Knowledge-guided automatic cleaning method for elevator vibration data

Abstract

The application relates to the field of elevator data analysis, and particularly discloses a knowledge-guided elevator vibration data automatic cleaning method, which comprises a training stage and a test stage. In the training stage, normal monitoring data is collected as training data; a convolutional autoencoder based on discrete wavelet transform is constructed to train the normal monitoring data and obtain a signal reconstruction model; in the test stage, the signal reconstruction model is used to calculate reconstruction errors of vibration monitoring data; according to a set reconstruction error threshold, the vibration monitoring data is judged to be normal monitoring data or abnormal monitoring data; spectrum analysis is carried out on the abnormal monitoring data to calculate a fault characteristic frequency theoretical value; if the detection amplitude peak value in the fault characteristic frequency theoretical value plus or minus 5% interval is within a normal level, the equipment is marked as normal equipment; otherwise, the equipment is a fault equipment; the vibration monitoring data corresponding to the normal equipment is time-sliced, the absolute average value of each time period is calculated, and the vibration monitoring data abnormal type is judged according to a box plot rule.

Description

Technical Field

[0001] This invention relates to the field of elevator data analysis, and more specifically, to a knowledge-guided automatic cleaning method for elevator vibration data. Background Technology

[0002] As a special type of equipment, the health status of elevators directly affects the safety of people's travel. Industry and academia have conducted research on elevator condition monitoring and intelligent operation and maintenance algorithms, achieving fruitful results and effectively ensuring the safe service of elevators. However, due to the "garbage in, garbage out" characteristic of artificial intelligence algorithms, they heavily rely on high-quality condition monitoring data. However, the uncertainty of the elevator service environment and the stability of sensor operation mean that the actual collected data often contains a large number of outliers, leading to decreased algorithm accuracy and frequent false alarms and missed alarms, increasing the difficulty of identifying elevator health status. Therefore, outlier identification and cleaning are the first steps in conducting efficient and high-quality intelligent operation and maintenance. Currently, the identification of outliers mainly utilizes statistical and artificial intelligence methods to perform statistical feature analysis on the data, including:

[0003] For data anomalies such as missing or duplicate values, simple interpolation for imputation and deletion are commonly used. Some researchers have attempted to recover missing values ​​using methods such as tensor decomposition, but these methods often have poor accuracy. Since such data anomalies are easy to identify in actual elevator operation and can be replaced by new monitoring data obtained through subsequent sampling, they will not be discussed here.

[0004] For abnormal mutation values, statistical methods such as the Raida criterion, absolute median difference, and box plots can be used for identification, as well as artificial intelligence methods such as clustering. However, these methods have limited applicability and are limited to specific scenarios, making them unsuitable for handling the complexities of elevator monitoring data cleaning. The main reasons for this are:

[0005] (1) The scenarios in which elevators generate abnormal monitoring data are more diverse. In addition to the scenarios mentioned above, there are also typical scenarios such as low-frequency interference and sensor detachment. Existing methods have failed to cover all types of sensor data anomalies.

[0006] (2) When elevator components malfunction or degrade, the amplitude of the monitoring signal often changes. Existing methods have failed to effectively distinguish between the abnormalities in normal and fault states.

[0007] (3) The elevator equipment has a variety of models and operating conditions, which leads to changes in the amplitude of the monitoring signal. Existing methods ignore the "individual differences" of elevators, resulting in poor generalization performance of the methods in cross-equipment scenarios.

[0008] In view of this, the present invention provides a knowledge-guided automatic cleaning method for elevator vibration data to solve the above problems. Summary of the Invention

[0009] To overcome the problems in the prior art, this invention proposes a knowledge-guided automatic cleaning method for elevator vibration data. The method uses a signal processing knowledge-guided convolutional autoencoder structure to determine the reconstruction error of the signal to identify normal and abnormal data. For abnormal data, spectral analysis is performed to determine whether a fault has occurred. For data that has not occurred, the abnormal data type of the sensor is identified through a quartile box plot. For data that has occurred, the fault is located by checking whether the amplitude at the fault characteristic frequency exceeds a threshold.

