Elevator running state cross-domain evaluation method fused with passive domain adaptive technology

Abstract

The elevator running state cross-domain evaluation method fusing the passive domain self-adaptive technology comprises the steps that elevator running time sequence data, elevator historical running logs and expert evaluation results are collected and integrated, and a source domain elevator data set is formed; the elevator operation time sequence data are preprocessed, feature extraction is conducted on the elevator operation time sequence data and elevator historical operation logs, and dynamic features and static features are obtained; building an elevator running state evaluation model, and training the model by taking the dynamic characteristics and the static characteristics as input; historical operation logs and elevator operation time sequence data of the elevator in the new scene are collected to serve as target domain data, and dynamic features and static features of the elevator in the new scene are extracted; and a label stability evaluation mechanism is utilized to set pseudo labels for elevator samples of the target domain, and the dynamic characteristics and static characteristics of the elevator in the new scene are adopted to carry out migration training on the elevator running state evaluation model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of elevator state evaluation, and particularly relates to an elevator running state cross-domain evaluation method fusing a passive domain adaptive technology. BACKGROUND

[0002] Under the background of accelerating urban-rural integration, the demand for elevators continues to rise, and the growing number of elevators has brought severe challenges to elevator maintenance work in China. At present, the elevator maintenance work in China is mainly carried out according to relevant national standards and local regulations, and in practice, there are problems of maintenance lag or excessive maintenance. Therefore, combined with the real-time running state of the elevator, developing a scientific and reasonable health recovery strategy has become an inevitable requirement for promoting the healthy development of the elevator industry.

[0003] At present, the research on elevator running state evaluation is still in its infancy. On the one hand, as a typical multi-coupling nonlinear complex system, the evaluation method based on the traditional decision analysis model performs poorly in dealing with the complex nonlinear relationship between the indicators, and there is a problem of excessive dependence on expert subjective experience and untimely model weight updating. On the other hand, the existing research on elevator running state evaluation is mostly limited to a single data source, and it is not feasible to collect running data of all types of elevators for model training in practice, which makes it difficult to ensure the cross-domain evaluation effect of the elevator state evaluation model. SUMMARY

[0004] To solve the above technical problems, the purpose of the present application is to provide an elevator running state cross-domain evaluation method fusing a passive domain adaptive technology.

[0005] The present application provides an elevator running state cross-domain evaluation method fusing a passive domain adaptive technology, comprising:

[0006] Step 1: collect elevator running time series data and integrate elevator historical running logs and expert evaluation results to form a source domain elevator data set;

[0007] Step 2: pre-process the elevator running time series data and extract features from the elevator running time series data and the elevator historical running logs to obtain dynamic features and static features;

[0008] Step 3: build an elevator running state evaluation model, take the dynamic features and the static features as inputs, and take the expert evaluation results as sample labels to train the model;

[0009] Step 4: collect the historical running logs and the elevator running time series data of the elevator under a new scenario as target domain data, and extract the dynamic features and the static features of the elevator under the new scenario according to the method of step 2;

[0010] Step 5: The target domain elevator sample is preset with labels at intervals of a certain period of time by using a label stability evaluation mechanism, and the dynamic characteristics and static characteristics of the elevator in the new scene are used to migrate and train the elevator operation state evaluation model.

[0011] The elevator operation state cross-domain evaluation method fusing the passive domain adaptive technology has the following beneficial effects:

[0012] The method of the present application aims at the pain points of the existing elevator evaluation system "static bias" and the difficulty of traditional models in dealing with the nonlinear relationship between indicators, establishes a dynamic and static combined evaluation index system, covers dynamic characteristics such as starting and stopping in the whole running stage, and breaks through the evaluation limitation of single static standard. And using ResNet embedded with CBAM attention mechanism as the feature fusion layer, through attention screening key indicators and inhibiting redundant information, through the residual module to avoid the loss of core information caused by multiple nonlinear transformations, effectively reduce the difficulty of dealing with the collinearity problem between indicators. At the same time, the passive domain adaptive method based on label reliability is innovatively used, the differential loss training is used to improve the elevator cross-scene evaluation effect, and the limitation of single data source is broken through; and according to the evaluation period, a parameter dynamic updating mechanism is established to avoid static weight deviation, and the long-term applicability of the model in multiple scenes is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The flowchart of the elevator operation state cross-domain evaluation method fusing the passive domain adaptive technology of the present application;

[0014] Figure 2 The comparison chart before and after the pre-processing of the elevator operation time series data of the embodiment of the present application;

