Elevator transmission component cross-domain fault diagnosis method and system based on deep learning

Abstract

The application provides an elevator transmission component cross-domain fault diagnosis method and system based on deep learning, vibration signal data is collected by a sensor installed at an elevator transmission component, in the offline training stage of the model, the source domain and target domain vibration signals are divided by a physical time length T, in the division process, the window point number is determined by the product of the sampling rate and T; a one-dimensional convolutional neural network containing an adaptive global average pooling layer is used as the backbone; progressive parameter freezing migration is adopted, combined with a gradient reversal layer, a domain classifier and a center loss to realize sub-domain adaptive alignment, the source domain samples participate in the domain adversarial loss after gradient truncation; in the online diagnosis stage of the method, a comprehensive health index is obtained through multi-window time series smoothing, when the comprehensive health index and the confidence exceed the corresponding threshold, a speed reduction linkage instruction is sent to an elevator main control system through an edge computing device; the application can realize fault diagnosis and safety linkage closed loop in the cross-working-condition, cross-sampling-rate and target domain few-sample scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault diagnosis technology for special equipment, and in particular to a cross-domain fault diagnosis method and system for elevator transmission components based on deep learning, involving the application of deep learning and transfer learning. Background Technology

[0002] Elevators are among the special equipment explicitly listed in the National Special Equipment Safety Law. Elevators are driven by a traction machine via traction sheaves and wire ropes, which raise and lower the car and counterweight. Door operators, guide wheels, and other components facilitate the opening and closing of the car doors and guide the elevator's movement. These transmission components generally rely on rolling bearings for rotational support. The rolling bearings in these transmission components are subjected to alternating loads over long periods, and the main types of failures include outer ring failures, inner ring failures, and rolling element failures. Failure to diagnose these failures in a timely manner can affect the normal operation of the elevator and even lead to safety accidents. Therefore, implementing 24 / 7 online condition monitoring and fault diagnosis for elevator transmission components is of significant engineering importance. In recent years, deep learning and transfer learning methods have been extensively studied in the fault diagnosis of rotating machinery across operating conditions and equipment, and a series of related technical solutions have emerged both domestically and internationally.

[0003] Chinese patent CN108414226A discloses a method for diagnosing rolling bearing faults under varying operating conditions based on feature transfer learning, achieving cross-condition diagnosis by extracting common features between the source and target domains. Chinese patent CN110220709B discloses a bearing fault diagnosis method based on convolutional neural networks and transfer learning, employing a pre-training and fine-tuning strategy to achieve cross-domain diagnosis. Chinese patent CN109902393B discloses a method for diagnosing bearing faults under varying operating conditions based on deep features and transfer learning, combining deep feature extraction and transfer mechanisms to improve cross-condition diagnosis performance. When processing signals with different sampling rates in the source and target domains, these three patents typically use downsampling or resampling to unify the signals from both domains to the same sampling rate, and often employ global domain alignment strategies such as maximum mean difference to achieve cross-domain adaptation. Chinese patent CN117370851B discloses an adaptive input length bearing fault diagnosis method based on unsupervised transfer learning. The input length is selected based on the bearing's rotation frequency, number of rolling elements, rolling element diameter, pitch circle diameter, contact angle, and other geometric parameters. The analysis bandwidth is determined by calculating the fault characteristic frequency through envelope spectrum analysis.

[0004] In practical applications of elevator transmission components, the existing technology described above has the following three aspects that need improvement:

[0005] Firstly, in cross-sampling rate signal processing of elevator drive components, existing cross-domain migration methods suffer from the loss of high-frequency fault features or difficulty in determining the input length. The aforementioned downsampling or resampling-based schemes, when unifying the sampling rates of the two domains, will cause a certain loss of high-frequency fault impact components in the high-sampling rate data; the adaptive input length scheme needs to be determined with reference to the internal geometric parameters of the bearing, but the bearings of elevator drive components are usually encapsulated inside the traction machine, and it is difficult to disassemble and measure the internal geometric parameters on site, and the geometric parameter libraries of bearings from different manufacturers and models are not yet complete.

[0006] Secondly, in scenarios involving three-stage variable operating conditions in elevators and a limited number of samples in the target domain, existing cross-domain transfer methods have limited cross-domain adaptability. A single elevator operation cycle consists of acceleration, constant speed, and deceleration phases, with continuously changing operating conditions and non-stationary vibration signals. Furthermore, the number of labeled fault samples available in the target domain during actual elevator deployments is relatively limited. The aforementioned global domain alignment-based schemes need further improvement in cross-domain adaptability when faced with significant differences in subdomain distribution and scarce target domain samples.

[0007] Third, existing fault diagnosis methods and elevator main control systems have not yet formed an effective closed loop for safety linkage between monitoring, diagnosis, and handling. The output of the aforementioned existing solutions is mainly the fault category identification result, and the interaction with the elevator main control system is limited to the presentation of diagnostic information. When transmission components malfunction, there is a lack of a mechanism to actively output safety linkage commands with physical quantification constraints to the elevator control system.

[0008] The aforementioned engineering scenario characteristics place higher demands on cross-domain migration fault diagnosis methods in terms of cross-sampling rate signal processing, adaptation to varying operating conditions and limited samples, and safety linkage and coordination. Summary of the Invention

[0009] This invention proposes a cross-domain fault diagnosis method and system for elevator transmission components based on deep learning, which can realize fault diagnosis and safety linkage closed loop under cross-operating conditions, cross-sampling rates, and target domain low-sample scenarios.

[0010] The present invention adopts the following technical solution.

[0011] A deep learning-based cross-domain fault diagnosis method for elevator transmission components is proposed. This method uses vibration sensors installed on the bearing housings or end covers of the elevator transmission components to synchronously collect vibration signal data for training the diagnostic model via multiple channels. During the offline training phase of the diagnostic model, the signal data is segmented into source and target domain vibration signals by a physical duration T. During segmentation, the number of window points is determined by the product of the sampling rate and T, independent of bearing geometric parameters. A one-dimensional convolutional neural network with an adaptive global average pooling layer serves as the backbone, eliminating the need for resampling of variable-length signals. Progressive parameter freezing transfer is employed, combined with a gradient inversion layer, a domain classifier, and a center loss to achieve adaptive alignment of subdomains. Source domain samples are truncated by gradients and only participate in domain adversarial loss. In the online diagnosis phase, a comprehensive health index is obtained through multi-window temporal smoothing. When both the comprehensive health index and confidence level exceed the corresponding thresholds, a speed reduction linkage command is sent to the elevator main control system via an edge computing device. This invention utilizes synergistic strategies to avoid high-frequency feature loss, improve diagnostic accuracy, and suppress false alarm rates, making it suitable for cross-operating conditions, cross-sampling rates, and low-sample scenarios.

[0012] The diagnostic model uses a source domain baseline model and a transfer model running on a computing device. The diagnostic method includes steps S1 to S5 for the offline training phase of the model, and also includes step S6 for performing online diagnosis using the model; specifically:

[0013] Step S1, Source Domain Data Preprocessing: Vibration signals of transmission components with fault category labels are collected as source domain data. The source domain data is segmented using a sliding window based on physical duration T to obtain the source domain training sample set.

[0014] Step S2, pre-training of the source domain baseline model: Construct a diagnostic model with a convolutional neural network as the backbone network. The backbone network includes at least two convolutional blocks connected in series, an adaptive global average pooling layer, and a fully connected classifier. The adaptive global average pooling layer is used to process input signals of different lengths. Supervised training of the backbone network is performed on the source domain training sample set to obtain the weights of the source domain pre-trained model.

[0015] Step S3, target domain data preprocessing: collect the target domain vibration signal of the elevator transmission component under variable working conditions, and use the same physical duration T as in step S1 to perform sliding window segmentation to obtain the target domain training sample set;

[0016] Step S4, Progressive parameter freeze transfer: Load the weights of the source domain pre-trained model, freeze the parameters of the shallow convolutional blocks in the backbone network, unfreeze the parameters of the deep convolutional blocks and the fully connected classifier, and perform fine-tuning training on the target domain training sample set;

[0017] Step S5, Subdomain Adaptive Domain Alignment: During the fine-tuning training process in step S4, a gradient reversal layer and a domain classifier are introduced at the output of the adaptive global average pooling layer, and the center loss function is added to the training objective to form the total training loss function; in each training iteration, source domain samples and target domain samples are loaded simultaneously, and the features of the source domain samples are extracted by gradient truncation, which only participates in the calculation of domain adversarial loss and does not participate in the gradient backpropagation of classification loss and center loss;

[0018] Step S6, Online Diagnostic Inference and Speed ​​Reduction Linkage Command Output: The vibration signal of the elevator transmission component to be diagnosed is segmented as described in step S3 and input into the transfer model trained in steps S4 and S5, and the fault category diagnosis result is output; a comprehensive health index is calculated based on the time-series smoothing result of the fault category diagnosis probability of 3-10 consecutive sliding windows, and the confidence level is determined according to the consistency ratio of fault categories in the consecutive sliding windows; when the comprehensive health index is lower than the preset threshold and the confidence level is higher than the preset value, the computer sends the speed reduction linkage command to the elevator main control system through the communication interface.

