A Fault Diagnosis Method for Elevator Brakes Based on Multi-Scale Feature Distillation

Abstract

This invention provides a fault diagnosis method for elevator brakes based on multi-scale feature distillation, relating to the field of elevator brake fault diagnosis. The method includes: acquiring multi-channel signals related to the braking event of the elevator brake, including PVDF strain, acceleration time-domain signals, and high-frequency acoustic emission transient pulse signals; processing these signals through time alignment, signal slicing, windowing, and normalization to obtain standardized signals; inputting these signals into a main diagnostic network, extracting and fusing multi-scale features to complete dual-task diagnosis of fault type and severity; training the network with real labels and generating soft labels; constructing a lightweight edge diagnostic network; training a deployable diagnostic model based on real and soft label distillation; and acquiring and processing real-time signals before inputting them into the diagnostic model to obtain the fault type and severity. This invention improves the accuracy of fault identification and severity assessment, while simultaneously enabling real-time online diagnosis of elevator brakes.

Description

Technical Field

[0001] This invention relates to the field of elevator brake fault diagnosis, and more specifically, to an elevator brake fault diagnosis method based on multi-scale feature distillation. Background Technology

[0002] Elevator brakes are used for normal parking braking, floor holding braking, and power failure / emergency braking. Common braking failures include insufficient braking force leading to slippage, incomplete release causing friction pad overheating, and delayed braking response leading to shock and vibration. In severe cases, this can lead to significant safety hazards such as slippage or holding failure. Elevator brakes operate in a complex environment for extended periods, facing challenges such as high noise, weak signals, non-stationarity, and multi-source coupling characteristics, making fault detection and diagnosis extremely difficult.

[0003] Traditional fault diagnosis methods, such as spectral analysis based on acceleration signals and empirical threshold methods, are typically suitable for diagnosing single faults under steady-state conditions. However, in non-steady-state multi-source coupled noise environments, the acquired signals are simultaneously mixed with multiple noise sources, which then superimpose through resonance or electromagnetic interference, causing the brake characteristics to be "submerged" and significantly reducing the diagnostic effectiveness. Furthermore, the complex physical relationships between compound faults, early fault signals, and different sensor channels result in low recognition rates and high false positive rates for these traditional methods. Therefore, achieving highly robust and engineered elevator brake fault diagnosis, especially in scenarios with strong noise and significant changes in operating conditions, remains a major challenge in the current technological field.

[0004] Current fault diagnosis methods primarily rely on vibration signal spectrum analysis and traditional machine learning techniques, such as SVM (Short Vector Machine), KNN (K-Nearest Neighbors), and Random Forest. These methods are highly dependent on expert experience and exhibit poor stability and accuracy under dynamic operating conditions. In particular, spectrum analysis methods often fail to effectively extract useful features in complex noisy environments, causing fault signals to be easily obscured by noise. While deep learning methods can automatically extract complex features, their "black box" nature limits their interpretability, posing significant challenges to their application in high-safety scenarios such as elevators.

[0005] In addition, existing diagnostic methods based on physical models, such as multibody dynamics models, can explain the fault mechanism from a physical perspective, but they suffer from high costs in model building and calibration. Furthermore, key parameters drift over time, resulting in insufficient real-time performance and adaptability of the model, making it difficult to apply widely to different types and operating conditions of equipment. Summary of the Invention

[0006] This invention provides a method, system, device, storage medium, and computer program product for diagnosing elevator brake faults based on multi-scale feature distillation. By acquiring multi-channel signals and combining multi-scale feature fusion and dual-task diagnosis, the accuracy of fault identification and severity assessment is improved. At the same time, a lightweight edge model is constructed by knowledge distillation and directly deployed on the device to achieve real-time online diagnosis of elevator brakes.

[0007] According to a first aspect of the present invention, an embodiment of the present invention provides a method for diagnosing elevator brake faults based on multi-scale feature distillation, comprising: acquiring multi-channel signals related to braking events of the elevator brake; performing windowing and normalization processing on the multi-channel signals to obtain standardized multi-channel signals for each window within a braking event segment; inputting the standardized multi-channel signals into a multi-scale feature extraction network of a main diagnostic network, fusing multi-channel and multi-scale features through a feature layer to obtain shared features, performing a dual-task diagnosis of fault type identification and fault severity assessment based on the shared features, and outputting the fault type probability distribution and fault severity probability distribution for each window; calculating a loss function based on the fault type probability distribution and fault severity probability distribution for each window, and according to... The loss function is used to train the main diagnostic network, and soft labels are generated for the dual-task diagnosis of the main diagnostic network. A lightweight edge diagnostic network corresponding to the main diagnostic network is constructed, and the lightweight edge diagnostic network is trained by distillation based on the real labels and the soft labels to obtain a fault diagnosis model that can be deployed on edge devices. The standardized real-time multi-channel signal of each window in the current braking event segment is obtained based on the real-time multi-channel signal of the elevator brake. The standardized real-time multi-channel signal is input into the fault diagnosis model, and the fault diagnosis model outputs the fault type probability value and fault severity probability value of each window in the current braking event segment. The fault type and degree of the elevator brake are determined based on the fault type probability value and fault severity probability value of each window in the current braking event segment.

[0008] According to the above embodiments of the present invention, a lightweight diagnostic model that can run on edge devices is obtained by multi-task, multi-scale feature modeling and knowledge distillation. The current elevator brake fault type and severity are output according to the lightweight diagnostic model, realizing online identification and severity assessment of brake faults. This ensures that high fault diagnosis robustness and accuracy are maintained even in noisy and fluctuating field environments.

[0009] In some embodiments of the present invention, acquiring multi-channel signals related to the braking event of the elevator brake includes: acquiring the original multi-channel signals of the elevator brake, the original multi-channel signals including: PVDF strain signals, acceleration time-domain signals, and high-frequency acoustic emission transient pulse signals; performing time alignment processing on the original multi-channel signals, and performing signal slicing processing on the time-aligned multi-channel signals according to the braking trigger point to obtain the multi-channel signals related to the braking event.

[0010] According to the above embodiments of the present invention, by jointly collecting signals such as acceleration, PVDF strain and acoustic emission of the elevator brake, the sensitivity to typical elevator brake faults such as jamming and wear is enhanced by using multi-source complementary means.

[0011] In some embodiments of the present invention, the fault diagnosis method further includes: if a braking control command of the elevator brake is obtained, the command edge of the braking control command is taken as the braking trigger point; if the braking control command of the elevator brake is not obtained, the braking trigger point is determined by joint mutation detection of the multi-channel signals.

[0012] In some embodiments of the present invention, the multi-scale feature extraction network is a three-branch parallel convolutional residual structure with multi-channel matching. The acceleration time-domain signal is fed into the short-scale branch of the three-branch parallel convolutional residual structure to extract low-frequency vibration features, the PVDF strain signal is fed into the medium-scale branch to obtain strain features, and the high-frequency acoustic emission transient pulse signal is fed into the long-scale branch to extract high-frequency pulse features.

[0013] In some embodiments of the present invention, the fault diagnosis method further includes: calculating a sluggish score based on the acceleration time-domain signal; calculating a friction interface degradation score based on the PVDF strain signal and the high-frequency acoustic emission transient pulse signal; and generating the real label based on the sluggish score and the friction interface degradation score combined with a preset threshold.