[0015] Figure 3 The structure chart of the elevator operation state evaluation model based on Attention-ResNet-MLP of the present application;

[0016] Figure 4a The t-SNE distribution chart of the dynamic characteristics and static characteristics of the embodiment of the present application;

[0017] Figure 4b The t-SNE distribution chart of the fused features of the embodiment of the present application;

[0018] Figure 5 The loss flow chart in the elevator cross-domain evaluation process of the embodiment of the present application;

[0019] Figure 6a The t-SNE distribution chart of the elevator operation state evaluation model based on Attention-ResNet-MLP of the embodiment of the present application before fine-tuning;

[0020] Figure 6bA t-SNE distribution diagram of the Attention-ResNet-MLP-based elevator operation state evaluation model after fine-tuning of an embodiment of the present application. DETAILED DESCRIPTION

[0021] As shown in Figure 1 A fusion elevator operation state cross-domain evaluation method based on passive domain adaptive technology, comprising:

[0022] Step 1: Collecting elevator operation time series data and integrating elevator historical operation logs and expert evaluation results to form a source domain elevator dataset.

[0023] For a specific elevator group in a fixed working scenario, use vibration, sound, and temperature sensors to collect time series data during elevator operation and transmit it to the upper computer. On this basis, further integrate the time series data with elevator historical operation logs and expert evaluation results to form a source domain elevator dataset.

[0024] The elevator operation time series data refers to a set of continuous signals collected by various sensors around key components such as brakes and traction machines, as well as core systems such as door systems and car systems during elevator operation. It includes: brake shoe temperature signals (directly reflecting the working state of the friction pair), coil current signals (reflecting the stability of the brake circuit excitation), traction machine motor end vibration signals, stator current signals, traction wheel speed signals, car door vibration signals, door machine current signals, car vibration signals, car acceleration signals, and noise signals.

[0025] The elevator historical operation logs include: elevator basic identification, fault and maintenance records, and inherent technical parameters; wherein the elevator basic identification includes: equipment model and installation time. The fault and maintenance records include: fault information and post-maintenance information during the past operation of the elevator, such as key component failure frequency and maintenance time, etc. The inherent technical parameters refer to the parameters strictly controlled by the state during the delivery and installation of the elevator, including: rated load, rated speed, balance coefficient, and sill height.

[0026] Step 2: Preprocessing the elevator operation time series data and extracting features from the elevator operation time series data and the elevator historical operation logs to obtain dynamic and static features.

[0027] In practical implementation, to address the noise interference and inherent trend term interference in some signals mixed into the elevator running sequence data, this invention employs a signal preprocessing method based on variable mode extraction and detrended fluctuation analysis to effectively remove noise components and trend term interference from the signal. Based on this, and in accordance with national standards, local regulations, and industry specifications, the invention extracts dynamic features characterizing the elevator's operating status from the preprocessed elevator running sequence data, and static features reflecting the elevator's health status from the elevator's historical operating logs. The specific implementation steps are as follows:

[0028] Step 2.1: Using the variational mode extraction algorithm, the elevator running sequence data is adaptively decomposed into multiple non-overlapping eigenmode components with definite center frequencies, specifically:

[0029] Step 2.11: Decompose according to the preset scale Perform VME decomposition, using each modal component Minimizing the sum of frequency domain bandwidths is the objective function. Superimposed and restored to As constrained, the following constrained variational mode extraction model is constructed:

[0030]

[0031] in, For Hilbert transform kernel, It is the Dirac function. Modal components The center frequency, through with Performing convolution operations will It is converted into an analytic signal (containing only the positive frequency domain portion) to facilitate frequency domain analysis. For time-domain differentiation operators, As a frequency shift operator, based on the derivative property of the Fourier transform, the time-domain derivative operator can apply a "frequency weighting" to the spectrum of an analytic signal; based on the shift property of the Fourier transform, the frequency shift operator can shift the spectrum of an analytic signal from its center frequency. The signal is shifted to the origin in the frequency domain. At the same time, according to the Parseval theorem of Fourier transform, the frequency domain bandwidth (second-order center distance of the spectrum) of the signal can be reflected by the time domain signal energy after differentiation and frequency shift modulation.

[0032] Step 2.12: Based on the constrained variational mode extraction model, introduce a quadratic penalty factor. Lagrange multiplication operators This transforms the constrained variational problem into an unconstrained variational problem:

[0033]

[0034] Step 2.13: Update sequentially by alternating direction method , , to solve the above equations iteratively;

[0035] First, fix , , update , get the update expression for:

[0036]

[0037] wherein is the kth modal component after the n+1th iteration.