[0019] The number of window points N for segmenting the selected domain of the source domain data in step S1. src Equal to the source domain sampling rate f ssrc The product of the physical duration T and the physical duration is rounded down.

[0020] The number of window points N for segmenting the target domain data in step S3. tgt Equal to the target domain sampling rate f stgt The product of the physical duration T and the physical duration is rounded down.

[0021] In steps S1 and S3, the segmented samples are subjected to dimensionless standardization.

[0022] In step S5, the total loss function L is trained. total The weighted sum of classification loss, domain adversarial loss, and centering loss is calculated using the following formula:

[0023]

[0024] Where L cls For classification cross-entropy loss based solely on samples from the target domain, L domain To achieve domain adversarial binary classification based on source and target domain samples, L center For the center loss based solely on samples from the target domain, λ domain With λ center These are weighting coefficients;

[0025] Step S5, subdomain adaptive domain alignment, consists of three loss functions working together. During training, both source domain sample batches and target domain sample batches are loaded simultaneously. The features of the source domain sample batches are extracted by the gradient truncation module and then input into the gradient inversion layer and the domain classifier. At the output of the adaptive global average pooling layer, three branches are generated: the first branch connects to the fully connected classifier and outputs the classification loss L. cls The second branch connects the gradient inversion layer to the domain classifier's output domain adversarial loss L. domain The third branch input center loss module outputs center loss L. center .

[0026] The three losses are weighted and summed by the training loss calculation module using the following formula:

[0027] L total = L cls + λ domain · L domain + λ center · L center

[0028] Where λ domain The value ranges from 0.05 to 0.2 (0.1 in this embodiment); λ center The value ranges from 0.005 to 0.05 (0.01 in this embodiment).

[0029] The specific calculation formulas for the three losses—classification loss, domain adversarial loss, and center loss—are as follows:

[0030]

[0031] Where N t y represents the number of training samples in the target domain, C represents the total number of fault categories (C=3 in this embodiment), and y i,c For one-hot real labels, p i,c To predict probabilities, L cls Calculations are based solely on samples from the target domain.

[0032]

[0033] Where N is the total number of samples involved in the calculation, d i For the binary label of the source domain (source domain is 0, target domain is 1), g(x) i ) is the eigenvector x i The output after gradient inversion layer processing, D(·), represents the probability that the domain classifier's output belongs to the target domain. The domain classifier consists of two fully connected layers (128→64, 64→1) and a Sigmoid activation layer;

[0034]

[0035] Where f(x) i Let be the feature vector of the i-th sample in the target domain, and c be the feature vector of the i-th sample. yi The category y to which this sample belongs i The learnable center vector is obtained. The center vector is updated using an independent auxiliary optimizer (learning rate 0.5), and the denominator in the update formula is m. y +1 to avoid division by zero anomalies. The initial value of the center vector is the mean of the feature vectors of each class in all training samples in the target domain;

[0036] The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a negative factor -α during backpropagation. p represents the training progress (p∈[0,1]), γ takes values ​​from 5 to 15 (10 in this embodiment); α increases smoothly from close to 0 at the beginning of training to close to 1 at the end according to an S-shaped curve.

[0037] Each training batch loads samples from both the source and target domains simultaneously: the features of the source domain samples, after being processed by the gradient truncation module, only participate in the L... domain Calculation, not involved in L cls With L center Gradient backpropagation; target domain samples participate in the calculation of three loss terms simultaneously; gradient truncation can be implemented by disabling the gradient calculation context provided by the deep learning framework (e.g., PyTorch's torch.no_grad or detach, TensorFlow's tf.stop_gradient), and this invention does not limit the specific implementation method;

[0038] The three strategies of gradient truncation, progressive parameter freezing, and subdomain adaptive domain alignment work together: progressive parameter freezing provides a stable source domain underlying feature base for subdomain alignment; source domain gradient truncation makes the classification loss and center loss focus on the target domain; and domain adversarial loss and center loss work together to achieve fine alignment at the class level.

[0039] The physical duration T mentioned in steps S1 and S3 ranges from 30 milliseconds to 300 milliseconds; when the source domain sampling rate f ssrc With the target domain sampling rate f stgt When the features are unequal, the adaptive global average pooling layer compresses convolutional feature maps of different lengths into fixed-length feature vectors of the same dimension; the source domain sampling rate f ssrc With the target domain sampling rate f stgt The value range is from 10 kHz to 200 kHz.

[0040] In step S2, the kernel size of each convolutional block decreases sequentially along the signal flow direction. The kernel size of each subsequent convolutional block is smaller than that of the previous convolutional block, and the number of output channels is not less than twice the number of output channels of the previous convolutional block. Each convolutional block contains a convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer. Each convolutional block is followed by a squeezed excitation channel attention module. In step S4, when freezing the parameters of the shallow convolutional block, the parameters of the squeezed excitation channel attention module following the shallow convolutional block are frozen simultaneously.

[0041] In step S6, the time-series smoothing adopts the exponential moving average method. The comprehensive health index is the time-series smoothed value of the normal category diagnosis probability. The confidence level is the proportion of windows with the same fault category in multiple consecutive sliding windows. The number of multiple consecutive sliding windows ranges from 3 to 10. The preset threshold of the comprehensive health index ranges from 0.5 to 0.7. The deceleration corresponding to the deceleration action indicated by the deceleration linkage command does not exceed 0.1 m / s², and the target running speed after deceleration is not less than 40% of the elevator's rated speed.

[0042] like Figure 4 As shown, the progressive parameter freeze transfer strategy in step S4 is as follows: After loading the source domain pre-trained model weights obtained in step S2, the parameters of the first convolutional block and its subsequent first squeeze excitation channel attention module are frozen through the parameter freeze module, while the parameters of the remaining convolutional blocks, attention modules, and fully connected classifiers are unfrozen through the parameter unfreeze module. The basic impact pulse and envelope features learned by shallow modules have cross-domain universality and are suitable for freezing and preserving source domain knowledge; the fine frequency features learned by deep modules are sensitive to operating conditions and equipment and need to be re-adapted in the target domain. Fine-tuning uses the Adam optimizer with an initial learning rate of 0.0001.

[0043] In step S3, the variable operating condition signals consisting of acceleration, constant speed and deceleration phases in a single elevator operation cycle are uniformly covered by a physical duration sliding window, and the transfer model directly outputs the fault category diagnosis results for each signal segment.

[0044] The speed reduction linkage command in step S6 includes verification information for the elevator main control system to perform safety verification. The verification information includes at least one of the comprehensive health index, the confidence level, the fault category diagnosis result, and the command timestamp. Step S6 further includes a data reliability judgment sub-step: when the signal-to-noise ratio of the vibration signals output by each sensor node is lower than a preset lower limit, the current data is determined to be unreliable and a retest process is triggered. The preset lower limit ranges from 5 dB to 10 dB. Specifically:

[0045] After the weights of the transfer model trained offline are distributed to the edge computing hardware platform, the online phase process in step S6 is as follows: The vibration signal acquisition module acquires the vibration signal to be diagnosed from the elevator transmission component, and the sliding window segmentation module segments the sample to be diagnosed according to the physical duration T=213 milliseconds. The sample is then input into the transfer model inference module to output the classification probability of various faults; the time-series smoothing module performs smoothing based on the diagnostic probability of N consecutive sliding windows (N is 3 to 10, and 5 in this embodiment); the comprehensive health index calculation module calculates the comprehensive health index H according to the exponential moving average formula.

[0046]

[0047] in Let H0 be the normal category diagnosis probability of the t-th window, β be the smoothing coefficient (ranging from 0.1 to 0.5, and 0.3 in this embodiment), and H0 be initially 1.0. The closer the comprehensive health index H is to 1, the better the health status. The confidence level is the proportion of windows that output the same fault category in N consecutive sliding windows.

[0048] When the dual-threshold judgment module determines that the comprehensive health index H is lower than the preset threshold (ranging from 0.5 to 0.7, and 0.6 in this embodiment) and the confidence level is higher than the preset value (ranging from 70% to 90%, and 80% in this embodiment), the deceleration linkage command generation module generates a deceleration linkage command and sends it to the elevator main control system via the communication interface. The command includes verification information for safety verification, including at least one of the following: comprehensive health index H, confidence level, fault type, and command timestamp. In this embodiment, the deceleration indicated by the deceleration linkage command does not exceed 0.1 m / s², and the target running speed after deceleration is not lower than 40% of the rated speed. The above upper and lower limits are determined with reference to the relevant engineering practices of GB / T 24474.1—2009 "Measurement of Passenger Quality - Part 1: Elevators" and GB / T 7588.1—2020 "Safety Specifications for Elevator Manufacturing and Installation - Part 1: Passenger Elevators and Freight Elevators". When the signal-to-noise ratio of the vibration signal at each sensing node is lower than the preset lower limit (5 dB to 10 dB), the data is deemed unreliable and a retest process is triggered.