[0014] According to the above embodiments of the present invention, the jamming score and friction interface degradation score are calculated based on the physical characteristics of each channel signal, forming a precise physical correspondence with the fault type (jamming, wear). This gives the generated real labels clear physical meaning, rather than simply abstract labels, further improving the targeting and accuracy of subsequent main diagnostic network training and providing reliable label support for fault type identification and severity assessment. Simultaneously, it enables the quantitative and systematic generation of real labels, overcoming the limitations of traditional labels relying on manual annotation, and making label generation more closely aligned with the actual fault evolution patterns of elevator brakes. Furthermore, this label generation method forms a closed loop with the logic of multi-channel signal processing and multi-scale feature extraction, ensuring that the entire fault diagnosis process, from signal acquisition and feature extraction to label generation and model training, revolves around the physical mechanism of brake faults. This further enhances the interpretability of the diagnostic method, overcomes the shortcomings of traditional deep learning "black box" diagnosis, and adapts to the needs of high-safety elevator application scenarios.

[0015] In some embodiments of the present invention, the fault types include three categories: normal, jamming, and friction plate wear, and the fault severity includes four levels: normal, minor, moderate, and severe.

[0016] In some embodiments of the present invention, determining the fault type and severity of the elevator brake based on the fault type probability value and fault severity probability value of each window within the current braking event segment includes: performing mean fusion on the fault type probability distribution of all windows within the current braking event segment, and taking the fault category corresponding to the maximum probability value after fusion as the final fault type of the elevator brake; taking the maximum probability value at the window level for each fault severity level according to the fault severity level, and then taking the fault severity level corresponding to the largest value among the maximum probability values ​​at the window level for each fault severity level as the final fault severity of the elevator brake.

[0017] According to a second aspect of the present invention, an embodiment of the present invention provides an elevator brake fault diagnosis system based on multi-scale feature distillation, wherein the adjustment system performs the fault diagnosis method described in any embodiment of the present invention.

[0018] According to a third aspect of the present invention, an embodiment of the present invention provides a computer-readable storage medium having computer-readable instructions stored thereon, which, when executed by a processor, cause a computer to perform the following operations: the operations include the steps included in the fault diagnosis method as described in any of the above embodiments.

[0019] According to a fourth aspect of the present invention, an embodiment of the present invention provides a computer device including a memory and a processor, wherein the memory is used to store one or more computer-readable instructions, wherein the one or more computer-readable instructions, when executed by the processor, can implement the fault diagnosis method as described in any of the above embodiments.

[0020] According to a fifth aspect of the present invention, an embodiment of the present invention provides a computer program product including a computer program, which, when executed by a processor, implements the fault diagnosis method as described in any of the above embodiments.

[0021] As described above, the elevator brake fault diagnosis method, system, device, storage medium, and computer program product based on multi-scale feature distillation provided by the embodiments of the present invention enhance the sensitivity to typical elevator brake faults such as jamming and wear by jointly acquiring signals such as acceleration, PVDF strain, and acoustic emission, and by using multi-source complementary means. Through multi-task, multi-scale feature modeling and knowledge distillation, a lightweight diagnostic model that can run on edge devices is obtained, and the fault type and severity of the current elevator brake are output according to the lightweight diagnostic model, realizing online identification and severity assessment of brake faults. This ensures that high fault diagnosis robustness and accuracy are maintained even in noisy and fluctuating field environments. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the elevator brake fault diagnosis method based on multi-scale feature distillation according to Embodiment 1 of the present invention.

[0023] Figure 2 This is a perspective view of the structure of a data acquisition device for acquiring raw multi-channel signals of an elevator caliper disc brake according to an embodiment of the present invention;

[0024] Figure 3 This is a flowchart illustrating the elevator brake fault diagnosis method based on multi-scale feature distillation according to Embodiment 2 of the present invention.

[0025] Figure 4 This is a schematic diagram of the architecture of TaskKD-ScaleNet, an elevator brake fault diagnosis network according to an embodiment of the present invention.

[0026] The reference numerals in the attached figures are explained as follows: 1- PVDF flexible piezoelectric strip, 2- AE sensor, 3- accelerometer. Detailed Implementation

[0027] The various aspects of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Well-known modules, units, and their connections, links, communications, or operations are not shown or described in detail. Furthermore, the described features, architectures, or functions can be combined in any way in one or more embodiments. Those skilled in the art should understand that the various embodiments described below are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. It will also be readily understood that the modules, units, or processing methods in the embodiments described herein and shown in the accompanying drawings can be combined and designed in various different configurations.

[0028] The following is a brief explanation of the terms used in this text.

[0029] PVDF flexible piezoelectric strip: A flexible functional device with PVDF piezoelectric polymer as the core, processed into a thin strip shape. It can realize bidirectional conversion of mechanical energy to electrical energy (piezoelectric effect), and has excellent flexibility, wide frequency response and chemical stability. It is an ideal sensing element in the fields of structural health monitoring, acoustic emission detection, strain sensing and other fields.

[0030] KL divergence (KLD), also known as relative entropy, is used to measure the difference between two probability distributions.

[0031]

Example 1

[0032] Figure 1 This is a flowchart illustrating the elevator brake fault diagnosis method based on multi-scale feature distillation according to Embodiment 1 of the present invention.

[0033] like Figure 1 As shown, in Embodiment 1 of the present invention, the elevator brake fault diagnosis method may include at least the following steps S11, S12, S13, S14, S15, S16, S17 and S18, which are described in detail below.

[0034] In step S11, a multi-channel signal related to the braking event of the elevator brake is acquired.

[0035] In some implementations, acquiring multi-channel signals related to the braking event of the elevator brake includes: acquiring the original multi-channel signals of the elevator brake, the original multi-channel signals including: PVDF strain signals, acceleration time-domain signals, and high-frequency acoustic emission transient pulse signals; performing time alignment processing on the original multi-channel signals, and performing signal slicing processing on the time-aligned multi-channel signals according to the braking trigger point to obtain the multi-channel signals related to the braking event.

[0036] In some implementations, strain signals are acquired using PVDF flexible piezoelectric strips, acceleration time-domain vibration signals are acquired using accelerometers, and high-frequency transient pulse signals are acquired using acoustic emission sensors.

[0037] This invention provides a specific example of a data acquisition device for acquiring raw multi-channel signals from a single caliper disc brake. The acquisition device includes: a PVDF (polyvinylidene fluoride) flexible piezoelectric strip, an accelerometer, and an AE (acoustic emission) sensor. The following is a detailed description... Figure 2 The example will be explained in detail.

[0038] Figure 2 This is a perspective view of the structure of a data acquisition device for acquiring raw multi-channel signals from an elevator caliper disc brake according to an embodiment of the present invention, as shown below. Figure 2 As shown, the PVDF flexible piezoelectric strip 1 serves as the main detection sensor, attached to the metal substrate on the back of the friction pair along the principal stress line of the brake caliper. The PVDF strip can monitor shear strain and friction contact pressure fluctuations in real time. When the friction pad thickness decreases or the contact becomes uneven, the amplitude and frequency characteristics of the strain signal will change. Furthermore, by combining the strain signal collected by the PVDF flexible piezoelectric strip with the signal collected by the AE sensor 2 (described later) for time-frequency analysis, the wear fault identification of the friction pad can be achieved.

[0039] In this embodiment, a PVDF strip structure with principal stress lines attached to the clamp body is used, which can capture stress and vibration characteristics with a higher signal-to-noise ratio. The distribution of the principal stress lines of the clamp body is determined through finite element analysis, and a flexible PVDF strip is attached along this direction. Through its piezoelectric effect, the minute mechanical strain of the brake clamp body is converted into an electrical signal, thereby achieving high-sensitivity dynamic monitoring and upgrading from localized point acquisition to continuous linear strain acquisition. The integral shear signal and acceleration signal output from the PVDF channel are complementary, which can effectively detect latent faults such as friction pad wear.