[0038] By Parseval Fourier equidistant transformation, convert the above formula to the frequency domain, replace with , and transform it into a half-space integral form on non-negative frequency, to obtain the update expression of each modal component in the frequency domain:

[0039]

[0040] wherein is the Fourier transform of the kth modal component after the n+1th iteration; , , respectively represent the Fourier transform of , , .

[0041] Then, fix , update , get the center frequency of each component update expression:

[0042]

[0043] Finally, fix , , update , get expression:

[0044]

[0045] In the formula, is the allowable parameter for noise, , , respectively represent the inverse Fourier transform of , , .

[0046] When the sum of the differences of all modal components of two adjacent iterations meets the following termination condition , , , the update iteration process of is exited:

[0047]

[0048] At this time, the inverse Fourier transform result of is the adaptive decomposition result of the elevator operation time series data .

[0049] Step 2.14: preset multiple decomposition scales Repeat the VME decomposition process of steps 2.11-2.13 on the elevator operation time series data , and then determine the final decomposition scale according to the center frequency distribution result of each group , and output the corresponding adaptive decomposition result of the elevator operation time series data .

[0050] Step 2.2: use the detrended fluctuation analysis method to calculate the scaling exponent of the above decomposed intrinsic mode components, specifically:

[0051] Step 2.21: calculate the cumulative deviation of the intrinsic mode component :

[0052]

[0053] Where T is the sequence length, is the mean value.

[0054] Step 2.22: divide into non-overlapping windows , each window contains S data, that is , assuming that each interval has a trend related to time t of order p, then the trend equation in each window is expressed as:

[0055] =1,2,..., , =1,2,...,

[0056] ​​​​Wherein, by least square fitting to the data of the window, the trend equation coefficient of the window is obtained .

[0057] Step 2.23: eliminating the trend item of each window time sequence , to obtain the time sequence , the second order wave function of the time sequence is calculated :

[0058]

[0059]

[0060] Step 2.24: the window size S is increased according to a certain step, and steps 2.22 and 2.23 are repeated to obtain the second order wave function curve varying with the window size S, if the curve obeys the power law, that is, there is:

[0061]

[0062] It is shown that has self-similar fractal characteristics, the logarithm of and is calculated respectively, to obtain and are linearly related, and the least square method is used to obtain:

[0063]

[0064] Wherein, the slope is the scaling exponent of , is the coefficient to be solved.

[0065] Step 2.3: combined with the characteristics of the elevator running signal, the scaling exponent threshold is set to 0.8, and the intrinsic mode components with the scaling exponent greater than the scaling exponent threshold are removed, and at the same time, for the remaining intrinsic mode components, the wavelet denoising algorithm is used for processing, which can filter out the noise interference to the maximum extent while retaining the effective information contained therein. Finally, the remaining intrinsic mode components after denoising are directly added for reconstruction, and the preprocessing work of the elevator running data is completed. The comparison before and after signal processing is shown in Figure 2 .

[0066] Step 2.4: According to national standards, local regulations and industry specifications, on the one hand, from the pretreated elevator operation time series data, dynamic features that can represent the elevator operation state are extracted, such as brake shoe temperature rise rate, effective value of brake coil current signal, peak-to-peak value of traction motor end vibration signal, effective value of stator current signal, gravity frequency of door system car door vibration signal, start-stop peak value of door machine current signal, root mean square of car system vibration signal, A95 acceleration of acceleration signal, and equivalent continuous A sound level and peak sound pressure level of noise signal.

[0067] Step 2.5: Extract static features that can reflect the health status of the elevator from the elevator historical operation log, including: fault frequency, daily average operation frequency, single day maximum operation frequency, average maintenance interval, service life of equipment, landing accuracy, rated load, rated operating speed and balance coefficient.

[0068] In this embodiment, the operation time series data and historical operation log of the elevator group of an office building are collected, and the evaluation results are given by experts in the industry to form the source domain elevator dataset. Part of the dynamic and static feature extraction results are shown in Table 1.

[0069] Table 1 Dynamic and static feature records of elevator group of an office building

[0070]

[0071] Step 3: Build an elevator operation state evaluation model, with dynamic features and static features as input, and expert evaluation results as sample labels to train the model.

[0072] In specific implementation, the elevator operation state evaluation model based on Attention-ResNet-MLP takes the residual network embedded with attention mechanism as the feature fusion layer, and takes MLP as the classification layer.