[0049] The variable operating conditions of a single elevator operation cycle consist of an acceleration phase, a constant speed phase, and a deceleration phase. The method of this invention does not divide the three-stage variable operating condition signal into operating stages. The diagnostic method directly uses a sliding window of the same physical duration to cover the signal segments of each stage, and the migration model directly outputs the fault category diagnosis results for each window segment.

[0050] In step S1, under the standard constant operating conditions of the elevator, the speed fluctuation does not exceed ±5% of the rated speed and the load fluctuation does not exceed ±10% of the rated load.

[0051] The deep learning-based cross-domain fault diagnosis system for elevator drive components uses the deep learning-based cross-domain fault diagnosis method for elevator drive components described above. The diagnosis system includes a processor and a memory; or it includes a processor, a memory, and at least one of a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), and a graphics processor (GPU). The processor and the memory, or the processor, the memory, and at least one of the FPGA, ASIC, and GPU, are configured to implement step S6 of the deep learning-based cross-domain fault diagnosis method for elevator drive components described above. The memory stores a computer program, which, when executed by the processor, implements the steps of the deep learning-based cross-domain fault diagnosis method for elevator drive components described above.

[0052] The diagnostic system also includes an edge computing device and an elevator main control system. The output of the edge computing device is connected to the input of the elevator main control system via a wired or wireless communication interface. The communication interface is configured inside the edge computing device or is set independently between the edge computing device and the elevator main control system. The edge computing device outputs a speed reduction linkage command to the elevator main control system. The speed reduction linkage command is used to instruct the elevator main control system to perform a speed reduction action on the elevator.

[0053] The elevator main control system controls the elevator transmission components. In the elevator transmission components, the traction machine drives the elevator car and counterweight to rise and fall via the traction sheave and wire rope, and the door operator, guide wheel and other components realize the opening and closing of the elevator car door and the running guidance. The rotating parts of the elevator transmission components are supported by rolling bearings to rotate.

[0054] The diagnostic system uses vibration sensors installed in the bearing housings or end caps of elevator drive components to synchronously acquire vibration signal data for training the diagnostic model via multiple channels.

[0055] The vibration signal is acquired by installing a vibration sensor on the bearing housing, end cover, or a nearby location that is rigidly connected to the bearing housing of the monitored elevator transmission component. The vibration sensor is at least one of a piezoelectric accelerometer, a MEMS accelerometer, or an eddy current displacement sensor. Piezoelectric accelerometer is preferred to retain high-frequency fault impact components.

[0056] The sensitive axis of the vibration sensor is arranged along the radial or axial direction of the bearing being measured; when monitoring a single transmission component, a single sensor is arranged to form a single sensing node.

[0057] When it is necessary to simultaneously monitor multiple transmission components such as traction machine bearings, gantry crane bearings, and guide wheel bearings, or to monitor the same component in multiple directions, multiple vibration sensors are arranged to form multiple sensing nodes. Each sensing node collects vibration signals at the same time reference through a multi-channel synchronous acquisition circuit.

[0058] After the vibration signal is output by the sensor, it is amplified and anti-aliasing filtered by the signal conditioning circuit, and then discretized into a digital vibration signal by the analog-to-digital converter at a preset sampling rate. The sampling rate is set according to the frequency range of the fault characteristics to be retained, ranging from 10 kHz to 200 kHz. The source domain data and the target domain data use unequal sampling rates, and the input signals of different lengths generated by different sampling rates are processed by an adaptive global average pooling layer. This invention does not limit the specific model of the vibration sensor, the installation and fastening method, or the specific implementation of the analog-to-digital converter. Any acquisition method that can obtain the vibration signal of the elevator transmission component within the above sampling rate range is applicable to this invention.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) This invention overcomes the problems of high-frequency fault feature loss or input length dependence on bearing geometric parameters in the cross-sampling rate signal processing of elevator transmission components by existing cross-domain migration methods, and can achieve lossless cross-domain adaptation of source and target domain signals. This invention proposes a cross-sampling rate signal alignment method based on physical duration T, combined with an adaptive global average pooling layer. The same backbone network can process signals with different sampling rates in the source and target domains without downsampling, resampling and interpolation processing, so that the high-frequency fault impact components in high sampling rate data can be completely preserved. The physical duration segmentation only depends on the product of sampling rate and physical duration to determine the number of window points, and does not depend on the geometric parameters of bearing such as rolling element diameter, number of rolling elements, pitch circle diameter, contact angle and fault feature frequency. It is suitable for elevator transmission components where the bearing is encapsulated inside the traction machine and it is difficult to disassemble and measure the internal geometric parameters on site.

[0061] (2) Improved the cross-domain adaptation capability of elevators in three-stage variable operating conditions and target domain with few samples, and suppressed cross-domain negative transfer. This invention proposes a synergistic combination scheme of source domain gradient truncation, progressive parameter freezing transfer, and subdomain adaptive domain alignment at the cross-domain transfer learning algorithm level: source domain samples are only used for domain adversarial loss calculation after gradient truncation, avoiding the discriminative bias of labeled samples in the source domain to the classifier and polluting the target domain classification boundary; the progressive parameter freezing strategy achieves a balance between the preservation of general features in the source domain and the adaptation of specific features in the target domain; the gradient inversion layer and the domain classifier work together with the central loss to form a refined alignment of subdomain distribution differences. The three work together: progressive parameter freezing provides a stable source domain low-level feature base for subdomain alignment, and gradient truncation makes the discriminative learning of subdomain alignment focus on the target domain, which is particularly significant in target domain with few samples. In addition, the variable operating condition signal consisting of acceleration, constant speed and deceleration in a single elevator operation cycle does not need to be segmented into operation stages. The transfer model directly outputs the fault category diagnosis results for each physical duration window segment.

[0062] (3) This invention fills the gap in the existing fault diagnosis method and the elevator main control system's safety linkage and coordination mechanism, and constructs a complete safety linkage closed loop for monitoring, diagnosis and handling. In the online diagnosis stage, this invention calculates a comprehensive health index based on the time-series smoothing results of multi-window fault category diagnosis probability. When the comprehensive health index is lower than a preset threshold and the confidence level is higher than a preset value, the computer sends a speed reduction linkage command to the elevator main control system through the communication interface. The comprehensive health index and confidence level triggering mechanism can suppress the false alarm rate, enabling the method of this invention to further form an active safety linkage command output capability on the basis of outputting fault category identification results, and connecting the elevator transmission component status monitoring and the safety control behavior of the elevator main control system into a closed-loop management and control link. Attached Figure Description

[0063] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0064] Appendix Figure 1 This is a schematic diagram of the overall process of the deep learning-based cross-domain fault diagnosis method for elevator transmission components provided in an embodiment of the present invention;

[0065] Appendix Figure 2 A schematic diagram of a one-dimensional convolutional neural network backbone network structure provided in an embodiment of the present invention;

[0066] Appendix Figure 3 A schematic diagram illustrating the principle of physical duration-aligned cross-sampling rate signal preprocessing provided in this embodiment of the invention;

[0067] Appendix Figure 4 This is a schematic diagram of the transfer learning training process provided in an embodiment of the present invention;

[0068] Appendix Figure 5 This is a bar chart showing the comparison between cross-domain diagnostic accuracy and F1 score of various transfer learning strategies provided in the embodiments of the present invention;

[0069] Appendix Figure 6 This is a schematic diagram comparing the confusion matrices of the target domain test sets for each transfer learning strategy provided in this embodiment of the invention.

[0070] Appendix Figure 7 This is a two-dimensional visualization comparison diagram of the t-SNE of the target domain test set features of various transfer learning strategies provided in the embodiments of the present invention;

[0071] Appendix Figure 8 A bar chart illustrating the diagnostic accuracy of the method of the present invention across three independent cross-domain migration paths, as provided in this embodiment of the invention.

[0072] Appendix Figure 9 A horizontal comparison bar chart of the method of the present invention and the publicly disclosed baseline transfer learning method provided for embodiments of the present invention;

[0073] Appendix Figure 10 This is a schematic diagram of the diagnostic accuracy gain comparison curve of the subdomain adaptive module in a low-sample scenario provided in an embodiment of the present invention.