[0010] In a first aspect, the present invention provides a knowledge-guided automatic cleaning method for elevator vibration data, including a training phase and a testing phase;

[0011] Step S1: During the training phase, normal monitoring data is collected as training data; a convolutional autoencoder based on discrete wavelet transform is used to perform convolutional training on the normal monitoring data to obtain a signal reconstruction model;

[0012] Step S2: During the testing phase, the reconstruction error of the vibration monitoring data is calculated based on the signal reconstruction model. A threshold is set according to the reconstruction error, and the vibration monitoring data is judged to be normal or abnormal based on the threshold. For abnormal monitoring data, spectrum analysis is performed to calculate the theoretical value of the fault characteristic frequency. The peak value of the detected amplitude within ±5% of the theoretical value of the fault characteristic frequency is marked as normal equipment, otherwise it is faulty equipment.

[0013] Vibration monitoring data for normal equipment is divided into time slices, and the absolute average value of each time segment is calculated. The abnormality type is determined by combining the box plot rules. The abnormality types include abnormal impact, sensor detachment, or low-frequency interference.

[0014] As a preferred embodiment of the first aspect of the present invention, the construction logic of the discrete wavelet transform convolutional autoencoder is as follows:

[0015] A convolutional autoencoder structure based on discrete wavelet transform is constructed to decompose and reconstruct the input signal;

[0016] The convolutional autoencoder structure adopts a 3-layer encoder-decoder symmetrical structure. The encoder decomposes the signal layer by layer through wavelet basis convolution kernels to extract low-frequency approximate components and high-frequency detail components. The decoder symmetrically reconstructs the signal to form a multi-scale analysis framework.

[0017] The convolution kernel is initialized using the scaling function and wavelet function of the wavelet function, the associated convolution kernel is calculated through the characteristics of the conjugate orthogonal filter, and a learnable hard threshold function is introduced to automatically optimize the signal decomposition during training.

[0018] With minimizing reconstruction error as the core objective, the Adam optimizer is used to adjust the learning rate in stages to drive the convolutional autoencoder structure to extract features of normal elevator vibration signals.

[0019] As a preferred technical solution of the first aspect of the present invention, the convolutional autoencoder structure based on discrete wavelet transform is a structure of stacked 3-layer encoder and 3-layer decoder.

[0020] As a preferred technical solution of the first aspect of the present invention, the application logic of the encoder is as follows:

[0021] Vibration monitoring data is used as the input signal. Low-frequency approximate components are extracted using a scaling function convolution kernel, and high-frequency detail components are extracted using a wavelet function convolution kernel. Noise in the high-frequency detail components is then filtered out using a hard threshold function.

[0022] Pooling is performed on low-frequency approximate components and high-frequency detail components to compress the data dimensionality.

[0023] Repeat the above process to decompose the low-frequency components layer by layer, forming multi-level time-frequency characteristics. The final low-frequency approximation components and high-frequency detail components at each level are the encoded vibration signals of the vibration monitoring data.

[0024] As a preferred technical solution of the first aspect of the present invention, the application logic of the decoder is as follows:

[0025] Using the encoded vibration signal as the input signal, the final low-frequency approximation component is upsampled, and the high-frequency detail components of the third layer are fused through deconvolution operation to initially reconstruct the low-frequency component;

[0026] By analyzing the high-frequency detail components at each level, the high-frequency signal is reconstructed to approximate the vibration monitoring data as closely as possible.

[0027] As a preferred embodiment of the first aspect of the present invention, the reconstruction error is calculated using a loss function:

[0028] If the reconstruction error is less than or equal to the threshold, the vibration monitoring data is determined to be normal monitoring data.

[0029] If the reconstruction error is greater than the threshold, the vibration monitoring data is determined to be abnormal monitoring data.