[0040] like Figure 2 As shown, AE sensor 2 is installed in a hole on the back of the brake caliper friction pair. The contact surface at the bottom of the hole has been deburred, leveled, and cleaned. A thin layer of coupling agent is applied to the probe end face to improve high-frequency signal transmission. AE sensor 2 is used to capture high-frequency pulse data of microcracks and friction problems on the brake surface. By combining this data with the signal acquired by the PVDF sensor for time-frequency analysis, wear on the friction pads can be identified.

[0041] Accelerometer 3 is fastened to the inner surface of the brake bracket, with priority given to locations on the bracket's inner surface that are "highly rigid and close to the transmission path of the clamp arm," avoiding weak cantilevered areas or edges prone to resonance, to improve mechanical coupling and signal-to-noise ratio. When the elevator brake is released normally, the vibration signal is a short pulse, while when it is stuck or malfunctioning, it generates periodic low-frequency vibrations and delayed impact signals. The time-domain waveform and spectral characteristics of the acceleration channel can directly reflect the degree of resistance in the clamp arm return mechanism, thereby detecting brake sticking.

[0042] The above-mentioned acquisition device of the present invention acquires multi-modal synchronous data on strain, vibration and high-frequency friction signals during the operation of elevator brake through PVDF flexible piezoelectric strip, accelerometer and acoustic emission sensor respectively. The information of the three complements each other, which can enhance the sensitivity to latent faults and improve the diagnostic robustness under on-site noise and operating condition fluctuations.

[0043] In some implementations, the braking trigger point is obtained through the following steps: if a braking control command of the elevator brake is obtained, the command edge of the braking control command is taken as the braking trigger point; if the braking control command of the elevator brake is not obtained, the braking trigger point is determined by joint mutation detection of the multi-channel signals.

[0044] In step S12, the multi-channel signal is windowed and normalized to obtain a standardized multi-channel signal for each window within the braking event segment.

[0045] In step S13, the standardized multi-channel signal is input into the multi-scale feature extraction network of the main diagnostic network. The multi-channel and multi-scale features of the feature layer are fused to obtain shared features. Based on the shared features, a dual-task diagnosis of fault type identification and fault severity assessment is performed, and the fault type probability distribution and fault severity probability distribution of each window are output.

[0046] The multi-scale feature extraction network is a three-branch parallel convolutional residual structure with multi-channel matching. The acceleration time-domain signal is fed into the short-scale branch of the three-branch parallel convolutional residual structure to extract low-frequency vibration features, the PVDF strain signal is fed into the medium-scale branch to obtain strain features, and the high-frequency acoustic emission transient pulse signal is fed into the long-scale branch to extract high-frequency pulse features.

[0047] In step S14, based on the probability distribution of fault type and fault severity for each window, a loss function is calculated by combining the true labels of fault type and fault severity. The main diagnostic network is then trained according to the loss function, and soft labels are generated for the dual-task diagnosis of the main diagnostic network. The fault types include three categories: normal, stuck, and friction plate wear; the fault severity includes four levels: normal, minor, moderate, and severe.

[0048] In some implementations, the real label is obtained by the following steps: calculating a hysteresis score based on the acceleration time-domain signal; calculating a friction interface degradation score based on the PVDF strain signal and the high-frequency acoustic emission transient pulse signal; and generating the real label based on the hysteresis score and the friction interface degradation score combined with a preset threshold.

[0049] In step S15, a lightweight edge diagnostic network corresponding to the main diagnostic network is constructed, and the lightweight edge diagnostic network is trained by distillation based on the real labels and the soft labels to obtain a fault diagnosis model that can be deployed on edge devices. The lightweight edge diagnostic network is a lightweight edge diagnostic network with consistent feature extraction and dual-task diagnostic logic with the main diagnostic network, but with lower model complexity.

[0050] In step S16, the standardized real-time multi-channel signal of each window in the current braking event segment is obtained based on the real-time multi-channel signal of the elevator brake collected in real time.

[0051] In step S17, the standardized real-time multi-channel signal is input into the fault diagnosis model, and the fault diagnosis model outputs the fault type probability value and fault severity probability value for each window within the current braking event segment.

[0052] In step S18, the fault type and severity of the elevator brake are determined based on the fault type probability value and fault severity probability value of each window within the current braking event segment.

[0053] In some implementations, determining the fault type and severity of the elevator brake based on the fault type probability value and fault severity probability value of each window within the current braking event segment includes: performing mean fusion on the fault type probability distribution of all windows within the current braking event segment, and taking the fault category corresponding to the maximum probability value after fusion as the final fault type of the elevator brake; taking the maximum probability value at the window level for each fault severity level according to the fault severity level, and then taking the fault severity level corresponding to the largest value among the maximum probability values ​​at the window level for each fault severity level as the final fault severity of the elevator brake.

[0054] The elevator brake fault diagnosis method described in Embodiment 1 of this invention enhances the sensitivity to typical elevator brake faults such as jamming and wear by jointly acquiring signals such as acceleration, PVDF strain, and acoustic emission, and using multi-source complementary methods. Through multi-task, multi-scale feature modeling and knowledge distillation, a lightweight diagnostic model that can run on edge devices is obtained. Based on this lightweight diagnostic model, the fault type and severity of the current elevator brake are output, realizing online identification and severity assessment of brake faults. This ensures high robustness and accuracy of fault diagnosis even in noisy and fluctuating field environments.

[0055]

Example 2

[0056] Figure 3 This is a flowchart illustrating the elevator brake fault diagnosis method based on multi-scale feature distillation according to Embodiment 2 of the present invention.

[0057] like Figure 3 As shown, in Embodiment 2 of the present invention, the fault diagnosis method may include at least the following steps S21, S22, S23, S24, S25 and S26, which are described in detail below.

[0058] Step S21: Multimodal sensors synchronously acquire strain, vibration, and high-frequency friction signals during the operation of the elevator brake. PVDF flexible piezoelectric strips, accelerometers, and acoustic emission sensors are used to synchronously acquire these signals in multiple modes during elevator brake operation. The complementary information from these three sensors enhances sensitivity to latent faults and improves diagnostic robustness under varying noise and operating conditions.

[0059] Step S22: Time alignment of the multi-channel signals and signal slicing and windowing preprocessing based on the braking trigger event.

[0060] Step S23: Perform multi-channel and multi-scale feature fusion processing on the standardized multi-channel signal.

[0061] Step S24, multi-task joint fault diagnosis. Steps S23 and S24 input the standardized multi-channel signal into the multi-scale residual feature extraction network, complete the fusion of multi-channel and multi-scale features at the feature layer, and jointly complete the multi-task diagnosis of fault type identification and severity assessment based on shared features.

[0062] Step S25, Knowledge Distillation and Lightweight Model Generation. During the training phase, the multi-scale, multi-task knowledge of the main diagnostic network is uniformly transferred to the lightweight edge diagnostic model through output layer soft label distillation and intermediate feature alignment. This allows the model to maintain recognition capabilities close to those of the main model with fewer parameters and less computing power, facilitating real-time online deployment. At the same time, temperature-sensitive soft labels are used to preserve the similarity relationships between categories, improving the integrity of the model's learning information and its generalization performance.

[0063] Step S26: Output the fault type and severity. During the inference phase, by performing event-level fusion on the window-level diagnostic results and combining them with the mechanism score constructed based on acceleration, PVDF, and acoustic emission signals, the final fault type, severity, and interpretable diagnostic results of the elevator brake are output.

[0064] Meanwhile, this invention proposes a unified elevator brake fault diagnosis network, TaskKD-ScaleNet. This diagnostic network system integrates multi-task learning, multi-scale feature representation, and hierarchical knowledge distillation into a single end-to-end trainable process, avoiding information loss and poor coordination caused by training these methods separately or stitching them together later. The following section combines... Figure 4 The elevator brake fault diagnosis network TaskKD-ScaleNet is described in detail.