[0074] Appendix Figure 11 A bar chart showing the ablation comparison between the physical duration segmentation strategy provided in this embodiment of the invention and the fixed sampling point segmentation and resampling segmentation baselines;

[0075] Appendix Figure 12 A bar chart showing the comparison between the source domain sample gradient truncation strategy and the ablation of the source domain participating in the entire loss baseline provided in the embodiments of the present invention;

[0076] Appendix Figure 13 This is a bar chart comparing the false alarm rate and missed alarm rate of comprehensive health index and confidence level triggering versus single threshold triggering, as provided in an embodiment of the present invention. Detailed Implementation

[0077] As shown in the figure, a deep learning-based cross-domain fault diagnosis method for elevator transmission components is presented. This method uses vibration sensors installed on the bearing housings or end covers of the elevator transmission components to synchronously collect vibration signal data for training the diagnostic model via multiple channels. During the offline training phase of the diagnostic model, the signal data is segmented into source and target domain vibration signals by a physical duration T. During segmentation, the number of window points is determined by the product of the sampling rate and T, independent of bearing geometric parameters. A one-dimensional convolutional neural network with an adaptive global average pooling layer serves as the backbone, eliminating the need for resampling of variable-length signals. A progressive parameter freezing transfer algorithm, combined with a gradient inversion layer, domain classifier, and center loss, achieves adaptive alignment of subdomains. Source domain samples are truncated by gradients and only participate in domain adversarial loss. In the online diagnosis phase, a comprehensive health index is obtained through multi-window temporal smoothing. When both the comprehensive health index and confidence level exceed the corresponding thresholds, a speed reduction linkage command is sent to the elevator main control system via an edge computing device. This invention utilizes synergistic strategies to avoid high-frequency feature loss, improve diagnostic accuracy, and suppress false alarm rates, making it suitable for scenarios involving multiple operating conditions, different sampling rates, and few samples.

[0078] The diagnostic model uses a source domain baseline model and a transfer model running on a computing device. The diagnostic method includes steps S1 to S5 for the offline training phase of the model, and also includes step S6 for performing online diagnosis using the model; specifically:

[0079] Step S1, Source Domain Data Preprocessing: Vibration signals of transmission components with fault category labels are collected as source domain data. The source domain data is segmented using a sliding window based on physical duration T to obtain the source domain training sample set.

[0080] Step S2, pre-training of the source domain baseline model: Construct a diagnostic model with a convolutional neural network as the backbone network. The backbone network includes at least two convolutional blocks connected in series, an adaptive global average pooling layer, and a fully connected classifier. The adaptive global average pooling layer is used to process input signals of different lengths. Supervised training of the backbone network is performed on the source domain training sample set to obtain the weights of the source domain pre-trained model.

[0081] Step S3, target domain data preprocessing: collect the target domain vibration signal of the elevator transmission component under variable working conditions, and use the same physical duration T as in step S1 to perform sliding window segmentation to obtain the target domain training sample set;

[0082] Step S4, Progressive parameter freeze transfer: Load the weights of the source domain pre-trained model, freeze the parameters of the shallow convolutional blocks in the backbone network, unfreeze the parameters of the deep convolutional blocks and the fully connected classifier, and perform fine-tuning training on the target domain training sample set;

[0083] Step S5, Subdomain Adaptive Domain Alignment: During the fine-tuning training process in step S4, a gradient reversal layer and a domain classifier are introduced at the output of the adaptive global average pooling layer, and the center loss function is added to the training objective to form the total training loss function; in each training iteration, source domain samples and target domain samples are loaded simultaneously, and the features of the source domain samples are extracted by gradient truncation, which only participates in the calculation of domain adversarial loss and does not participate in the gradient backpropagation of classification loss and center loss;

[0084] Step S6, Online Diagnostic Inference and Speed ​​Reduction Linkage Command Output: The vibration signal of the elevator transmission component to be diagnosed is segmented as described in step S3 and input into the transfer model trained in steps S4 and S5, and the fault category diagnosis result is output; a comprehensive health index is calculated based on the time-series smoothing result of the fault category diagnosis probability of 3-10 consecutive sliding windows, and the confidence level is determined according to the consistency ratio of fault categories in the consecutive sliding windows; when the comprehensive health index is lower than the preset threshold and the confidence level is higher than the preset value, the computer sends the speed reduction linkage command to the elevator main control system through the communication interface.

[0085] The number of window points N for segmenting the selected domain of the source domain data in step S1. src Equal to the source domain sampling rate f ssrc The product of the physical duration T and the physical duration is rounded down.

[0086] The number of window points N for segmenting the target domain data in step S3. tgt Equal to the target domain sampling rate f stgt The product of the physical duration T and the physical duration is rounded down.

[0087] In steps S1 and S3, the segmented samples are subjected to dimensionless standardization.

[0088] In step S5, the total loss function L is trained. total The weighted sum of classification loss, domain adversarial loss, and centering loss is calculated using the following formula:

[0089]

[0090] Where L cls For classification cross-entropy loss based solely on samples from the target domain, L domain To achieve domain adversarial binary classification based on source and target domain samples, L center For the center loss based solely on samples from the target domain, λ domain With λ center These are weighting coefficients;

[0091] Step S5, subdomain adaptive domain alignment, consists of three loss functions working together. During training, both source domain sample batches and target domain sample batches are loaded simultaneously. The features of the source domain sample batches are extracted by the gradient truncation module and then input into the gradient inversion layer and the domain classifier. At the output of the adaptive global average pooling layer, three branches are generated: the first branch connects to the fully connected classifier and outputs the classification loss L. cls The second branch connects the gradient inversion layer to the domain classifier's output domain adversarial loss L. domain The third branch input center loss module outputs center loss L. center .

[0092] The three losses are weighted and summed by the training loss calculation module using the following formula:

[0093] L total = L cls + λ domain · L domain + λ center · L center

[0094] Where λ domain The value ranges from 0.05 to 0.2 (0.1 in this embodiment); λ center The value ranges from 0.005 to 0.05 (0.01 in this embodiment).

[0095] The specific calculation formulas for the three losses—classification loss, domain adversarial loss, and center loss—are as follows:

[0096]

[0097] Where N t y represents the number of training samples in the target domain, C represents the total number of fault categories (C=3 in this embodiment), and y i,c For one-hot real labels, p i,c To predict probabilities, Lcls Calculations are based solely on samples from the target domain.

[0098]

[0099] Where N is the total number of samples involved in the calculation, d i For the binary label of the source domain (source domain is 0, target domain is 1), g(x) i ) is the eigenvector x i The output after gradient inversion layer processing, D(·), represents the probability that the domain classifier's output belongs to the target domain. The domain classifier consists of two fully connected layers (128→64, 64→1) and a Sigmoid activation layer;

[0100]

[0101] Where f(x) i Let be the feature vector of the i-th sample in the target domain, and c be the feature vector of the i-th sample. yi The category y to which this sample belongs i The learnable center vector is obtained. The center vector is updated using an independent auxiliary optimizer (learning rate 0.5), and the denominator in the update formula is m. y +1 to avoid division by zero anomalies. The initial value of the center vector is the mean of the feature vectors of each class in all training samples in the target domain;

[0102] The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a negative factor -α during backpropagation. p represents the training progress (p∈[0,1]), γ takes values ​​from 5 to 15 (10 in this embodiment); α increases smoothly from close to 0 at the beginning of training to close to 1 at the end according to an S-shaped curve.

[0103] Each training batch loads samples from both the source and target domains simultaneously: the features of the source domain samples, after being processed by the gradient truncation module, only participate in the L... domain Calculation, not involved in L cls With L center Gradient backpropagation; target domain samples participate in the calculation of three loss terms simultaneously; gradient truncation can be implemented by disabling the gradient calculation context provided by the deep learning framework (e.g., PyTorch's torch.no_grad or detach, TensorFlow's tf.stop_gradient), and this invention does not limit the specific implementation method;

[0104] The three strategies of gradient truncation, progressive parameter freezing, and subdomain adaptive domain alignment work together: progressive parameter freezing provides a stable source domain underlying feature base for subdomain alignment; source domain gradient truncation makes the classification loss and center loss focus on the target domain; and domain adversarial loss and center loss work together to achieve fine alignment at the class level.

[0105] The physical duration T mentioned in steps S1 and S3 ranges from 30 milliseconds to 300 milliseconds; when the source domain sampling rate f ssrc With the target domain sampling rate f stgt When the features are unequal, the adaptive global average pooling layer compresses convolutional feature maps of different lengths into fixed-length feature vectors of the same dimension; the source domain sampling rate f ssrc With the target domain sampling rate f stgt The value range is from 10 kHz to 200 kHz.

[0106] In step S2, the kernel size of each convolutional block decreases sequentially along the signal flow direction. The kernel size of each subsequent convolutional block is smaller than that of the previous convolutional block, and the number of output channels is not less than twice the number of output channels of the previous convolutional block. Each convolutional block contains a convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer. Each convolutional block is followed by a squeezed excitation channel attention module. In step S4, when freezing the parameters of the shallow convolutional block, the parameters of the squeezed excitation channel attention module following the shallow convolutional block are frozen simultaneously.

[0107] In step S6, the time-series smoothing adopts the exponential moving average method. The comprehensive health index is the time-series smoothed value of the normal category diagnosis probability. The confidence level is the proportion of windows with the same fault category in multiple consecutive sliding windows. The number of multiple consecutive sliding windows ranges from 3 to 10. The preset threshold of the comprehensive health index ranges from 0.5 to 0.7. The deceleration corresponding to the deceleration action indicated by the deceleration linkage command does not exceed 0.1 m / s², and the target running speed after deceleration is not less than 40% of the elevator's rated speed.