[0030] As a preferred embodiment of the first aspect of the present invention, the logic for determining the abnormal type is as follows:

[0031] The vibration monitoring data is divided into time slices according to the sampling frequency and sampling duration, and the absolute average value within each time slice is calculated; the absolute average values ​​of all time slices are combined into an absolute average value sequence.

[0032] After arranging all absolute average sequences in ascending order, the 75th and 25th percentile numbers are marked as the first threshold and the second threshold, respectively.

[0033] If the absolute average sequence contains an event exceeding the first threshold, it is marked as an abnormal shock.

[0034] If any absolute average value in the sequence exceeds the second threshold, it is marked as sensor detachment;

[0035] If neither of the above two situations applies, it is marked as low-frequency interference.

[0036] The specific advantages of this invention are as follows:

[0037] This invention uses wavelet basis functions to initialize convolution kernels and constructs a three-layer encoder-decoder structure. It effectively extracts the essential features of vibration signals through multi-scale decomposition and combines an adaptive hard threshold function to significantly improve noise suppression capabilities, achieving a breakthrough in elevator vibration data cleaning and anomaly detection.

[0038] This invention not only enables accurate early warning of faults, but also effectively distinguishes between equipment faults and sensor anomalies, providing a high-precision, high-efficiency, and highly generalizable solution for intelligent elevator operation and maintenance. The automated processing significantly reduces the need for manual intervention. Attached Figure Description

[0039] Figure 1 This is a flowchart of an automatic elevator vibration data cleaning method guided by the knowledge of this invention.

[0040] Figure 2 This is a diagram showing the overall architecture of the convolutional autoencoder network based on discrete wavelet transform in this invention.

[0041] Figure 3 This is a schematic diagram of the encoder framework corresponding to the convolutional autoencoder based on discrete wavelet transform of the present invention;

[0042] Figure 4 This is a schematic diagram of the decoder framework corresponding to the convolutional self-decoder based on discrete wavelet transform of the present invention;

[0043] Figure 5 This is a schematic diagram of abnormal impact waveforms in the vibration monitoring data of this invention;

[0044] Figure 6 This is a schematic diagram of the waveform of the vibration monitoring data sensor detachment in this invention;

[0045] Figure 7 This is a schematic diagram of the loose waveform of the vibration monitoring data sensor of the present invention;

[0046] Figure 8 This is a schematic diagram of the low-frequency interference waveform of the vibration monitoring data of the present invention;

[0047] Figure 9 This is a schematic diagram of the waveform of the abnormal data segment in this invention;

[0048] Figure 10 This is a schematic diagram of a normal data segment waveform of the present invention;

[0049] Figure 11 This is a schematic diagram of the waveform of the data anomaly segment identification result of the present invention;

[0050] Figure 12 This is a schematic diagram of the waveform of the normal data segment recognition result of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1

[0053] Please see Figure 1 The present invention provides a technical solution: a knowledge-guided automatic cleaning method for elevator vibration data, including a training phase and a testing phase.

[0054] I. During the training phase, normal monitoring data is collected as training data; a convolutional autoencoder based on discrete wavelet transform is used to train the normal monitoring data.

[0055] (1) Normal monitoring data refers to vibration monitoring data during normal elevator operation. Normal monitoring data is x = {x1, x2, ..., x}. N}, where: x is x i The data represents the normal monitoring data collected at time i; the subscripts 1 to N represent the sampling times.

[0056] (2) Design as follows Figure 2-4 The convolutional autoencoder structure based on discrete wavelet transform shown uses convolutional training on normal monitoring data to obtain a signal reconstruction model. Based on this model, the input signal is decomposed and reconstructed, where: Figure 2 The paper presents a convolutional autoencoder structure based on discrete wavelet transform, consisting of a stacked 3-layer encoder and a 3-layer decoder. Figure 3 For encoder frame structure, Figure 4 The proposed method presents a decoder framework. Unlike the random initialization of convolutional kernels in traditional convolutional networks, this method uses the scaling function h[n] and wavelet function g[n] from wavelet transform as convolutional kernels. The encoder and decoder are symmetrically configured. The encoder decomposes the signal layer by layer using wavelet-based convolutional kernels to extract low-frequency approximate components and high-frequency detail components. The decoder symmetrically reconstructs the signal, forming a multi-scale analysis framework.