[0073] The feature fusion layer takes the residual module-CBAM attention module as the basic unit, and takes the output of the previous basic unit as the input of the next basic unit to form a chain structure of "unit 1→ unit 2→…→ unit N". At the same time, through the step-by-step increase of the channel number of the convolution layer in the basic unit (for example, from 64→128→256→512), the feature dimension is gradually improved.

[0074] In the residual module, a dual-path structure with a main branch and a shortcut branch running in parallel is adopted: the main branch achieves nonlinear feature transformation through a series connection of convolutional layers → batch normalization layers → activation functions, enhancing the model's ability to handle complex relationships between multi-dimensional evaluation indicators of elevators; the shortcut branch is flexibly designed based on whether the feature dimensions of the module's input and output match. When the number of channels and feature sizes of the input and output are consistent, the branch directly connects the module's input and output, passing the original features downstream; when the dimensions do not match, the branch introduces a convolutional layer to adjust the input feature dimensions, ensuring that they are element-wise added to the output features of the main branch. The specific mathematical expression is as follows:

[0075]

[0076]

[0077] in, and They represent the first Inputs and outputs of each residual module. This is the residual mapping function. When the model runs to the redundant module, this function is set to zero, so that the redundant module is identically mapped to the output of the previous module. Represents a linear projection matrix, when and When dimensions are different, through Change Dimension.

[0078] In the attention module, the feature map output by the residual module is first input into the channel attention module to generate a channel attention weight vector. Then, the channel attention weight vector is multiplied element-wise with the feature output by the residual module to obtain the channel weighted feature, thereby quantifying the independent contribution of each feature channel to the elevator evaluation result. Subsequently, the channel weighted feature is input into the spatial attention module to generate spatial attention weights, and then multiplied element-wise with the spatial attention weights to obtain the final feature. This module can adaptively learn the importance of different features within the same channel in the elevator operation status evaluation. The specific mathematical expression is as follows:

[0079]

[0080]

[0081] In the formula, and These are global average pooling and global max pooling operations, respectively. For activation function, It is a multilayer perceptron. For convolution operations, The residual module outputs features. attention for a channel, weighting features for a channel, attention for a space.

[0082] In the training process of the elevator operation state evaluation model based on Attention-ResNet-MLP, label smoothing cross entropy is used as a loss function. During training, the dynamic features and static features of the elevator are used as model inputs, and the expert evaluation results are used as elevator sample labels. The elevator sample labels are converted from original one-hot labels to smooth labels. Then, by minimizing the cross entropy loss between the smooth labels and the output results of the model Softmax layer, the model parameter iterative update is driven until the loss converges.

[0083]

[0084] wherein, is the number of elevator samples, is the number of elevator state categories, is the elevator smooth label, which is related to the set smoothing coefficient, is the output result of the model Softmax layer; the elevator state includes four categories of excellent, good, general and poor.

[0085] The present application selects label smoothing cross entropy as the loss function, and introduces a small disturbance to the real label to effectively guide the model to learn more robust feature representation. At the same time, in order to enhance the perception ability of the model to the feature area, the present application introduces the CBAM attention mechanism in the residual module, that is, the normalized output result of a certain convolution layer of the residual module is multiplied by the attention weight vector generated by CBAM element by element, and then the product result is directly combined with the normalized output of the convolution layer through residual connection as the input of the subsequent module. The model structure is shown in Figure 3 The network structure and parameters of the feature fusion layer are shown in Table 2.

[0086] Table 2 Network structure and parameters of the model feature fusion module

[0087]

[0088] The model parameter adjustment mechanism based on the model performance takes the alignment degree of the actual elevator fault and the poor result of the model evaluation in each period as the trigger basis of parameter adjustment. When the alignment degree does not reach the preset standard and the parameter adjustment is triggered, in order to effectively reduce the safety risk of model false negative faults, the proportion of fault class loss needs to be increased in a targeted manner to ensure that it occupies a dominant position in the total loss function, so as to guide the model parameter update process to be more inclined to minimize the error related to the fault class.

[0089] In order to verify the effect of the model feature fusion layer, the t-SNE dimension reduction algorithm is used to visualize the elevator dynamic and static feature vectors and the model fusion feature vectors, and the results are shown in Figure 4a and 4b As shown in Figure 4a The results show that in the t-SNE distribution of dynamic and static features, there are obvious overlapping areas between different categories of samples, and the boundaries between categories are blurred. Figure 4b Looking back at the model fusion features, the feature clusters of different categories of samples are obviously expanded, and the class discrimination is greatly improved. This result fully proves the effectiveness of the synergistic effect of the loss function, attention mechanism and residual module designed in the present application.