[0108] like Figure 4 As shown, the progressive parameter freeze transfer strategy in step S4 is as follows: After loading the source domain pre-trained model weights obtained in step S2, the parameters of the first convolutional block and its subsequent first squeeze excitation channel attention module are frozen through the parameter freeze module, while the parameters of the remaining convolutional blocks, attention modules, and fully connected classifiers are unfrozen through the parameter unfreeze module. The basic impact pulse and envelope features learned by shallow modules have cross-domain universality and are suitable for freezing and preserving source domain knowledge; the fine frequency features learned by deep modules are sensitive to operating conditions and equipment and need to be re-adapted in the target domain. Fine-tuning uses the Adam optimizer with an initial learning rate of 0.0001.

[0109] In step S3, the variable operating condition signals consisting of acceleration, constant speed and deceleration phases in a single elevator operation cycle are uniformly covered by a physical duration sliding window, and the transfer model directly outputs the fault category diagnosis results for each signal segment.

[0110] The speed reduction linkage command in step S6 includes verification information for the elevator main control system to perform safety verification. The verification information includes at least one of the comprehensive health index, the confidence level, the fault category diagnosis result, and the command timestamp. Step S6 further includes a data reliability judgment sub-step: when the signal-to-noise ratio of the vibration signals output by each sensor node is lower than a preset lower limit, the current data is determined to be unreliable and a retest process is triggered. The preset lower limit ranges from 5 dB to 10 dB. Specifically:

[0111] After the weights of the transfer model trained offline are distributed to the edge computing hardware platform, the online phase process in step S6 is as follows: The vibration signal acquisition module acquires the vibration signal to be diagnosed from the elevator transmission component, and the sliding window segmentation module segments the sample to be diagnosed according to the physical duration T=213 milliseconds. The sample is then input into the transfer model inference module to output the classification probability of various faults; the time-series smoothing module performs smoothing based on the diagnostic probability of N consecutive sliding windows (N is 3 to 10, and 5 in this embodiment); the comprehensive health index calculation module calculates the comprehensive health index H according to the exponential moving average formula.

[0112]

[0113] in Let H0 be the normal category diagnosis probability of the t-th window, β be the smoothing coefficient (ranging from 0.1 to 0.5, and 0.3 in this embodiment), and H0 be initially 1.0. The closer the comprehensive health index H is to 1, the better the health status. The confidence level is the proportion of windows that output the same fault category in N consecutive sliding windows.

[0114] When the dual-threshold judgment module determines that the comprehensive health index H is lower than the preset threshold (ranging from 0.5 to 0.7, and 0.6 in this embodiment) and the confidence level is higher than the preset value (ranging from 70% to 90%, and 80% in this embodiment), the deceleration linkage command generation module generates a deceleration linkage command and sends it to the elevator main control system via the communication interface. The command includes verification information for safety verification, including at least one of the following: comprehensive health index H, confidence level, fault type, and command timestamp. In this embodiment, the deceleration indicated by the deceleration linkage command does not exceed 0.1 m / s², and the target running speed after deceleration is not lower than 40% of the rated speed. The above upper and lower limits are determined with reference to the relevant engineering practices of GB / T 24474.1—2009 "Measurement of Passenger Quality - Part 1: Elevators" and GB / T 7588.1—2020 "Safety Specifications for Elevator Manufacturing and Installation - Part 1: Passenger Elevators and Freight Elevators". When the signal-to-noise ratio of the vibration signal at each sensing node is lower than the preset lower limit (5 dB to 10 dB), the data is deemed unreliable and a retest process is triggered.

[0115] The variable operating conditions of a single elevator operation cycle consist of an acceleration phase, a constant speed phase, and a deceleration phase. The method of this invention does not divide the three-stage variable operating condition signal into operating stages. The diagnostic method directly uses a sliding window of the same physical duration to cover the signal segments of each stage, and the migration model directly outputs the fault category diagnosis results for each window segment.

[0116] In step S1, under the standard constant operating conditions of the elevator, the speed fluctuation does not exceed ±5% of the rated speed and the load fluctuation does not exceed ±10% of the rated load.

[0117] The deep learning-based cross-domain fault diagnosis system for elevator drive components uses the deep learning-based cross-domain fault diagnosis method for elevator drive components described above. The diagnosis system includes a processor and a memory; or it includes a processor, a memory, and at least one of a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), and a graphics processor (GPU). The processor and the memory, or the processor, the memory, and at least one of the FPGA, ASIC, and GPU, are configured to implement step S6 of the deep learning-based cross-domain fault diagnosis method for elevator drive components described above. The memory stores a computer program, which, when executed by the processor, implements the steps of the deep learning-based cross-domain fault diagnosis method for elevator drive components described above.

[0118] The diagnostic system also includes an edge computing device and an elevator main control system. The output of the edge computing device is connected to the input of the elevator main control system via a wired or wireless communication interface. The communication interface is configured inside the edge computing device or is set independently between the edge computing device and the elevator main control system. The edge computing device outputs a speed reduction linkage command to the elevator main control system. The speed reduction linkage command is used to instruct the elevator main control system to perform a speed reduction action on the elevator.

[0119] The elevator main control system controls the elevator transmission components. In the elevator transmission components, the traction machine drives the elevator car and counterweight to rise and fall via the traction sheave and wire rope, and the door operator, guide wheel and other components realize the opening and closing of the elevator car door and the running guidance. The rotating parts of the elevator transmission components are supported by rolling bearings to rotate.

[0120] The diagnostic system uses vibration sensors installed in the bearing housings or end caps of elevator drive components to synchronously acquire vibration signal data for training the diagnostic model via multiple channels.

[0121] The vibration signal is acquired by installing a vibration sensor on the bearing housing, end cover, or a nearby location that is rigidly connected to the bearing housing of the monitored elevator transmission component. The vibration sensor is at least one of a piezoelectric accelerometer, a MEMS accelerometer, or an eddy current displacement sensor. Piezoelectric accelerometer is preferred to retain high-frequency fault impact components.

[0122] The sensitive axis of the vibration sensor is arranged along the radial or axial direction of the bearing being measured; when monitoring a single transmission component, a single sensor is arranged to form a single sensing node.

[0123] When it is necessary to simultaneously monitor multiple transmission components such as traction machine bearings, gantry crane bearings, and guide wheel bearings, or to monitor the same component in multiple directions, multiple vibration sensors are arranged to form multiple sensing nodes. Each sensing node collects vibration signals at the same time reference through a multi-channel synchronous acquisition circuit.

[0124] After the vibration signal is output by the sensor, it is amplified and anti-aliasing filtered by the signal conditioning circuit, and then discretized into a digital vibration signal by the analog-to-digital converter at a preset sampling rate. The sampling rate is set according to the frequency range of the fault characteristics to be retained, ranging from 10 kHz to 200 kHz. The source domain data and the target domain data use unequal sampling rates, and the input signals of different lengths generated by different sampling rates are processed by an adaptive global average pooling layer. This invention does not limit the specific model of the vibration sensor, the installation and fastening method, or the specific implementation of the analog-to-digital converter. Any acquisition method that can obtain the vibration signal of the elevator transmission component within the above sampling rate range is applicable to this invention.

[0125] Example:

[0126] The elevator transmission components mentioned in this example refer to rotating components that participate in power transmission or motion guidance during elevator operation, are subjected to alternating loads for a long time, and whose internal rolling bearings are prone to fatigue failure, including but not limited to at least one of traction machine bearings, door machine bearings, and guide wheel bearings.

[0127] Taking an elevator traction system as an example, the traction machine drives the traction sheave to rotate via an electric motor. The traction sheave is connected to the car and counterweight via a wire rope through a friction transmission connection. The traction motor bearing and gearbox bearing (or the motor spindle bearing in a gearless traction machine) inside the traction machine bear the alternating loads of the motor torque and the car-counterweight system. In the door operator system, the door operator motor drives the door knife and door panel to open and close via a transmission belt or lead screw. The door operator bearing bears the reciprocating start and stop loads of the door body. In the guide system, the guide wheel is mounted on the car or counterweight frame via an axle and rolls along the guide rail. The guide wheel bearing bears the lateral guide loads. When the rolling bearings of the above-mentioned transmission components experience fatigue spalling, pitting, or other faults in the outer ring, inner ring, or rolling elements, they will all produce periodic impact characteristics in the vibration signal. The method of this invention is applicable to all of the above-mentioned components. This embodiment uses the traction machine bearing as the preferred diagnostic object for explanation.

[0128] In terms of specifications, the types of rolling bearings used in elevator transmission components include deep groove ball bearings, cylindrical roller bearings, and angular contact ball bearings, with common inner diameter specifications ranging from 20 mm to 120 mm, corresponding to bearing operating speeds of approximately 50 rpm to 2000 rpm. The method of this invention does not rely on the specific specifications and bearing internal geometric parameters mentioned above. The above specifications are only used to illustrate typical applications of this invention and do not constitute a limitation on the scope of protection of this invention.

