The source domain classification loss, the distribution difference loss, and the domain adversarial loss are weighted and summed to obtain the joint loss. The network parameters of the source domain feature extractor, the target domain feature extractor, the wear classifier, and the domain discriminator are updated by backpropagation based on the joint loss until the network converges. The latent features of the target domain obtained after network convergence are input into the trained wear classifier. The wear classifier is kept in an active state with random deactivation in the fully connected layer. Multiple independent forward propagations are performed to obtain multiple sets of prediction results. The mean of the multiple sets of prediction results is calculated as the wear state point estimate output. The variance of the multiple sets of prediction results is calculated as the cognitive uncertainty estimate. The wear state prediction confidence interval is constructed based on the mean and the cognitive uncertainty. The wear level prediction result and the confidence interval are output.
2. The method according to claim 1, characterized in that, During the process of angular domain interpolation resampling of the source domain vibration time-domain signal and the target domain vibration time-domain signal, it is determined whether the instantaneous angular velocity has a sudden change exceeding a preset threshold. If a sudden change occurs, the time of the sudden change is used as the segmentation node, and the signals before and after the sudden change are resampled by angular domain interpolation.
3. The method according to claim 1, characterized in that, The angular domain interpolation resampling is implemented using the cubic spline interpolation method. When performing discrete Fourier transform on the source domain equal angular domain signal and the target domain equal angular domain signal, the amplitude of the transformed spectrum is used as the order spectrum. The order in the order spectrum represents the number of harmonics per revolution relative to the principal axis.
4. The method according to claim 1, characterized in that, The receptive field of the convolution kernel in the deep one-dimensional residual network is larger than that of the convolution kernel in the target domain feature extractor. The target domain feature extractor adopts a depth-separable one-dimensional convolution structure to reduce the number of network parameters and computational cost, so as to adapt to the computing power constraints of the on-site edge computing hardware.
5. The method according to claim 1, characterized in that, The distribution difference loss is obtained by calculating the distance between the mean embeddings of the source domain latent features and the target domain latent features after they are mapped to the regenerating kernel Hilbert space.
6. The method according to claim 1, characterized in that, The gradient inversion layer keeps the input features unchanged during forward propagation, and multiplies the gradient by a negative coefficient before passing it to the source domain feature extractor and the target domain feature extractor during backward propagation, so as to realize adversarial training between the source domain feature extractor, the target domain feature extractor and the domain discriminator.
7. The method according to claim 1, characterized in that, After outputting the wear level prediction result and the confidence interval, it is determined whether the width of the confidence interval exceeds a preset threshold. If it does, a low confidence prompt message is output, and the wear level prediction result, the confidence interval, and the low confidence prompt message are transmitted to the predictive maintenance decision unit.
8. The method according to claim 1, characterized in that, After the network converges, the trained target domain feature extractor and wear classifier are deployed to the edge computing gateway installed at the processing equipment site. The edge computing gateway performs angular domain interpolation resampling, feature extraction, and confidence assessment on the target domain vibration time domain signal collected in real time at the site.
9. An intelligent monitoring device for implementing the method according to any one of claims 1 to 8, characterized in that, It includes a source domain data acquisition unit, a target domain data acquisition unit, an order resampling unit, an asymmetric feature extraction unit, a latent space alignment unit, and a Bayesian uncertainty assessment unit; The source domain data acquisition unit includes a laboratory high-frequency piezoelectric accelerometer sensor for acquiring time-domain signals of source domain vibration; The target domain data acquisition unit includes a microelectromechanical system vibration sensor installed on the bearing seat of the saw belt wheel of the processing equipment, which is used to acquire the time domain vibration signal of the target domain. The order resampling unit is connected to the source domain data acquisition unit and the target domain data acquisition unit respectively, and is used to perform angular domain interpolation resampling and discrete Fourier transform on the source domain vibration time domain signal and the target domain vibration time domain signal according to the rotational speed correlation signal to obtain the source domain order features and the target domain order features. The asymmetric feature extraction unit includes a source domain feature extractor and a target domain feature extractor with different structures. The source domain feature extractor is connected to the order resampling unit and is used to extract latent features in the source domain. The target domain feature extractor is connected to the order resampling unit and is used to extract latent features in the target domain. The latent space alignment unit is connected to the source domain feature extractor and the target domain feature extractor respectively, and is used to calculate the source domain classification loss, distribution difference loss and domain adversarial loss and construct a joint loss, and update the network parameters of the asymmetric feature extraction unit according to the joint loss; The Bayesian uncertainty assessment unit is connected to the target domain feature extractor and is used to perform multiple independent forward propagations on the aligned target domain latent features, outputting wear level prediction results and confidence intervals.
10. The intelligent monitoring device according to claim 9, characterized in that, It also includes a predictive maintenance decision output unit, which is connected to the Bayesian uncertainty assessment unit. The predictive maintenance decision output unit is used to receive the wear level prediction result and the confidence interval, and output a low confidence prompt message when the width of the confidence interval exceeds a preset threshold.