[0065] Figure 4 This is a schematic diagram of the architecture of TaskKD-ScaleNet, an elevator brake fault diagnosis network according to an embodiment of the present invention. Figure 4 As shown, the training process of the elevator brake fault diagnosis network includes two stages: the first stage is the main diagnosis network training, and the second stage is the lightweight edge diagnosis network training.

[0066] The first stage of training the main diagnostic network includes: inputting elevator brake fault signals (i.e., the aforementioned PVDF strain signal, acceleration time-domain signal, and high-frequency acoustic emission transient pulse signal) into a multi-scale residual feature extraction module; setting two task output heads based on the extracted shared features: Task A: fault type (identification), and Task B: fault severity (assessment). The cross-entropy (CE) loss is calculated for each task and then summed to obtain the total CE loss for this stage; simultaneously, soft labels with temperature coefficients are generated to provide guidance for subsequent knowledge distillation.

[0067] The second stage of training the lightweight edge diagnostic network involves constructing an edge model with fewer parameters and lower computational cost, still outputting two results: "type" and "severity". During training, on the one hand, the CE loss is calculated using real labels to ensure the model's basic recognition ability; on the other hand, distillation loss is introduced: a soft distribution obtained using temperature softmax is applied to the output layer, and KL divergence is used to make the lightweight edge diagnostic model closely resemble the soft labels of the main diagnostic model; alignment constraints are added to the intermediate feature layer, which specifically refers to the output features of several layers within the multi-scale residual feature extraction module (convolution + residual block) and its multi-scale fusion mapping of the main and edge diagnostic networks, i.e., the feature representation located before global pooling and the task output head, enabling the edge diagnostic network to learn the multi-scale representation method / representation ability of the main diagnostic network at the representation layer. Finally, the CE loss and the hierarchical distillation loss are weighted and fused to form the total loss for edge network training.

[0068] In this embodiment, the elevator brake fault diagnosis network TaskKD-ScaleNet includes: an event localization and slicing module, a window construction and normalization module, a multi-scale residual feature extraction module, a TaskKD-ScaleNet diagnostic network training module, a mechanism scoring and label generation module, and an event-level fusion and diagnostic result output module.

[0069] The multi-scale residual feature extraction module is used to accurately extract multi-modal and multi-scale features based on the temporal characteristics differences of multi-channel signals from elevator brakes. It mainly includes multi-scale parallel convolution branches, residual blocks, and feature fusion and dimensionality reduction components. Specifically, the main diagnostic network uses multi-scale parallel convolution kernels (kernels of different lengths) to cover different time scales. For example, in the multi-scale residual feature extraction module, the acceleration signal is processed through a short-scale branch (convolution kernel k=7, stride s=2, padding p=3) to extract periodic low-frequency vibration features; the PVDF sensor signal is processed through a medium-scale branch (convolution kernel k=15, stride s=2, padding p=7) to capture brake strain features; and the AE signal is processed through a long-scale branch (convolution kernel k=31, stride s=2, padding p=15) to extract high-frequency pulse data features from the brake. Each branch is followed by two residual blocks (kernel k=3, stride s=1, padding p=1) for feature extraction, increasing the number of channels from 32 to 64. The feature maps of the three branches are then concatenated along the channel dimension, resulting in 192 channels. These channels are then mapped to 128 channels via a 1×1 convolution (kernel k=1, stride s=1, padding p=0), and global average pooling is performed to obtain a 128-dimensional feature vector as input for subsequent tasks.

[0070] The three scale branches constructed are represented by the following formula (1):

[0071]

[0072] in, This represents a stacked structure of "convolution + several residual blocks", with different scale branches corresponding to convolution kernels of different lengths to match the temporal scale features of the signal; To synchronously acquire and fuse multi-channel elevator brake fault signals; branch Representing the PVDF signal, during the clamping and releasing process of an elevator brake, changes in clamping force and contact stress induce micro-strain in the clamp body. The PVDF voltage amplitude and fluctuation characteristics of the PVDF sensor can effectively characterize changes in contact conditions such as friction pad wear; branch Representing acceleration signals, when the elevator brake malfunctions, it leads to an enhancement of low-frequency vibration signals and a (significant) change in vibration and impact characteristics. This branch can specifically capture these fault characteristics. Representing the AE signal, when defects such as cracks or wear occur in the brake friction pads, a high-frequency transient pulse signal will be generated. This branch can accurately extract such high-frequency fault characteristics.

[0073] Furthermore, multi-scale feature fusion is achieved through feature concatenation and linear mapping, as shown in the following formula (2):

[0074]

[0075] in, This is a channel-dimensional splicing operation used to merge multi-scale features extracted from three scale branches, achieving preliminary fusion of multimodal features; This is a fusion mapping function, such as a 1×1 convolution operation, used to perform linear mapping and dimensionality reduction on the concatenated high-dimensional features; The shared features obtained after fusion take into account multimodal and multi-scale information.

[0076] The fused shared features are then processed by the pooling operation described in formula (3) to obtain a single-channel feature signal. :

[0077]

[0078] in, This is a pooling operator (global pooling) used to compress feature dimensions, retain key feature information, and reduce subsequent computational complexity.

[0079] Then, the system synchronizes and aligns the pooled signals obtained from the pooling process. The expression for the synchronized multi-channel discrete signal is as follows:

[0080]

[0081] in, The pooled characteristic signals corresponding to the three channels of PVDF strain signal, acceleration signal and AE (acoustic emission) signal are synchronized and used as the input for subsequent fault diagnosis dual tasks.

[0082] In this embodiment, the fault diagnosis dual tasks are Task A (fault type) and Task B (severity), where Task A is represented as:

[0083]

[0084] Among them, 0 indicates normal, 1 indicates jamming fault, and 2 indicates friction plate wear fault;

[0085] Task B is represented as:

[0086]

[0087] In this system, 0 represents normal, 1 represents minor fault, 2 represents moderate fault, and 3 represents serious fault.

[0088] The brake fault information is processed by the network and then output as follows:

[0089]

[0090] in, For window The probability distribution of the corresponding type (brake failure type); It is a three-dimensional real number space; For window The probability distribution of the corresponding severity (severity of brake failure); It is a four-dimensional real number space.

[0091] In this embodiment, the event localization and slicing module performs the following steps S41 and S42: Step S41, using data processing and a three-mechanism consistency algorithm, determines the braking action trigger point (discrete moment). Specifically, if a braking control command for the elevator brake is received, the command edge of the braking control command shall be used as the trigger point for the braking action. If no braking control command is received from the elevator brake, the braking trigger point is determined based on the joint abrupt change detection of PVDF strain fluctuation, acceleration impact response, and AE high-frequency pulse energy. Braking action trigger point The calculation formulas are shown in formulas (8) and (9) below:

[0092]

[0093]

[0094] in, For the first Each channel at discrete time The sampled values; For the first Each channel at discrete time The sampled values; For discrete time The joint mutation detection statistic (characterizing the overall mutation degree of multi-channel signals). The candidate search range for trigger points (must be preset according to the actual scenario, such as the time range in which braking action may occur); To obtain the discrete time index that maximizes the expression.

[0095] Step S42, around the braking action trigger point The expression for capturing an event fragment is:

[0096]

[0097] in, The multi-channel discrete sequence corresponding to the braking event segment can be represented as a sequence with dimension . The matrix (M is the number of channels, L is the number of sampling points for the event segment); X is the complete multi-channel sampling data matrix with dimensions M×N (N is the total number of sampling points); notation This indicates retrieving all channels and time indices from... arrive Slices; For trigger point Take a fixed reserved time forward from the center (for example, 0.2s). For trigger point Centered on the sample, a fixed event duration is taken forward (e.g., 1.0s, which needs to be converted to the number of sampling points based on the sampling rate; for example, when the sampling rate is 1000Hz). =1000).