[0057] Specifically, by selecting a wavelet, the scaling function h[n] can be initialized. Based on the characteristics of the conjugate orthogonal filter, the wavelet function g[n] and the reconstruction filter of the scaling function h[n] and the wavelet function g[n] can be determined. and The calculation can be performed according to formula (1.1):

[0058]

[0059] In the above formula, h[n] and It is a reconstruction filter. Through this structure, the network can achieve reasonable separation of signals in different frequency bands and at different times. Therefore, only the scaling function h[n] is optimized in the entire network.

[0060] Taking the encoding process as an example, Figure 3 The method of combining discrete wavelet transform with the encoder structure in convolutional autoencoders is given, and the specific convolution calculation process is as follows:

[0061]

[0062] In the above formula, a l and a l-1 These are the approximate signals for layers l and l-1, respectively. The values ​​for the initial layer are x and d. l This represents the detail signal of layer l. Pool represents the pooling operation, * represents the convolution operation, and f(·) is a learnable hard thresholding function, calculated as follows:

[0063] f(x) = x·[σ(-10·(x+b)] - ))+σ(-10·(x+b + (1.3)

[0064] Where σ is the Sigmoid function, b + and b - These are two learnable parameters that optimize along with the overall network.

[0065] Accordingly, the combination of discrete wavelet transform and the decoder structure in convolutional autoencoders is described in [reference needed]. Figure 4 .

[0066] The Adams algorithm is used to optimize the entire network, and the loss function is calculated as follows:

[0067]

[0068] In the above formula, This is the reconstructed data of the i-th sample. Specifically, the number of training epochs is 100, the initial learning rate is set to 0.001, and the learning rate is adjusted to 0.0005 in the 50th epoch. The number of encoder and decoder layers is set to 3 each.

[0069] (3) When the number of training iterations reaches the specified value or the reconstruction error is lower than the threshold, stop training and output the convolutional autoencoder for use as a signal reconstruction model.

[0070] This embodiment only requires monitoring data from normal elevator operation for training, which better reflects the actual situation in engineering where there is a large amount of normal monitoring data but a small sample size for faults and abnormal states. Meanwhile, the data reconstruction quality based on autoencoders is highly dependent on the network structure. Traditional convolutional autoencoder structures learn features that lack physical meaning and struggle to find local optima. The proposed convolutional autoencoder based on discrete wavelet transform fully incorporates prior knowledge from signal processing, making it easier for the autoencoder to extract vibration signal features. The proposed process can locate normal states, abnormal values ​​within normal states (abnormal impacts, low-frequency interference, sensor detachment), and fault states. The method is simple to use, has strong generalization ability, and demonstrates good engineering application results.

[0071] II. Testing Phase

[0072] (1) Input the newly acquired vibration monitoring data into the convolutional autoencoder based on discrete wavelet transform that was trained in the training phase, and calculate the reconstruction error using formula (1.4).

[0073] (2) If the reconstruction error is less than the set threshold ε, the current input vibration monitoring data is considered normal. Specifically, the threshold is set to 1.5 times the average reconstruction error of the autoencoder trained on normal monitoring data.

[0074] (3) If the reconstruction error is greater than the set threshold ε, then the following formula is used to perform spectrum analysis.

[0075]

[0076] In the above formula, x(kΔt) is the newly acquired vibration monitoring data, N is the number of sequence points, Δt is the sampling interval, n is the index of the frequency domain discrete value, and k is the index of the time domain discrete value.