[0129] The vibration signal acquisition method is as follows: a vibration sensor is installed at the bearing housing, end cover, or a nearby location rigidly connected to the bearing housing of the monitored elevator transmission component. The vibration sensor is at least one of a piezoelectric accelerometer, a MEMS accelerometer, or an eddy current displacement sensor, with a piezoelectric accelerometer preferred to retain high-frequency fault impact components. The sensitive axis of the sensor is arranged radially or axially along the bearing being measured. When monitoring a single transmission component, a single sensor can be arranged to form a single sensing node. When multiple transmission components such as traction machine bearings, gantry machine bearings, and guide wheel bearings need to be monitored simultaneously, or when the same component needs to be monitored in multiple directions, multiple sensors can be arranged to form multiple sensing nodes. Each sensing node acquires vibration signals at the same time reference through a multi-channel synchronous acquisition circuit.

[0130] After the vibration signal is output by the sensor, it is amplified and anti-aliasing filtered by the signal conditioning circuit, and then discretized into a digital vibration signal by the analog-to-digital converter at a preset sampling rate. The sampling rate is set according to the frequency range of the fault characteristics to be retained, ranging from 10 kHz to 200 kHz. The source domain data and the target domain data can use different sampling rates. The method of this invention processes the input signals of different lengths generated by different sampling rates through an adaptive global average pooling layer. This invention does not limit the specific model of the vibration sensor, the installation and fastening method, or the specific implementation of the analog-to-digital converter. Any acquisition method that can obtain the vibration signal of the elevator transmission component within the above sampling rate range is applicable to this invention.

[0131] In this embodiment, the Case Western Reserve University (CWRU) rolling bearing dataset is used as the surrogate source domain, which is a standard constant condition dataset. The CWRU dataset contains four categories: normal, inner race fault, outer race fault, and rolling element fault. This embodiment selects the first three categories.

[0132] The source domain data preprocessing parameters in step S1 are: sliding window physical duration T = 213 milliseconds, source domain sampling rate f ssrc =12 kHz, N source domain window points src =int(f ssrc×T)=int(12000×0.213)=2556 points (int(·) represents rounding down), the sliding step size is 50% of the window length overlap ratio; after segmentation, 3960 samples are obtained (1320 per class), and class equalization is performed by oversampling; Z-score standardization is performed independently on each window sample. norm =(x-μ) / σ; The data is randomly partitioned into categories based on 80% training set, 10% validation set, and 10% test set. The method for obtaining labels for the source domain data is determined by the annotation system of the dataset itself, and this invention does not limit this.

[0133] like Figure 2 As shown, the backbone network in this embodiment consists of a first convolutional block, a first squeezed excitation channel attention module, a second convolutional block, a second squeezed excitation channel attention module, a third convolutional block, a third squeezed excitation channel attention module, an adaptive global average pooling layer, and a fully connected classifier composed of a first fully connected layer, a random deactivation layer, a second fully connected layer, and a normalized exponential function layer, connected in series. The squeezed excitation channel attention module is a preferred feature; in a simplified implementation without this module, the backbone network only includes convolutional blocks, an adaptive global average pooling layer, and a fully connected classifier, and the core inventive contribution of the present invention still holds.

[0134] The kernel sizes of the first, second, and third convolutional blocks are k=7, k=5, and k=3, respectively, with 32, 64, and 128 output channels. Each convolutional block is followed by a batch normalization layer, a ReLU activation layer, and a max pooling layer. The squeezed activation channel attention module includes a global average pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a Sigmoid activation layer, used to adaptively adjust the weights of each channel. The adaptive global average pooling layer compresses convolutional feature maps of arbitrary length into 128-dimensional fixed-length feature vectors. In the fully connected classifier, the first fully connected layer maps the 128 dimensions to 256 dimensions. After a random deactivation layer (dropout rate of 0.5), the second fully connected layer maps the 256 dimensions to 3 dimensions. Finally, a normalized exponential function layer outputs the classification probabilities of various fault types.

[0135] When the fault category in the target domain is inconsistent with that in the source domain (e.g., adding a rolling element fault), the last layer dimension needs to be adjusted and training re-initialized when loading the pre-trained weights in step S4. Step S2: Source domain baseline model pre-training hyperparameters: optimizer Adam, initial learning rate 0.001, cosine annealing scheduling, batch size 64, total epochs 100, early stopping mechanism, data augmentation using random phase shift and Gaussian noise superposition; after training, the source domain pre-trained model weights are obtained.

[0136] like Figure 3As shown, step S3, target domain data preprocessing, uses physical duration T as the alignment reference. The source domain signal is segmented by a source domain sliding window, with N points in the source domain window. src =int(f ssrc ×T) = 2556 points; the sampling rate f of the target domain signal stgt =32 kHz, divided by a sliding window of the target domain, with N points in the target domain window. tgt =int(f stgt ×T) = 6816 points. When the number of points in the two domain windows is not equal, the backbone network uses an adaptive global average pooling layer to uniformly compress input signals of different lengths into a 128-dimensional fixed-length feature vector, without the need for downsampling, resampling, or interpolation, thus preserving high-frequency fault feature information in the high-sampling-rate data. For other common sampling rates such as 16 kHz, 25.6 kHz, 51.2 kHz, 64 kHz, and 200 kHz, N is used as the basis for compression. tgt =int(f stgt The number of points in the window is calculated using the ×T method; the samples after the target domain is segmented are also independently subjected to Z-score standardization.

[0137] To verify the effectiveness, stability, and universality of the method of the present invention, this embodiment employs three sets of independent cross-domain migration experiments. Experiment 1 used the CWRU dataset (12 kHz, constant operating condition) as the source domain and the Paderborn University (PU) dataset (64 kHz) as the target domain proxy to verify the ability to migrate across sampling rates and experimental benches. Experiment 2 used CWRU as the source domain and the University of Ottawa (Ottawa) constant speed dataset as the target domain proxy, serving as a control group that migrated across datasets but maintained constant speed conditions to verify migration stability. Experiment 3 used CWRU as the source domain and the Ottawa variable speed dataset (publicly released on the IEEE DataPort platform, named "Bearing Vibration Dataunder Time-varying Rotational Speed ​​Conditions") as the target domain proxy, selecting a subset of typical acceleration-uniform speed-deceleration complete operating cycles (600 to 1500 rpm, sampling rate 200 kHz) from this dataset to verify the effectiveness of the migration condition under which the source domain is constant speed and the target domain is variable speed, which is closest to the three-stage variable operating condition scenario of an elevator.

[0138] The three sets of experiments used the same network structure, hyperparameters, and training process. Each set of experiments was conducted 10 independent replicates under four progressive schemes, with each experiment using a different random seed. The four schemes are as follows: Exp0 Direct Testing (No Transfer), which directly uses the source domain pre-trained model for inference on the target domain test set without any fine-tuning; Exp1 Fine-tuning the Classifier Only, which only unfreezes the fully connected classifier, freezes other parameters of the backbone network, and then fine-tunes it on the target domain; Exp2 Progressive Parameter Freeze Transfer (Without Subdomain Alignment), which performs parameter freeze transfer according to step S4 of this invention but does not introduce subdomain adaptive domain alignment loss; Exp3 The Complete Method of this Invention, which is a collaborative scheme of progressive parameter freeze + subdomain adaptive domain alignment + source domain gradient truncation.

[0139] like Figure 5 As shown, with CWRU as the source domain and PU as the target domain, the mean diagnostic accuracy of the four progressive schemes is as follows: Exp0 approximately 21.82%, Exp1 approximately 55.90%, Exp2 approximately 88.42%, and Exp3 approximately 96.07%; the corresponding macro-mean F1 scores are approximately 17.21%, 55.32%, 88.41%, and 96.14%, respectively. Exp3 improves by approximately 7.65 percentage points compared to Exp2, with standard deviations of approximately 1.17% for Exp3 and 2.18% for Exp2. The paired-samples t-test yields a t-statistic of approximately 12.4 and a p-value < 0.001, indicating a highly statistically significant difference.

[0140] like Figure 6 As shown, the diagonal elements of the Exp3 confusion matrix are all relatively high while the off-diagonal elements are relatively small, and the accuracy of the three types of fault identification is higher than 96%, with balanced category identification. In contrast, the diagonal elements of Exp2 are significantly reduced, and the off-diagonal misclassification is significant.

[0141] like Figure 7 As shown, in the t-SNE visualization of Exp3, samples of the same type cluster together to form multiple compact local clusters, and each local cluster is composed of samples of the same type, with no obvious class overlap between local clusters of different types; while the feature space of Exp2 has obvious boundary blurring and class overlap.

[0142] like Figure 8As shown, the method of this invention achieves high diagnostic accuracy on three independent cross-domain migration paths. The mean accuracy of Exp3 (in 10 repeated experiments) on the CWRU→PU, CWRU→Ottawa constant speed, and CWRU→Ottawa variable speed paths is approximately 96.07%, 97.84%, and 96.05%, respectively, with standard deviations of approximately 1.17%, 0.95%, and 1.42%, respectively, representing improvements of approximately 7.65, 5.04, and 10.8 percentage points compared to Exp2. The improvement is most significant in experimental group three (constant speed in the source domain and variable speed in the target domain), demonstrating that the method of this invention is effective in migration scenarios that most closely resemble the actual deployment conditions of elevators.