[0098] In this embodiment, the window construction and normalization module performs the following operations: setting the length of each signal segment to a fixed length, setting the overlap ratio between adjacent samples, and performing zero-mean unit variance standardization.

[0099] For example, let the window length L = 2048 (number of sampling points), and the overlap rate of (adjacent windows) (Overlap ratio), window sliding step size is For event segment signals If a sliding window is used, the sample of the k-th window is represented as follows:

[0100]

[0101]

[0102] in, Let be the sample of the k-th window; k is the window number; the notation [:,ku:ku+L] represents a slice of all channels and time indices from ku to ku+L in the multi-channel matrix.

[0103] Then, for each window Each channel in The standardized input is obtained by standardizing the input according to the following formula (13). :

[0104]

[0105] in, This refers to the nth sampling point in the mth channel of the kth window; This represents the mean of the m-th channel in the k-th window; Let be the sample standard deviation of the m-th channel in the k-th window; The standardized sampling points are obtained by standardizing the m-th sensor channel, the k-th sliding window, and the n-th sampling point. After standardization of all channels, the standardized input of the k-th window is obtained. .

[0106] In this embodiment, the mechanism scoring and label generation module is used to output the jamming score and the friction interface degradation score, and to generate fault type / severity labels based on the jamming score and the friction interface degradation score, supporting model training, online fault interpretation, and maintenance closed loop. The calculation of the jamming score and the friction interface degradation score will be explained in detail below.

[0107] (a) Score for sluggishness calculate

[0108] Score The degree of brake sticking is characterized, with the sticking score primarily determined by acceleration and calculated based on the proportion of low-frequency energy in the acceleration channel window signal. Let... For the acceleration channel window, the low-frequency energy ratio and hysteresis score are calculated according to the following formulas (14) to (16). :

[0109]

[0110]

[0111]

[0112] in, For the first The low-frequency energy of each window represents the acceleration signal within the frequency range. Internal energy and This refers to the frequency range for low-frequency energy calculations, for example. Take 0.5Hz. Take 1Hz; For acceleration standardization signal The frequency domain representation of , that is, the spectrum obtained through Fourier transform; For the first The total energy of the window, that is, the energy of the acceleration signal across all frequency ranges; To prevent division by zero of extremely small constants, for example, a value of 100 can be used. ~ To avoid total energy The value becomes unstable when it is zero.

[0113] (ii) Scoring on the degradation of the friction interface calculate

[0114] The friction interface degradation score is composed of PVDF strain fluctuations and AE high-frequency pulses, characterizing the degree of failure such as wear, uneven contact, and crack propagation at the friction interface. The specific calculation steps are as follows:

[0115] (1) Physical characterization of PVDF signal (linear piezoelectric constitutive model)

[0116] The normalized signal of PVDF (with 3 polarization thickness directions and 1 principal stress direction) satisfies the following linear piezoelectric constitutive relation:

[0117]

[0118]

[0119]

[0120] in, For the first The first window Standardized signals from a single sensor (PVDF); The equivalent stress is along the principal stress line (direction 1); The piezoelectric constant can be -25. The dielectric constant can be taken as 30; This is an approximate voltage, i.e., the equivalent electric field strength in three directions; For PVDF strain; The Young's modulus of PVDF is approximately 2.5 GPa. The sensitivity constant for PVDF can be taken as 0.5. It can be seen that the equivalent stress caused by frictional contact pressure fluctuations and shear strain changes... It will be directly mapped to voltage. The amplitude and spectral characteristics of the material change, thus making it sensitive to wear / contact unevenness.

[0121] (2) Calculation of RMS (Root Mean Square) of PVDF

[0122] The strain fluctuation amplitude of PVDF is quantified by root mean square (RMS) according to the following formula (20):

[0123]

[0124] in, For the first The RMS value of strain fluctuation in each window is used to measure the fluctuation amplitude of the PVDF signal. For the first The first window Standardized signals from a single sensor (PVDF); The window length (usually set to 2048 points) is the number of sampling points contained in each window.

[0125] (3) AE (high-frequency pulse) energy calculation

[0126] The high-frequency energy of the AE signal, which characterizes the micro-cracks and slippage of the friction interface, is calculated according to the following formula (21):

[0127]

[0128] in, For the first The high-frequency energy of each window represents the frequency range of the AE signal. Internal energy and This refers to the frequency range for calculating high-frequency pulse energy. The possible value is 10kHz. The possible value is 100kHz; The square of the spectral amplitude of the AE signal represents the energy of that frequency component.

[0129] (4) Calculation of friction interface degradation score

[0130] The friction interface degradation score is obtained by fusing the PVDF normalized RMS and AE normalized high-frequency energy weighted average using the following formula (22):

[0131]

[0132] in, For the first The friction interface degradation score of each window combines the strain fluctuation of PVDF and the high-frequency pulse characteristics of AE. A high score indicates that the friction interface may have problems such as degradation, wear, crack propagation or slip friction. This is the normalized RMS value of the PVDF signal, used to measure the amplitude of strain fluctuations. Normalized high-frequency energy of the AE signal, used to measure the energy of high-frequency pulses in the AE signal; weights of 0.45 / 0.55 are set based on the sensitivity of the two types of signals to triboelectric degradation.

[0133] Then, based on the lag score Friction interface degradation score Generate fault type / severity labels. For example, fault labels are generated through the following steps:

[0134] (1) Threshold setting

[0135] First, calculate the mean mechanism score of the normal braking window according to the following formula (23). :

[0136]

[0137] in, The mean score of the mechanism.

[0138] Secondly, the threshold is set according to the following formula (24):

[0139]

[0140] in, , These are the normal braking windows. Mean and standard deviation; This is the first threshold for the fault type. This is the second threshold for fault type.

[0141] (2) Generate the tags for task A

[0142] Generate the tags for task A according to the tag generation rules in the following formula (25):

[0143]

[0144] in, For the first Discrete labels for each window on task A. ∈{0,1,2}, which correspond to no fault, stuck fault, and friction interface degradation fault, respectively.

[0145] (3) Generate the tags for task B

[0146] First, the comprehensive anomaly intensity, i.e. the comprehensive score, is defined according to the following formula (26):

[0147]

[0148] Secondly, set the segment threshold according to the following formula (27). :

[0149]

[0150] Among them, the segmented threshold ; This is the threshold for minor faults; The threshold for a moderate fault is [not specified]. This is the threshold for severe faults.

[0151] Then, the labels for task B are generated according to the following formula (28):

[0152]

[0153] in, For the first Discrete labels for each window on task B; This is the segmentation threshold, used to divide different severity intervals.

[0154] Furthermore, the shared features of multiple tasks are extracted according to the following formula (29), and the fault type and severity are output according to the following formulas (30) to (33):

[0155]

[0156]

[0157]

[0158]

[0159]

[0160] in, ( ) represents a shared feature extraction network; This is a shared feature representation used by the output headers of tasks A and B. Output logits (raw score) for Task A (elevator brake failure type); ) represents the output logits of Task B (elevator brake failure severity level); These are the output headers for tasks A and B, respectively; the Softmax function is used to convert the unranged raw scores into normalized probabilities between 0 and 1. , These are the category probability distributions for tasks A and B, respectively.

[0161] In this embodiment, the TaskKD-ScaleNet diagnostic network training module calculates the loss through the following steps and trains the diagnostic network based on the calculated loss:

[0162] (1) Calculate the hard supervision loss (label loss)

[0163] The total hard supervision loss generated after the signal enters the main diagnostic network is a weighted sum of the cross-entropy of the two tasks. The hard label supervision loss is calculated according to the following formulas (34) and (35):

[0164]

[0165]

[0166] in, To diagnose the hard-label supervision loss of the primary diagnostic network, For the true label of task A, The true label for Task B; This represents the cross-entropy loss.