[0090] Step 4: Collect the historical running log and elevator running time series data of the elevator in the new scene as the target domain data, and extract the dynamic and static features of the elevator in the new scene according to the method of step 2.

[0091] In this embodiment, the running time series data and historical running log of a certain residential elevator group are collected to form a target domain elevator dataset, which lacks state labels given by experts. Part of the dynamic and static feature extraction results are shown in Table 3.

[0092] Table 3 Dynamic and static feature records of a certain residential elevator group

[0093]

[0094] Step 5: Use the label stability evaluation mechanism to preset labels for the elevator samples in the target domain every certain period of time, and use the dynamic and static features of the elevator in the new scene to migrate and train the elevator running state evaluation model, specifically:

[0095] Step 5.1: Design a label stability evaluation mechanism to set pseudo-labels for the elevator samples in the target domain.

[0096] In specific implementation, the dynamic and static features of the residential elevator are input into the evaluation model trained based on the elevator data of a certain office building to generate preset pseudo-labels for the corresponding residential elevator. Then, based on the original elevator dataset, a data enhancement sample set is constructed by introducing white noise of different intensities, and the dynamic and static features of the sample set are input into the evaluation model trained based on the elevator data of a certain office building to generate preset pseudo-labels for each enhanced sample. Finally, the label stability is determined by consistency test: if the preset pseudo-labels of the original sample and the preset pseudo-labels of the sample after data enhancement are consistent for more than half, the original sample is defined as a stable sample; otherwise, it is determined as an unstable sample. The specific steps are as follows:

[0097] Step 5.11: Perform D different data enhancement transformations on each elevator sample in the target domain to obtain an expanded dataset , is denoted as:

[0098]

[0099] wherein, is a set of original elevator samples of a target domain, including elevator running time series data and elevator historical running logs, is denoted as is a set of elevator samples obtained by D different data augmentation methods, the number of samples of is consistent with

[0100] Step 5.12: calculating the class feature centers of each sample in the set of original elevator samples :

[0101]

[0102] wherein, is the number of original elevator samples of a target domain, is a fusion feature vector of an elevator sample of a target domain obtained by a feature fusion layer of an elevator running state evaluation model based on Attention-ResNet-MLP, is a probability that the elevator state of a sample obtained by a classification layer belongs to a class c.

[0103] After obtaining the class feature centers of each sample in the set , according to the similarity measurement results of a sample in the set and the class sample feature centers , a preset pseudo label of the sample is obtained:

[0104]

[0105]

[0106] wherein, is a preset pseudo label of an original elevator sample of a target domain, is a cosine similarity between a fusion feature vector of a sample and a feature center , is a Euclidean norm; a threshold determines whether to assign a preset pseudo label to a sample, i.e., when the maximum cosine similarity of each class feature center is lower than​​ When the value is unreliable, it is labeled as -1; otherwise, it is assigned the category label corresponding to the maximum cosine similarity. Each original elevator sample... The preset pseudo-labels constitute the preset pseudo-label set of the original elevator sample. The same method as in step 5.12 is used to obtain the preset pseudo-label set for each data-augmented elevator sample. .

[0107] Step 5.13: Calculate the set based on Step 5.12 Preset pseudo-tag set :

[0108]

[0109] Use majority voting strategy on the pre-set pseudo-label set Aggregation is performed to obtain the original elevator sample set for the final target domain. Middle Elevator Sample pseudo-tags It was identified as:

[0110]

[0111] Where D represents the number of data augmentation methods used in the voting process. It is used to determine the set medium sample Preset pseudo tags Whether it is a category c indicator function; if the condition is true, the output is 1; otherwise, it is 0. Samples at corresponding positions in the set The pre-defined pseudo-labels are consistent with each other for more than half of them, indicating that the elevator sample set in the target domain is consistent. Samples at corresponding positions If stable, use the preset pseudo-label as the category label for the sample; otherwise, label it as -1.

[0112] Step 5.2: Set different loss functions for samples with different stability types, and train the elevator operation status evaluation model based on the Attention-ResNet-MLP model built on the source domain data using the target domain data. Fine-tune the parameters of the Attention-ResNet-MLP elevator operation status evaluation model using the corresponding loss functions, specifically:

[0113] Step 5.21: Evaluate the stability of the elevator samples in the target domain using the label stability evaluation method in step 5.1, divide the samples into stable sample group and unstable sample group; at the same time, generate corresponding pseudo labels for stable samples; for unstable samples, mark them as "-1".