[0077] (3-1) Calculate the theoretical value of the fault characteristic frequency and search for the corresponding maximum amplitude value in the spectrum within its ±5% range. If the value exceeds the normal level, it indicates the presence of a fault characteristic frequency component, and the equipment is identified as faulty. If the signal contains a large amount of modulation components (e.g., gearbox), the signal can be preprocessed using methods such as fast spectral kurtosis filtering before searching for the maximum amplitude value in the spectrum (the maximum amplitude value is the maximum amplitude value in the vibration monitoring data spectrum). The frequency band with the largest kurtosis can be automatically selected for bandpass filtering, thereby reducing the interference of noise components in the signal.

[0078] (3-2) Otherwise, the vibration monitoring data is time-sliced ​​according to the sampling frequency and sampling duration. The absolute average value within each time slice is calculated (the absolute average value is the absolute average value of the vibration monitoring data in each time slice). The absolute average values ​​of all time slices are combined into an array or sequence called the absolute average value sequence. Specifically, the number of slices can usually be set to [4, 10]. The higher the sampling duration or sampling frequency, the larger this value can be. Based on the absolute average value sequence, the box curve can be... Figure 4 The quantile method is used to determine abnormal sensor conditions.

[0079] Q3 + 1.5 × (Q3 - Q1) (1.6)

[0080] Q1 - 1.5 × (Q3 - Q1) (1.7)

[0081] In the above formula, Q1 and Q3 are the 75th and 25th percentiles of all absolute mean sequences arranged from smallest to largest, respectively. The selection of this threshold is based on statistical experience and is applicable to most common data distributions. For example, when the data presents an approximately normal distribution, about 99.3% of the data are within this threshold, and those outside the threshold are considered abnormal.

[0082] If there is a value in the absolute average sequence that exceeds the value in equation (1.6), it is considered an abnormal shock.

[0083] If the absolute average value sequence contains a value exceeding that of equation (1.7), it is considered that the sensor has detached.

[0084] Otherwise, it is considered to be low-frequency interference.

[0085] This embodiment systematically considers the scenario of abnormal monitoring data generated by elevator equipment during operation, and realizes the comprehensive identification of abnormal situations such as abnormal impact, low frequency interference, sensor detachment and other abnormal conditions and fault states under normal equipment conditions. The proposed convolutional autoencoder structure based on discrete wavelet transform is used to characterize the degree of deviation of abnormal signals and fault signals from normal signals, and the presence of faults is determined based on spectrum analysis. The proposed time slice absolute average quartile method is used to determine the applicability of the solution in scenarios with diverse elevator equipment models.

[0086] Experiments were conducted on escalators, with vibration acceleration sensors placed on the motors. Abnormal and fault data were collected from the sensors under normal and healthy conditions. The fault data was collected when the bearing experienced inner ring, outer ring, or rolling element failures. The "abnormal impact" state was simulated by tapping near the sensor, the "sensor detachment" state was simulated by manually removing the sensor, and the "low-frequency interference" state was simulated by adding low-frequency components to the signal. Figure 5-8 The waveforms of vibration monitoring data collected by the sensor under various abnormal conditions are presented, each lasting 80 seconds. Table 1 compares the accuracy of the proposed method with other methods. Due to the randomness of the methods, each method was trained 5 times, and the average and standard deviation were included in Table 1. It can be seen that the proposed method has higher accuracy than existing methods. Figure 9-12 The results of using the box plot approach to identify abnormal data are presented, showing that the sensor can effectively identify abnormal impacts.