[0167] (2) Definition of temperature Softmax and soft tag

[0168] A temperature factor τ is introduced into the Softmax function, and the temperature Softmax is defined as follows:

[0169]

[0170] in, It is the first Logit (raw score) for each category; It is the logits of all categories; This represents the number of elevator brake failure types output. It is the distillation temperature; the predicted probability. Represented as the first in the Softmax classification task The predicted probability of fault categories. Specifically, let the output probability distribution of the main diagnostic network be... and This is called a soft label, and the output probability distribution of a lightweight edge diagnostic network is as follows: and This is called soft prediction.

[0171] (3) Calculate distillation loss

[0172] Calculate the distillation losses for tasks A and B using the following formulas (37) to (40):

[0173]

[0174]

[0175]

[0176]

[0177] in, , These represent the distillation losses for Task A and Task B, respectively. This is due to distillation losses; , The soft distribution of the main diagnostic network on task A is obtained by temperature softmax. The soft distribution of the lightweight edge diagnostic network on task A is obtained by temperature Softmax; KL is the Kullback-Leibler scattering loss function; This is the distillation loss scaling factor, ensuring that the magnitude of the loss matches the hard-supervised loss.

[0178] (4) Calculate the hard supervision loss for the lightweight edge diagnostic network.

[0179] The hard supervision loss of the lightweight edge diagnostic network is calculated according to the following formula (41):

[0180]

[0181] in, Hard-label supervision loss for lightweight edge diagnostic networks; The true label (fault type) for task A. The true label (severity) for task B.

[0182] (5) Calculate the total loss of knowledge distillation

[0183] First, calculate the total loss of knowledge distillation according to the following formula (42):

[0184]

[0185] in, ∈[0,1] represents the weight; This is the final optimization objective during edge network training.

[0186] Secondly, hierarchical distillation is performed, and L2 constraints are applied to the intermediate feature layers of the main / edge networks to ensure feature distribution alignment. The L2 constraint form for intermediate feature alignment is as follows:

[0187]

[0188] Then, the final total distillation loss is calculated according to the following formula (44):

[0189]

[0190] in, The primary diagnostic network Features of an alignment layer Features of the lightweight edge diagnostic network at the corresponding layer; For the number of layers involved in alignment, generally 2 or 3 layers are selected; The square of the L2 norm; The feature alignment weights.

[0191] Based on the final distillation loss training diagnostic network obtained from the above steps, a lightweight edge diagnostic network / fault diagnosis model that can be deployed on edge devices is obtained. This lightweight edge diagnostic network / fault diagnosis model can be used for fault diagnosis of elevator brakes.

[0192] In this embodiment, the event-level fusion and diagnostic result output module performs the following diagnostic steps:

[0193] First, the aforementioned processing is performed on the collected real-time multi-channel signals of the elevator brake to obtain the standardized real-time multi-channel signals for each window within the current braking event segment. Then, the standardized real-time multi-channel signals are input into a lightweight edge diagnostic network for elevator brake fault classification and severity assessment to obtain the window-level prediction results, i.e., the first prediction result output by the lightweight edge diagnostic network. Probability distribution of window types and the output of the lightweight edge diagnostics network The first window in the severity task class probability A single braking event is processed through windowing to generate a... Window, therefore the fault diagnosis model output Group window-level prediction results.

[0194] Then, event-level fusion is performed on the window-level prediction results to reduce misjudgments caused by single-window noise. Specifically, this includes fault type fusion and fault severity fusion:

[0195] (1) The fault type probabilities are fused according to the following formulas (45) and (46):

[0196]

[0197]

[0198] in, This represents the total number of windows in this event; The final fault type label is determined by averaging the probabilities across all windows. The final failure type prediction for the event is obtained by taking the category containing the maximum value.

[0199] (2) The severity is fused according to the following formula (47):

[0200]

[0201] in, For all windows in this event In the The maximum probability value for a class, here using the window-level maximum value, represents the maximum likelihood of each event in a certain severity category; The final severity label is obtained by finding the category with the highest probability across all windows to obtain the final severity prediction for the event.

[0202] The window-level results are fused into event-level results using the above fusion method. The final fault type and final severity are the fault diagnosis results of the elevator brake.

[0203] The fault diagnosis method described in Embodiment 2 of this invention combines acceleration, PVDF strain, and acoustic emission signals in the hardware section to enhance the sensitivity to typical elevator brake faults such as jamming and wear by using multi-source complementary methods. In the algorithm section, a lightweight diagnostic model that can run on edge devices is obtained by multi-task, multi-scale feature modeling and knowledge distillation. By using "multi-modal physical correspondence (PVDF-wear, AE-wear, vibration-jamming) + mechanism score (PVDF-RMS, AE high-frequency energy, low-frequency energy ratio) + event-level fusion", the model provides recalcible mechanism evidence while outputting the fault type and severity, thereby achieving online identification and severity assessment of brake faults. This maintains good robustness and accuracy even in noisy and fluctuating environments.

[0204]

Example 3

[0205] Embodiment 3 of the present invention provides a specific example of determining the fault diagnosis result of an elevator brake according to the elevator brake fault diagnosis method of Embodiment 2. The example specifically includes the following steps:

[0206] Step 1: Acquire input data via hardware acquisition

[0207] During a single "brake-holding" action, a certain elevator brake collects three fault signals from three sensors: PVDF, acceleration, and AE signals. The sampling frequency is set to... The total acquisition time for a single event is approximately 1.5 seconds; the three channels of signals respectively reflect the strain of the friction pair (PVDF), the structural vibration impact (acceleration) caused by jamming, and the high-frequency transient pulse (AE) caused by friction plate wear.

[0208] Step 2: Multi-scale residual feature extraction of TaskKD-ScaleNet

[0209] To address the differences in frequency characteristics among the three types of input signals, a multi-branch parallel residual feature extraction network (TaskKD-ScaleNet) is designed. Its core structure is "branch convolution + residual block + feature fusion + global pooling". The three branches correspond to the three channels, and the three channels are fed into the three branches to extract features in parallel.

[0210] In this embodiment, the acceleration signal is sent to the short-scale branch, the PVDF signal is sent to the medium-scale branch, and the AE signal is sent to the long-scale branch. The specific parameters of each branch are as follows:

[0211] Acceleration signals are processed through a short-scale branch (kernel k=7, stride s=2, padding p=3) to extract periodic low-frequency vibration features. PVDF sensor signals are processed through a medium-scale branch (kernel k=15, stride s=2, padding p=7) to capture brake strain features. AE signals are processed through a long-scale branch (kernel k=31, stride s=2, padding p=15) to extract high-frequency pulse data features of the brake. Each branch is followed by two residual blocks (kernel k=3, stride s=1, padding p=1) for feature extraction, increasing the number of channels from 32 to 64. The feature maps from the three branches are then concatenated along the channel dimension, resulting in 192 channels. These are then mapped to 128 channels via a 1×1 convolution (kernel k=1, stride s=1, padding p=0) and global average pooling is performed to obtain a 128-dimensional feature vector, which serves as input for subsequent tasks.

[0212] Step 3: Event Location and Slicing

[0213] First, determine the trigger point. Assuming no available control command edge is available, a joint mutation detection method is used to locate the trigger point within the candidate interval: considering the mutation degree of PVDF strain fluctuations, acceleration impacts, and AE high-frequency pulse energy, the time index with the largest joint mutation index is taken as the trigger point. In this example, the maximum mutation point is found to be... =240000, and the corresponding time is: .