[0143] like Figure 9 As shown, this embodiment compares the method of the present invention with three publicly available transfer learning baseline methods—DANN (Domain Adversarial Neural Network), Deep CORAL (Deep Relevance Alignment), and DAN / MMD (Deep Adaptive Network / Maximum Mean Difference)—across three transfer learning paths. Under the same dataset partitioning and the same number of training epochs (10 repeated experiments), the average accuracy of DANN on the three paths is approximately 82.52%, 87.53%, and 84.50%, respectively; Deep CORAL is approximately 80.21%, 85.43%, and 80.52%, respectively; DAN / MMD is approximately 81.44%, 86.21%, and 82.84%, respectively; and the accuracy of the method of the present invention is approximately 96.07%, 97.84%, and 96.04%, respectively, representing an improvement of 10 to 16 percentage points compared to the three baseline methods.

[0144] like Figure 10 As shown, the impact of the target domain training sample size on the gain of the subdomain adaptive module was examined under the CWRU→Ottawa variable speed path (10 repeated experiments). When the training sample ratio was 10%, 30%, 50%, 70%, and 100%, the mean accuracy difference between Exp3 and Exp2 was approximately 20.9, 16.4, 13.5, 12.0, and 10.8 percentage points, respectively. The results indicate that the scarcer the available samples in the target domain, the more significant the gain of the subdomain adaptive method of this invention.

[0145] like Figure 11As shown, ablation comparisons were performed on the physical duration segmentation strategies under the CWRU→PU path: Baseline A segmented the signal with a fixed number of sampling points N=2560 (the target domain only took a 40-millisecond physical duration segment due to its higher sampling rate); Baseline B first downsampled the target domain signal from 64 kHz to 12 kHz and then segmented it with 2560 sampling points; the proposed solution segmented the signal with a physical duration T=213 milliseconds (2556 points in the source domain and 13632 points in the target domain, combined with an adaptive global average pooling layer to handle variable-length input). The results of 10 repeated experiments showed that the mean diagnostic accuracy of Baseline A, Baseline B, and the proposed solution was approximately 82.43%, 85.16%, and 96.07%, respectively, with standard deviations of approximately 2.31%, 1.98%, and 1.17%, respectively. The proposed solution improved the accuracy by approximately 13.64 and 10.91 percentage points compared to Baseline A and Baseline B, respectively, and the p-values ​​of the paired-samples t-test were all less than 0.01. The performance degradation of baseline A is due to the target domain signal segment being too short and failing to cover the complete fault characteristic cycle; the performance degradation of baseline B is due to downsampling filtering out high-frequency fault impulse components.

[0146] like Figure 12 As shown, an ablation comparison of the source domain gradient truncation strategy is performed under the condition that the available training samples in the CWRU→PU path and the target domain account for 20% (approximately 480 samples), which is a small sample condition: the baseline C is the source domain sample that participates in L cls L domain With L center Gradient backpropagation (no gradient truncation in the source domain); the solution of this invention is that the source domain only participates in L... domain Calculation (gradient truncation performed in step S5). Results from 10 repeated experiments show that the mean accuracy of baseline C and the proposed scheme is approximately 82.32% and 91.18%, respectively, with standard deviations of approximately 2.85% and 1.34%. The proposed scheme improves upon baseline C by approximately 8.86 percentage points, and the p-value in the paired-samples t-test is less than 0.01. This is because baseline C has far more labeled samples in the source domain than available training samples in the target domain, and the source domain samples dominate the gradient backpropagation of the classification loss, causing the classifier to favor the source domain distribution. The proposed scheme calculates the gradient backpropagation of the classification loss and center loss based solely on target domain samples, avoiding source domain discrimination bias that contaminates the target domain classifier.

[0147] like Figure 13As shown, the dual-threshold triggering mechanism was compared under the CWRU→Ottawa variable speed path. During the online inference phase, a total of 1000 test windows were conducted (600 healthy, 400 faulty). The results for the three triggering schemes were as follows: Baseline D (H < 0.6 only) had a false alarm rate of 7.83% and a false alarm rate of 3.25%; Baseline E (confidence > 80% only) had a false alarm rate of 12.50% and a false alarm rate of 5.00%; the scheme of this invention (both thresholds are met simultaneously) had a false alarm rate of 1.67% and a false alarm rate of 3.50%. The scheme of this invention reduces the false alarm rate by approximately 6.16 and 10.83 percentage points compared to baselines D and E, respectively, while maintaining a false alarm rate at a level comparable to the baselines.

[0148] This embodiment uses elevator traction machine bearings as the preferred implementation object. The method of this invention is also applicable to other elevator transmission components such as elevator door operator bearings and guide wheel bearings, and can also be applied to rotating equipment scenarios with cross-condition and cross-sampling rate fault diagnosis requirements, such as industrial rotating machinery, wind turbines, and rail transit vehicles. The core algorithm of this invention can be independently deployed on an edge computing hardware platform, embedded controller, industrial computer, or cloud server. The communication protocol between the edge computing device and the elevator main control system includes, but is not limited to, industrial communication protocols such as CAN bus, Modbus, and Profinet.

[0149] Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention shall be included within the scope of protection of this invention.

Claims

1. A cross-domain fault diagnosis method for elevator transmission components based on deep learning, characterized in that: The method uses vibration sensors installed on elevator transmission components to collect vibration signal data for training the diagnostic model. During the offline training phase of the diagnostic model, the signal data is segmented into source and target domain vibration signals by physical duration T. During the segmentation process, the number of window points is determined by the product of the sampling rate and T. A one-dimensional convolutional neural network with an adaptive global average pooling layer is used as the backbone. Progressive parameter freezing transfer is adopted, combined with gradient inversion layer, domain classifier and center loss to achieve subdomain adaptive alignment. Source domain samples are truncated by gradient to participate in domain adversarial loss. In the online diagnostic phase of the method, a comprehensive health index is obtained through multi-window temporal smoothing. When both the comprehensive health index and confidence level exceed the corresponding threshold, a speed reduction linkage command is sent to the elevator main control system through the edge computing device.

2. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 1, characterized in that: The diagnostic model uses a source domain baseline model and a transfer model running on a computing device. The diagnostic method includes steps S1 to S5 for the offline training phase of the model, and also includes step S6 for performing online diagnosis using the model; specifically: Step S1, Source Domain Data Preprocessing: Vibration signals of transmission components with fault category labels are collected as source domain data. The source domain data is segmented using a sliding window based on physical duration T to obtain the source domain training sample set. Step S2, Source Domain Baseline Model Pre-training: Construct a diagnostic model with a convolutional neural network as the backbone network. The backbone network includes at least two convolutional blocks connected in series, an adaptive global average pooling layer, and a fully connected classifier. The adaptive global average pooling layer is used to process input signals of different lengths. The backbone network is trained under supervision on the source domain training sample set to obtain the source domain pre-trained model weights. Step S3, target domain data preprocessing: collect the target domain vibration signal of the elevator transmission component under variable working conditions, and use the same physical duration T as in step S1 to perform sliding window segmentation to obtain the target domain training sample set; Step S4, Progressive parameter freeze transfer: Load the weights of the source domain pre-trained model, freeze the parameters of the shallow convolutional blocks in the backbone network, unfreeze the parameters of the deep convolutional blocks and the fully connected classifier, and perform fine-tuning training on the target domain training sample set; Step S5, Subdomain Adaptive Domain Alignment: During the fine-tuning training process in step S4, a gradient reversal layer and a domain classifier are introduced at the output of the adaptive global average pooling layer, and the center loss function is added to the training objective to form the total training loss function; in each training iteration, source domain samples and target domain samples are loaded simultaneously, and the features of the source domain samples are extracted by gradient truncation, which only participates in the calculation of domain adversarial loss and does not participate in the gradient backpropagation of classification loss and center loss; Step S6, Online Diagnostic Inference and Speed ​​Reduction Linkage Command Output: The vibration signal of the elevator transmission component to be diagnosed is segmented as described in step S3 and input into the transfer model trained in steps S4 and S5, and the fault category diagnosis result is output; a comprehensive health index is calculated based on the time-series smoothing result of the fault category diagnosis probability of multiple consecutive sliding windows, and the confidence level is determined according to the consistency ratio of the fault categories in the multiple consecutive sliding windows; when the comprehensive health index is lower than a preset threshold and the confidence level is higher than a preset value, the computer sends the speed reduction linkage command to the elevator main control system through the communication interface.

3. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: The number of window points N for segmenting the selected domain of the source domain data in step S1. src Equal to the source domain sampling rate f ssrc The product of the product with the physical duration T is rounded down; The number of window points N for segmenting the target domain data in step S3. tgt Equal to the target domain sampling rate f stgt The product of the product with the physical duration T is rounded down; In steps S1 and S3, the segmented samples are subjected to dimensionless standardization.

4. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: In step S5, the total loss function L is trained. total The weighted sum of classification loss, domain adversarial loss, and centering loss is calculated using the following formula: Where L cls For classification cross-entropy loss based solely on samples from the target domain, L domain To achieve domain adversarial binary classification based on source and target domain samples, L center For the center loss based solely on samples from the target domain, λ domain With λ center These are weighting coefficients; The specific calculation formulas for the three losses—classification loss, domain adversarial loss, and center loss—are as follows: Where N t Let C be the number of training samples in the target domain, and y be the total number of fault categories. i,c For one-hot real labels, p i,c To predict probabilities, L cls Calculations are based solely on samples from the target domain. Where N is the total number of samples involved in the calculation, d i For the source domain binary label, g(x) i ) is the eigenvector x i The output after gradient inversion layer processing, D(·), represents the probability that the domain classifier's output belongs to the target domain. The domain classifier consists of two fully connected layers and a sigmoid activation layer; Where f(x) i Let be the feature vector of the i-th sample in the target domain, and c be the feature vector of the i-th sample. yi The category y to which this sample belongs i The learnable center vector is obtained. The center vector is updated using an independent auxiliary optimizer, and the denominator in the update formula is m. y +1 to avoid division by zero anomalies. The initial value of the center vector is the mean of the feature vectors of each class in all training samples in the target domain; The gradient reversal layer performs an identity mapping during forward propagation and multiplies the gradient by a negative factor -α during backpropagation. p represents the training progress (p∈[0,1]), γ ranges from 5 to 15, and α increases smoothly from close to 0 at the beginning of training to close to 1 at the end according to an S-shaped curve. Each training batch loads both source and target domain samples simultaneously: source domain sample features, after being processed by the gradient truncation module, only participate in L... domain Calculation, not involved in L cls With L center Gradient backpropagation; target domain samples participate in the calculation of three loss terms simultaneously; gradient truncation is implemented using the gradient computation disabled context provided by the deep learning framework. The three strategies of gradient truncation, progressive parameter freezing, and subdomain adaptive domain alignment work together: progressive parameter freezing provides a stable source domain low-level feature base for subdomain alignment; Source domain gradient truncation focuses the classification loss and center loss on the target domain; domain adversarial loss and center loss work together to achieve fine alignment at the class level.

5. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: The physical duration T mentioned in steps S1 and S3 ranges from 30 milliseconds to 300 milliseconds; when the source domain sampling rate f ssrc With the target domain sampling rate f stgt When the features are unequal, the adaptive global average pooling layer compresses convolutional feature maps of different lengths into fixed-length feature vectors of the same dimension; the source domain sampling rate f ssrc With the target domain sampling rate f stgt The value range is from 10 kHz to 200 kHz.

6. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: In step S2, the kernel size of each convolutional block decreases sequentially along the signal flow direction. The kernel size of each subsequent convolutional block is smaller than that of the previous convolutional block, and the number of output channels is not less than twice the number of output channels of the previous convolutional block. Each convolutional block contains a convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer. Each convolutional block is followed by a squeezed excitation channel attention module. In step S4, when freezing the parameters of the shallow convolutional block, the parameters of the squeezed excitation channel attention module following the shallow convolutional block are frozen simultaneously.

7. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: The time-series smoothing in step S6 adopts the exponential moving average method. The comprehensive health index is the time-series smoothed value of the normal category diagnosis probability. The confidence level is the proportion of windows with the same fault category in the multiple consecutive sliding windows. The number of the multiple consecutive sliding windows ranges from 3 to 10. The preset threshold value of the comprehensive health index ranges from 0.5 to 0.7; the deceleration corresponding to the deceleration action indicated by the deceleration linkage command does not exceed 0.1 m / s², and the target running speed after deceleration is not less than 40% of the elevator's rated speed; The progressive parameter freeze transfer strategy in step S4 is as follows: After loading the source domain pre-trained model weights obtained in step S2, the parameters of the first convolutional block and its subsequent first squeeze excitation channel attention module are frozen through the parameter freeze module, while the parameters of the remaining convolutional blocks, attention modules, and fully connected classifiers are unfrozen through the parameter unfreeze module. The basic impulse and envelope features learned by shallow modules have cross-domain universality and are suitable for freezing and preserving source domain knowledge.

8. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: In step S3, the variable operating condition signals consisting of acceleration, constant speed and deceleration phases in a single elevator operation cycle are uniformly covered by a physical duration sliding window, and the transfer model directly outputs the fault category diagnosis results for each signal segment.

9. The method for cross-domain fault diagnosis of elevator transmission components based on deep learning according to claim 2, characterized in that: The speed reduction linkage command in step S6 includes verification information for the elevator main control system to perform safety verification. The verification information includes at least one of the comprehensive health index, the confidence level, the fault category diagnosis result, and the command timestamp. Step S6 further includes a data reliability judgment sub-step: when the signal-to-noise ratio of the vibration signals output by each sensor node is lower than a preset lower limit, the current data is determined to be unreliable and a retest process is triggered. The preset lower limit ranges from 5 dB to 10 dB. Specifically: After the weights of the transfer model trained offline are distributed to the edge computing hardware platform, the online phase process in step S6 is as follows: The vibration signal acquisition module acquires the vibration signal to be diagnosed from the elevator transmission component, and the sliding window segmentation module segments the sample to be diagnosed according to the physical duration T. The sample is then input into the transfer model inference module to output the classification probabilities of various faults; the time-series smoothing module performs smoothing based on the diagnostic probabilities of N consecutive sliding windows; the comprehensive health index calculation module calculates the comprehensive health index H according to the exponential moving average formula. in Let H be the normal category diagnosis probability for the t-th window, and β be the smoothing coefficient; the closer the comprehensive health index H is to 1, the better the health status. The confidence level is the percentage of windows that output the same fault category in N consecutive sliding windows. When the dual threshold judgment module determines that the comprehensive health index H is lower than the preset threshold and the confidence level is higher than the preset value, the speed reduction linkage command generation module generates a speed reduction linkage command and sends it to the elevator main control system through the communication interface; the command contains verification information for safety verification, including at least one of the following: comprehensive health index H, confidence level, fault type and command timestamp; The variable operating conditions of a single elevator operation cycle consist of an acceleration phase, a constant speed phase, and a deceleration phase. The diagnostic method uses a sliding window of the same physical duration to cover the signal segments of each stage, and the transfer model directly outputs the fault category diagnosis results for each window segment.

10. A deep learning-based cross-domain fault diagnosis system for elevator transmission components, using the deep learning-based cross-domain fault diagnosis method for elevator transmission components as described in any one of claims 2 to 9, characterized in that: The diagnostic system includes a processor and a memory; or includes a processor, a memory, and at least one of a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), and a graphics processor (GPU); the processor and the memory, or the processor, the memory, and at least one of the FPGA, ASIC, and GPU are configured to implement step S6 of the deep learning-based cross-domain fault diagnosis method for elevator drive components as described in any one of claims 2 to 9; the memory stores a computer program, which, when executed by the processor, implements the steps of the deep learning-based cross-domain fault diagnosis method for elevator drive components as described in any one of claims 2 to 9; The diagnostic system also includes an edge computing device and an elevator main control system. The output of the edge computing device is connected to the input of the elevator main control system via a wired or wireless communication interface. The communication interface is configured inside the edge computing device or is set independently between the edge computing device and the elevator main control system. The edge computing device outputs a speed reduction linkage command to the elevator main control system. The speed reduction linkage command is used to instruct the elevator main control system to perform a speed reduction action on the elevator. The elevator main control system controls the elevator transmission components. In the elevator transmission components, the traction machine drives the elevator car and counterweight to rise and fall via the traction sheave and wire rope, and the door operator, guide wheel and other components realize the opening and closing of the elevator car door and the running guidance. The rotating parts of the elevator transmission components are supported by rolling bearings to rotate. The diagnostic system uses vibration sensors installed in the bearing housings or end caps of elevator transmission components to synchronously acquire vibration signal data for training the diagnostic model via multiple channels. The vibration signal is acquired by installing a vibration sensor at a location near the bearing housing, end cover, or rigidly connected to the bearing housing of the monitored elevator transmission component. The vibration sensor is at least one of a piezoelectric accelerometer, a MEMS accelerometer, or an eddy current displacement sensor. The sensitive axis of the vibration sensor is arranged along the radial or axial direction of the bearing being measured; when monitoring a single transmission component, a single sensor is arranged to form a single sensing node. When it is necessary to simultaneously monitor multiple transmission components such as traction machine bearings, gantry crane bearings, and guide wheel bearings, or to monitor the same component in multiple directions, multiple vibration sensors are arranged to form multiple sensing nodes. Each sensing node collects vibration signals at the same time reference through a multi-channel synchronous acquisition circuit. After the vibration signal is output by the sensor, it is amplified and anti-aliasing filtered by the signal conditioning circuit, and then discretized into a digital vibration signal by the analog-to-digital converter at a preset sampling rate. The sampling rate is set according to the frequency range of the fault features to be retained, and the value range is 10 kHz to 200 kHz. The source domain data and the target domain data use different sampling rates, and the input signals of different lengths generated by different sampling rates are processed by an adaptive global average pooling layer.

Images