[0087] Table 1 Comparison of Method Accuracy

[0088] method Recognition accuracy (%) Density-based spatial clustering (DBSCAN) 89.39±3.7 A proposed method (using a common convolutional autoencoder) is presented. 95.24±2.3 Proposed Method 98.18±2.8

[0089] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0090] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge-guided automatic cleaning method for elevator vibration data, characterized in that, Includes the following steps: Step S1: During the training phase, normal monitoring data is collected as training data; a convolutional autoencoder based on discrete wavelet transform is constructed, and the normal monitoring data is used to train the signal reconstruction model. Step S2: During the testing phase, the reconstruction error of the vibration monitoring data is calculated based on the signal reconstruction model. According to the set reconstruction error threshold, the vibration monitoring data is determined to be normal or abnormal monitoring data. For abnormal monitoring data, spectrum analysis is performed to calculate the theoretical value of the fault characteristic frequency. If the peak value of the detected amplitude is within the normal level within ±5% of the theoretical value of the fault characteristic frequency, it is marked as normal equipment; otherwise, it is faulty equipment. Vibration monitoring data for normal equipment is divided into time slices, and the absolute average value of each time segment is calculated. The abnormality type of vibration monitoring data is determined by combining the box plot rules. The abnormality type includes abnormal impact, sensor detachment, or low-frequency interference.

2. The knowledge-guided automatic cleaning method for elevator vibration data according to claim 1, characterized in that, The construction logic of the discrete wavelet transform convolutional autoencoder is as follows: A convolutional autoencoder structure based on discrete wavelet transform is constructed to decompose and reconstruct the input signal; The convolutional autoencoder structure adopts an encoder-decoder symmetrical structure. The encoder decomposes the signal layer by layer through wavelet-based convolution kernels to extract low-frequency approximate components and high-frequency detail components. The decoder reconstructs the signal through the symmetrical structure to form a multi-scale analysis framework. The convolution kernel is initialized using the scaling function and wavelet function in wavelet transform, and the associated convolution kernel is calculated through the characteristics of conjugate orthogonal filters. A self-learning hard threshold function is introduced to automatically optimize during training and reduce the impact of signal noise interference. With minimizing reconstruction error as the core objective, the Adam optimizer is used to adjust the learning rate in stages to drive the convolutional autoencoder structure to extract the features of normal elevator vibration signals.

3. The knowledge-guided automatic cleaning method for elevator vibration data according to claim 2, characterized in that, Encoder application logic: Vibration monitoring data is used as the input signal. Low-frequency approximate components are extracted using a scaling function convolution kernel, and high-frequency detail components are extracted using a wavelet function convolution kernel. Noise in the high-frequency detail components is then filtered out using a hard threshold function. Pooling is performed on low-frequency approximate components and high-frequency detail components to compress the data dimensionality. Repeat the above process to decompose the low-frequency components layer by layer, forming multi-level time-frequency characteristics. The final low-frequency approximate components and high-frequency detail components at each level are the encoded vibration signals of the vibration monitoring data.

4. The knowledge-guided automatic cleaning method for elevator vibration data according to claim 3, characterized in that, Decoder application logic: Using the encoded vibration signal as the input signal, the final low-frequency approximation component is upsampled, and the high-frequency detail components of the third layer are fused through deconvolution operation to initially reconstruct the low-frequency component; By reconstructing the high-frequency signal from each layer of high-frequency detail components, the vibration monitoring data can be approximated.

5. The knowledge-guided automatic cleaning method for elevator vibration data according to claim 4, characterized in that, Calculate the reconstruction error using the loss function: If the reconstruction error is less than or equal to the threshold, the vibration monitoring data is determined to be normal monitoring data. If the reconstruction error is greater than the threshold, the vibration monitoring data is determined to be abnormal monitoring data.

6. The knowledge-guided automatic cleaning method for elevator vibration data according to claim 5, characterized in that, The logic for determining the type of exception is as follows: The vibration monitoring data is divided into time slices according to the sampling frequency and sampling duration, and the absolute average value within each time slice is calculated; the absolute average values ​​of all time slices are combined into an absolute average value sequence. After arranging all absolute average sequences in ascending order, the 75th and 25th percentile numbers are marked as the first threshold and the second threshold, respectively. If the absolute average sequence contains an event exceeding the first threshold, it is marked as an abnormal shock. If any absolute average value in the sequence exceeds the second threshold, it is marked as sensor detachment; If there are no other cases, they are marked as low-frequency interference.

Images