[0114] Step 5.22: For the stable sample group, use the label smoothing cross entropy loss and information maximization loss co-training to iteratively update the network parameters of the feature fusion module and classification module of the model; at the same time, introduce cohesion-repulsion loss separately, only for targeted training and optimization of the parameters of the feature fusion module. The cohesion-repulsion loss function can effectively enhance the clustering effect of similar sample features, and through multi-loss collaborative optimization, the model evaluation results and reliable sample categories are accurately aligned.

[0115] Step 5.23: For the unstable sample group, use the entropy maximization loss to synchronously train and update the network parameters of the feature fusion layer and classification layer of the model, and maximize the evaluation uncertainty of the model for unreliable samples to avoid the interference of unreliable information on model parameter update. The mathematical expressions of the losses are as follows, and the loss flow in the model parameter update process is shown in Figure 5 .

[0116]

[0117] wherein, is the elevator sample, is the target domain sample dataset, is the conditional activation function, which is zero when the elevator sample pseudo label is -1, and the rest is 1. is the smooth pseudo label of the elevator sample , is the probability that the model considers the sample to be of class .

[0118]

[0119]

[0120]

[0121] represents the average probability of all samples belonging to class in the target domain data after being processed by the model.

[0122]

[0123] For a sample set with stable labels, In order to be in line with the current A stable set of target samples with pseudo-labels of the same category. For the present A stable set of target samples that do not have pseudo-labels of the same category. For the sample The fusion characteristics This indicates the inner product operation.

[0124]

[0125] To verify the cross-domain evaluation effect of the fine-tuned elevator operation status evaluation model of this invention, the sample fusion feature distribution before and after model fine-tuning was visualized and analyzed using the t-SNE dimensionality reduction algorithm. The visualization results show that in the t-SNE distribution map of the fused features before model fine-tuning, there is significant overlap in the features of samples from different categories, and the inter-class boundaries are blurred. Figure 6a As shown in Figure 6b, in the t-SNE distribution diagram of the fused features after model fine-tuning, the interval between feature clusters of samples from different categories is significantly increased, and the inter-class discrimination is substantially improved. This result fully demonstrates that the method used in this invention to fine-tune the model through the label stability assessment mechanism and the differential loss training strategy can significantly enhance the model's ability to discriminate elevator operating status features under different scenarios.

[0126] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-domain evaluation method for elevator operating status integrating passive domain adaptive technology, characterized in that, include: Step 1: Collect elevator runtime sequence data, integrate it with elevator historical operation logs and expert evaluation results to form a source domain elevator dataset; Step 2: Preprocess the elevator running sequence data and extract features from the elevator running sequence data and elevator historical running logs to obtain dynamic and static features; Step 3: Construct an elevator operation status evaluation model based on Attention-ResNet-MLP, using dynamic and static features as inputs and expert evaluation results as sample labels, to train the model; Step 4: Collect the elevator's historical operation logs and elevator operation sequence data under the new scenario as target domain data, and extract the dynamic and static features of the elevator under the new scenario according to the method in Step 2; Step 5: At regular intervals, use the label stability assessment mechanism to pre-label the elevator samples in the target domain, and use the dynamic and static features of the elevator in the new scenario to transfer train the elevator operation status evaluation model based on Attention-ResNet-MLP.

2. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 1, characterized in that: The elevator operation sequence data includes: brake shoe temperature signal, coil current signal, traction machine motor end vibration signal, stator current signal, traction sheave speed signal, door vibration signal of the door system, door operator current signal, car vibration signal of the car system, car acceleration signal, and noise signal. The elevator's historical operation log includes: elevator foundation identification, fault and maintenance records, and inherent technical parameters; among which, elevator foundation identification includes: equipment model and installation time; fault and maintenance records include: fault information and subsequent maintenance information during past operation of the elevator; inherent technical parameters include: rated load capacity, rated speed, balance coefficient, and sill height.

3. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: Using the variational mode extraction algorithm, the elevator running sequence data is adaptively decomposed into multiple non-overlapping intrinsic mode components with definite center frequencies; Step 2.2: Calculate the scaling exponent for the intrinsic mode components of the above decomposition using the detrended fluctuation analysis method; Step 2.3: Set the scaling exponent threshold, remove intrinsic mode components with scaling exponents greater than the scaling exponent threshold, and process the remaining intrinsic mode components using a wavelet denoising algorithm. Then, directly add the remaining denoised intrinsic mode components to reconstruct the data, thus completing the preprocessing of the elevator operation data. Step 2.4: Extract the following from the preprocessed elevator running sequence data: brake shoe temperature rise rate, effective value of coil current signal, peak-to-peak value of traction machine motor end vibration signal, effective value of stator current signal, center-of-gravity frequency of door system car door vibration signal, start-stop peak value of door machine current signal, root mean square of car system vibration signal, A95 acceleration / deceleration of acceleration signal, and equivalent continuous A-weighted sound level and peak sound pressure level of noise signal. Step 2.5: Extract static features that reflect the health status of the elevator from the elevator's historical operation log, including: failure frequency, average daily operation frequency, maximum daily operation frequency, average maintenance interval, equipment service life, leveling accuracy, rated load capacity, rated operating speed, and balance coefficient.

4. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 3, characterized in that, Step 2.1 specifically involves: Step 2.11: Decompose according to the preset scale Perform VME decomposition, using each modal component Minimizing the sum of frequency domain bandwidths is the objective function. Superimposed and restored to As constrained, the following constrained variational mode extraction model is constructed: in, For Hilbert transform kernel, It is the Dirac function. Modal components The center frequency, For time-domain differentiation operators, For frequency shifting operators; Step 2.12: Based on the constrained variational mode extraction model, introduce a quadratic penalty factor. Lagrange multiplication operators This transforms the constrained variational problem into an unconstrained variational problem: Step 2.13: Update sequentially using the alternating direction multiplier method , , To solve the above equations iteratively; First, fix , ,renew ,get The update expression is: in, It is the kth modal component after the (n+1)th iteration; By using Parseval Fourier isometric transform, the above equation is converted to the frequency domain. Replace with Then, transforming it into a half-space integral form at non-negative frequencies, we obtain the update expressions for each modal component in the frequency domain as follows: in, It is the Fourier transform of the k-th modal component after the (n+1)th iteration; , , They represent , , Fourier transform; Then, fix ,renew The center frequencies of each component are obtained. The update expression is: Finally, fix , ,renew ,get The expression is: In the formula, These are the tolerance parameters for noise. , , They represent , , The inverse Fourier transform; The termination condition is determined when the sum of the differences of all modal components in two consecutive iterations satisfies the following condition. At that time, jump out , , The update and iteration process: at this time, Inverse Fourier transform results That is, elevator running sequence data Adaptive decomposition results; Step 2.14: Preset multiple sets of decomposition scales Elevator running sequence data Repeat steps 2.11-2.13 of the VME decomposition process, and then according to each group The center frequency distribution results determine the final decomposition scale. It also outputs the corresponding elevator runtime sequence data. Adaptive decomposition results .

5. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 4, characterized in that, Step 2.2 specifically involves: Step 2.21: Calculate the intrinsic modal components Cumulative deviation: Where T is Sequence length, yes Mean; Step 2.22: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would Divided into equal parts Non-overlapping windows Each window contains S data items, that is... Assuming each interval has a p-th order trend with respect to time t, then the trend equation within each window is... Expressed as: =1,2,..., , =1,2,..., The trend equation coefficients for this window are obtained by performing a least-squares fit on the data within that window. ; Step 2.23: Eliminate the trend term in each window of the time series. To obtain the time series Calculate time series Second-order wave function : Step 2.24: Increment the window size S by a certain step size, and repeat steps 2.22 and 2.23 to obtain the second-order oscillation function. A curve that varies with window size S, if the curve follows a power law relationship, i.e., it exists: show It has self-similar fractal characteristics, calculate respectively and The logarithm of the equation yields the following result. and The relationship is linear, and can be obtained using the least squares method: The slope is scaling index , These are the coefficients to be solved.

6. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 1, characterized in that, In step 3, the elevator operation status evaluation model based on Attention-ResNet-MLP uses a residual network with an embedded attention mechanism as the feature fusion layer and an MLP as the classification layer. The feature fusion layer uses the residual module-CBAM attention module as the basic unit and takes the output of the previous basic unit as the input of the next basic unit. The feature dimension is gradually increased by increasing the number of channels in the convolutional layer in the basic unit. In the residual module, a dual-path structure with the main branch and the shortcut branch running in parallel is adopted: the main branch realizes nonlinear transformation of features through the convolutional layer → batch normalization layer → activation function in series connection, which enhances the model's ability to handle the complex relationship between multiple evaluation indicators of elevators; the shortcut branch is flexibly designed according to whether the feature dimensions of the module input and output match. When the number of channels and feature size of the input and output are the same, the branch directly connects the module input and output and passes the original features downstream. When the dimensions do not match, the branch introduces a convolutional layer to adjust the dimensions of the input features, ensuring that they are element-wise added to the output features of the main branch; In the attention module, the feature map output by the residual module is first input into the channel attention module to generate the channel attention weight vector. Then, the channel attention weight vector is multiplied element-wise with the feature output by the residual module to obtain the channel weighted feature, thereby quantifying the independent contribution of each feature channel to the elevator evaluation result. Subsequently, the channel weighted feature is input into the spatial attention module to generate the spatial attention weight, and then multiplied element-wise with the spatial attention weight to obtain the final feature.

7. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 1, characterized in that, In step 3, during the training process, label smooth cross-entropy is used as the loss function. During training, the dynamic and static features of the elevator are used as the model input, and the expert evaluation results are used as the elevator sample labels. The elevator sample labels are transformed from the original one-hot labels to smooth labels. Then, by minimizing the cross-entropy loss between the smooth labels and the output of the model's Softmax layer, the parameters of each module of the model are driven to be updated iteratively until the loss converges. in, For the number of elevator samples, Number of elevator status categories The elevator smoothing label is related to the set smoothing coefficient. The output of the model's Softmax layer is as follows: the elevator status includes four categories: excellent, good, average, and poor.

8. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 1, characterized in that, Step 5 specifically involves: Step 5.1: Design a label stability evaluation mechanism to set pseudo-labels for elevator samples in the target domain; Step 5.2: Set different loss functions for samples with different stability types, and train the elevator operation status evaluation model based on the Attention-ResNet-MLP model built on the source domain data using the target domain data. Fine-tune the parameters of the Attention-ResNet-MLP elevator operation status evaluation model using the corresponding loss function.

9. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 8, characterized in that, Step 5.1 specifically involves: Step 5.11: Perform D different data augmentation transformations on each elevator sample in the target domain to obtain the extended dataset. , is represented as: in, This is the original elevator sample set for the target domain, including elevator runtime sequence data and historical elevator operation logs. express Elevator sample sets obtained through D different data augmentation methods The number of samples and Consistent; Step 5.12: Calculate the original elevator sample set The various feature centers of each sample : in, The original number of elevator samples in the target domain. This represents elevator samples in the target domain obtained through the feature fusion layer of the Attention-ResNet-MLP elevator operation status evaluation model. The fused feature vector, Indicates sample The probability that the elevator state belongs to class c is obtained through the classification layer; According to the set medium sample With various sample feature centers Similarity measurement results yield samples Preset pseudo tags : in, Original elevator samples for the target domain Preset pseudo tags, Indicates sample fused feature vector With feature center Cosine similarity between them Represents the Euclidean norm; threshold Decide whether to assign a pre-defined pseudo-label to the sample, i.e., when The maximum cosine similarity with various feature centers is lower than When the value is unreliable, it is labeled as -1; otherwise, it is assigned the category label corresponding to the maximum cosine similarity. Each original elevator sample... The preset pseudo-labels constitute the preset pseudo-label set of the original elevator sample. The same method as in step 5.12 is used to obtain the preset pseudo-label set for each data-augmented elevator sample. ; Step 5.13: Calculate the set based on Step 5.12 Preset pseudo-tag set : Use majority voting strategy on the pre-set pseudo-label set Aggregation is performed to obtain the original elevator sample set for the final target domain. Middle Elevator Sample pseudo-tags It was identified as: Where D represents the number of data augmentation methods used in the voting process. It is used to determine the set medium sample Preset pseudo tags Whether it is a category c indicator function; if the condition is true, the output is 1; otherwise, it is 0. Samples at corresponding positions in the set The pre-defined pseudo-labels are consistent with each other for more than half of them, indicating that the elevator sample set in the target domain is consistent. Samples at corresponding positions If stable, use the preset pseudo-label as the category label for the sample; otherwise, label it as -1.

10. The elevator operation status cross-domain evaluation method integrating passive domain adaptive technology according to claim 8, characterized in that, Step 5.2 specifically involves: Step 5.21: Use the label stability assessment method in Step 5.1 to assess the stability of the elevator samples in the target domain, and divide the samples into stable sample groups and unstable sample groups; at the same time, generate corresponding pseudo labels for stable samples; for unstable samples, mark them uniformly as "-1"; Step 5.22: For stable sample groups, use label-smoothed cross-entropy loss. With information maximization loss Collaborative training iteratively updates the network parameters of the model's feature fusion and classification modules; simultaneously, a cohesion-repulsion loss is introduced separately. Only the parameters of the feature fusion module are specifically trained and optimized; Step 5.23: For unstable sample groups, use entropy maximization loss. The network parameters of the feature fusion layer and the classification layer of the model are trained and updated synchronously.

Images