[0214] Secondly, determine the left and right boundaries. , This example uses a method of "fixed reserved duration before the trigger point + fixed event duration after the trigger point" to determine the slice boundary. Taking the reserved duration as 0.2s and the event duration as 1.0s, we get: the number of reserved points before the trigger point is... =40000, therefore -40000 = 200000; the number of event points after the trigger point is... =200000, therefore +200000=440000.

[0215] Therefore, the event fragment is obtained as follows =[200000,440000], fragment length is If +1=240001 points, then the PVDF, acceleration, and AE signals will be synchronously extracted according to the index range of [200000, 440000].

[0216] Step 4: Window Construction and Normalization

[0217] The event segment is divided into sliding windows with a window length of L = 2048 points and an overlap ratio of 0.5. The step size is then... =1024. For demonstration purposes, this example only selects four windows closest to the trigger point: Window#1, Window#2, Window#3, and Window#4. The three-channel sequences within each window are standardized with zero mean and unit variance to obtain the network input. ,in, Number the window. .

[0218] Step 5: Obtain window-level probabilities from the outputs of the two tasks.

[0219] Set up two tasks, A and B: Task A (fault type) outputs a category set of {0 Normal, 1 Stuck, 2 Friction plate wear}, and outputs a window-level probability distribution. Task B (Severity) output level set is {0,1,2,3}, output window-level probability distribution. .

[0220] In this example, the edge diagnostic network outputs the following results for the four windows:

[0221] Task A Window Level Probability :

[0222] W1: [0.10, 0.22, 0.68]

[0223] W2: [0.05, 0.15, 0.80]

[0224] W3: [0.08, 0.18, 0.74]

[0225] W4: [0.15, 0.20, 0.65]

[0226] Task B Window Level Probability :

[0227] W1: [0.05, 0.20, 0.55, 0.20]

[0228] W2: [0.02, 0.10, 0.38, 0.50]

[0229] W3: [0.04, 0.18, 0.50, 0.28]

[0230] W4: [0.10, 0.35, 0.40, 0.15]

[0231] Step 6: Calculate the mechanism score used for interpretability and rule-based support.

[0232] This invention synchronously calculates the scoring mechanism during inference to interpret and stabilize the network output. This example provides reproducible numerical values ​​and introduces a minimal constant ϵ to prevent division by zero. .

[0233] (1) Calculate the stuttering score Acceleration low-frequency energy ratio (frequency band) )

[0234] The low-frequency range of the acceleration signal is taken as Perform frequency domain energy statistics for each window, assuming the low-frequency energy is... The total energy is Therefore, according to the aforementioned formulas (14) to (16), the lag scores for each window in this example are calculated as follows:

[0235] W1:

[0236] W2:

[0237] W3:

[0238] W4:

[0239] As can be seen, the proportion of low-frequency energy in W2 / W3 has increased significantly, suggesting that there may be signs of enhanced low-frequency vibration due to obstruction, dragging, or poor return to position.

[0240] (2) Calculate PVDF channel quantities: equivalent stress and approximate voltage, and PVDF-RMS

[0241] The material parameters for the effective working area of ​​PVDF are: Young's modulus. piezoelectric constant dielectric constant And set an approximate proportionality coefficient for the electromechanical conversion. =0.5. Within the window, the fluctuation intensity (e.g., RMS) of the PVDF voltage sequence characterizes the degree of contact pressure fluctuation and strain change. The PVDF-RMS results for each window in this example are as follows:

[0242] W1: PVDF-RMS = 0.62

[0243] W2: PVDF-RMS = 0.95

[0244] W3: PVDF-RMS = 0.80

[0245] W4: PVDF-RMS = 0.55

[0246] It is evident that the PVDF-RMS of W2 / W3 is significantly increased, indicating stronger strain / contact pressure fluctuations during the braking process, which is consistent with the physical intuition of friction interface degradation or uneven contact.

[0247] (3) Calculate AE channel quantity: high-frequency pulse energy (frequency band) )

[0248] In this example, the high-frequency band range of AE is selected as follows: =[10kHz,100kHz], calculate the energy of this frequency band for each window according to formula (21). The following results were obtained:

[0249] W1:

[0250] W2:

[0251] W3:

[0252] W4:

[0253] It is evident that the high-frequency pulse energy of W2 / W3 is significantly enhanced, which is consistent with the phenomenon that increased AE pulses are caused by friction slippage, accelerated wear, or microcracks.

[0254] (4) Calculate the degradation score of the friction interface. PVDF-RMS and AE high-frequency energy fusion

[0255] The PVDF-RMS and AE high-frequency energy are normalized to 0~1. In this example, the maximum value within the event is used for normalization, and the calculation is as follows: =[0.65,1.00,0.84,0.58] (normalized to 0.95); =[0.34,1.00,0.74,0.28] (normalized to 0.88), then perform the equal-weighted fusion calculation formula " "Obtain the friction interface degradation score within each window:"

[0256] W1:

[0257] W2:

[0258] W3:

[0259] W4:

[0260] (5) Overall score based on calculation mechanism

[0261] Overall score for definition mechanism: The overall mechanism score for each window is calculated based on this formula:

[0262] W1:

[0263] W2:

[0264] W3:

[0265] W4:

[0266] Step 7: Threshold Setting and Window-Level Label Generation

[0267] (1) Threshold setting and label A generation

[0268] First, based on formulas (23) and (24), the distribution characteristics of the mechanism's comprehensive score are statistically obtained from historical normal braking data (the specific values ​​of historical normal braking data are omitted here): Then the two thresholds are calculated as follows:

[0269]

[0270]

[0271] Secondly, the discrete labels of window-level task A are determined according to formula (25). The generation rules are as follows:

[0272] like and ,but (Stuck);

[0273] like and ,but (Wear / interface degradation);

[0274] otherwise (normal).

[0275] Substituting the values ​​from this example yields the following result:

[0276] W1: ,and ⇒

[0277] W2: and ⇒

[0278] W3: and ⇒

[0279] W4: and ⇒

[0280] It is evident that the event was detected by the "wear and tear / interface degradation" mechanism in the key windows (W2, W3), accompanied by some signs of lag (although...). Not achieved ,but (The ratio of W2 to W3 increases relatively).

[0281] (2) Combining anomaly intensity with task B label generation

[0282] First, according to formula (26) " The overall anomaly intensity of each window was calculated, and the following results were obtained:

[0283] W1:

[0284] W2:

[0285] W3:

[0286] W4:

[0287] Secondly, based on formula (27) and the aforementioned The segmented thresholds are calculated as follows: .

[0288] Then, the discrete labels of window-level task B are determined according to formula (28). The generation rules are as follows:

[0289] like ,but ;

[0290] like ,but ;

[0291] like ,but ;

[0292] like ,but .

[0293] Substituting the values ​​from this example yields the following result:

[0294] W1:

[0295] W2:

[0296] W3:

[0297] W4:

[0298] Step 8: Training Phase

[0299] This step, based on hard supervision, soft distillation, and intermediate feature alignment constraints, enables transfer training from the main diagnostic network to the edge diagnostic network. This example training employs a two-stage strategy: first, the main diagnostic network (Teacher network) is trained, followed by the lightweight edge diagnostic network (Student network). During the Student network training process, it simultaneously receives hard supervision signals and distillation constraints, ensuring that the performance approximates that of the main network while maintaining a lightweight design.

[0300] 1) Main diagnostic network training: For the target event in this example, the manually labeled fault type is "friction plate wear" (corresponding to task A=2), and the fault severity is "3" (corresponding to task B=3). The main network training adopts a weighted summation of cross-entropy loss between the two tasks to optimize the classification accuracy of the training network for fault type and severity.

[0301] 2) Edge diagnostic network distillation training: Assume the distillation temperature is... The output layer uses a soft distribution of temperature Softmax for KL distillation, and the distillation loss weight is denoted as... Simultaneously, a feature alignment constraint is added to the "intermediate feature layer," and the feature alignment loss weight is denoted as... Alignment layer number K (The example selects two feature layers for alignment: "representative layer features before fusion + features before pooling after fusion"). Taking window W2 as an example, if the Teacher's soft distribution is... The soft distribution of Student is The output layer KL distillation yields a positive loss value (approximately 0.093 in the example), and task B similarly yields a KL value (approximately 0.075 in the example). For feature alignment, if the squared L2 distances of the two aligned layers are 0.40 and 0.28 respectively, the intermediate feature alignment loss is 0.68. The final total loss for Student is calculated as "hard supervision (two-task CE) + ... · Distilled KL + • Feature alignment enables edge networks to approximate the performance of the main network under constraints of parameters and computing power.

[0302] Step 9: Event-level fusion (Window → Events) and output the final diagnostic results.

[0303] (1) Predicting Fault Type (Task A): For each window within the event... Calculate the average, then select the category with the highest probability as the event type. In this example:

[0304] The average result for event level of task A is:

[0305] Normal mean:

[0306] Mean lag:

[0307] Average wear:

[0308] The maximum value is the average wear value. Therefore, the event fault type output is: Task A=2 (friction plate wear).

[0309] (2) Predict the severity (Task B): For each window within the event... Take the maximum value for each category, and then take the largest value among the four severity categories as the severity of the event.

[0310] The maximum event level for Task B is:

[0311] Maximum value for level 0: max(0.05, 0.02, 0.04, 0.10) = 0.10

[0312] Maximum value for Level 1: max(0.20, 0.10, 0.18, 0.35) = 0.35

[0313] Maximum value for level 2: max(0.55, 0.38, 0.50, 0.40) = 0.55

[0314] Maximum value for level 3: max(0.20, 0.50, 0.28, 0.15) = 0.50

[0315] The maximum value appears at level 2 (0.55), therefore, the event severity output is: Task B=2 (Medium).

[0316] Therefore, the final diagnosis is: Fault type: Friction plate wear; Fault severity: Moderate.

[0317] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software combined with a hardware platform. Based on this understanding, all or part of the technical solution of the present invention that contributes to the background art can be embodied in the form of a computer software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0318] Correspondingly, embodiments of the present invention also provide a computer-readable storage medium storing computer-readable instructions or programs thereon. When executed by a processor, the computer-readable instructions or programs cause a computer to perform the following operations, which include the steps included in the fault diagnosis method described in any of the above embodiments, and will not be repeated here. The storage medium may include, for example, an optical disc, a hard disk, a floppy disk, flash memory, magnetic tape, etc.

[0319] In addition, embodiments of the present invention also provide a computer device including a memory and a processor. The memory is used to store one or more computer-readable instructions or programs, wherein the one or more computer-readable instructions or programs, when executed by the processor, can implement the fault diagnosis method described in any of the above embodiments. The computer device may be, for example, a server, a desktop computer, a laptop computer, a tablet computer, etc.

[0320] This invention also provides a computer program product including a computer program containing program code for executing the fault diagnosis method shown in the flowchart. When the computer program product is run in a computer system, the program code enables the computer system to implement the fault diagnosis method provided in the embodiments of this disclosure.

[0321] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on a user's computing device, partially on a user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0322] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A method for diagnosing elevator brake faults based on multi-scale feature distillation, characterized in that, The fault diagnosis method includes: Acquire multi-channel signals related to braking events of the elevator brake; The multi-channel signals are windowed and normalized to obtain standardized multi-channel signals for each window within the braking event segment. The standardized multi-channel signal is input into the multi-scale feature extraction network of the main diagnostic network. The multi-channel and multi-scale features of the feature layer are fused to obtain shared features. Based on the shared features, a dual-task diagnosis of fault type identification and fault severity assessment is performed, and the fault type probability distribution and fault severity probability distribution of each window are output. Based on the probability distribution of fault type and fault severity of each window, a loss function is calculated by combining the real labels of fault type and fault severity, and the main diagnostic network is trained according to the loss function. At the same time, soft labels are generated for the dual-task diagnosis of the main diagnostic network. A lightweight edge diagnostic network corresponding to the main diagnostic network is constructed, and the lightweight edge diagnostic network is trained by distillation based on the real labels and the soft labels to obtain a fault diagnosis model that can be deployed on edge devices. Based on the real-time multi-channel signals of the elevator brake collected in real time, the standardized real-time multi-channel signals of each window within the current braking event segment are obtained; The standardized real-time multi-channel signal is input into the fault diagnosis model, and the fault diagnosis model outputs the probability value of the fault type and the probability value of the fault severity for each window within the current braking event segment. The fault type and severity of the elevator brake are determined based on the fault type probability value and fault severity probability value of each window within the current braking event segment; Among them, acquiring multi-channel signals related to the braking event of the elevator brake includes: The original multi-channel signals of the elevator brake are acquired, including: PVDF strain signal, acceleration time-domain signal, and high-frequency acoustic emission transient pulse signal; The original multi-channel signal is time-aligned, and the time-aligned multi-channel signal is sliced ​​according to the braking trigger point to obtain the multi-channel signal related to the braking event. The multi-scale feature extraction network is a three-branch parallel convolutional residual structure with multi-channel matching. The acceleration time-domain signal is fed into the short-scale branch of the three-branch parallel convolutional residual structure to extract low-frequency vibration features, the PVDF strain signal is fed into the medium-scale branch to obtain strain features, and the high-frequency acoustic emission transient pulse signal is fed into the long-scale branch to extract high-frequency pulse features.

2. The fault diagnosis method as described in claim 1, characterized in that, The fault diagnosis method further includes: If the braking control command of the elevator brake is obtained, the command edge of the braking control command shall be the braking trigger point; If the braking control command of the elevator brake is not obtained, the braking trigger point is determined by joint mutation detection of the multi-channel signals.

3. The fault diagnosis method as described in claim 1, characterized in that, The fault diagnosis method further includes: The lag score is calculated based on the acceleration time-domain signal. The friction interface degradation score is calculated based on the PVDF strain signal and the high-frequency acoustic emission transient pulse signal. The true label is generated based on the stickiness score and the friction interface degradation score combined with a preset threshold.

4. The fault diagnosis method as described in claim 1, characterized in that, The fault types include three categories: normal, jamming, and friction plate wear. The fault severity includes four levels: normal, minor, moderate, and severe.

5. The fault diagnosis method as described in claim 4, characterized in that, The process of determining the fault type and severity of the elevator brake based on the fault type probability value and fault severity probability value of each window within the current braking event segment includes: The fault type probability distribution of all windows within the current braking event segment is fused by mean fusion, and the fault category corresponding to the maximum probability after fusion is taken as the final fault type of the elevator brake. The fault severity probability distribution of all windows within the current braking event segment is categorized by fault severity level, and the maximum probability value at the window level is taken for each fault severity level. Then, the fault severity level corresponding to the largest maximum probability value at the window level for each fault severity level is taken as the final fault severity of the elevator brake.

6. A fault diagnosis system for elevator brakes based on multi-scale feature distillation, characterized in that, Perform the fault diagnosis method as described in any one of claims 1-5.

7. A computer-readable storage medium storing computer-readable instructions, characterized in that, The computer-readable instructions are executed by a processor to implement the fault diagnosis method as described in any one of claims 1-5.

8. A computer device comprising a memory and a processor, The memory stores computer-readable instructions, characterized in that, The processor executes the computer-readable instructions to implement the fault diagnosis method as described in any one of claims 1-5.

9. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the fault diagnosis method as described in any one of claims 1-5.

Images