An elevator fault early warning method and system based on voiceprint recognition

Abstract

The application discloses an elevator fault early warning method and system based on voiceprint recognition, and relates to the field of elevator detection. The method comprises the following steps: obtaining the traction machine operation voiceprint of a full-cycle operation condition of a to-be-detected elevator under different elevator loads; extracting voiceprint features through wavelet transform; taking the operation condition as the X-axis, the elevator load as the Y-axis, and the voiceprint features as the Z-axis to construct a first point cloud space, and inputting the first point cloud space into an elevator fault risk assessment model; extracting horizontal time sequence features along the X-Z plane, vertical space features along the Y-Z plane, and joint distribution features along the X-Y-Z space; fusing the horizontal time sequence features, the vertical space features, and the joint distribution features to generate first fusion features and determine a fault risk value; and issuing a warning if the fault risk value is greater than a preset fault threshold. Through implementation of the application, the limitations of traditional machine learning methods that only learn the relationship between voiceprints and faults are solved, and the fault recognition accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of elevator detection, and in particular to an elevator fault early warning method and system based on voiceprint recognition. BACKGROUND

[0002] In recent years, with the acceleration of urbanization, the safety and reliability of elevators, as indispensable vertical transportation tools in high-rise buildings, have attracted widespread attention. Among various types of elevators, traction elevators have become the mainstream choice in residential buildings, shopping malls and other places due to their stable performance, efficient operation and large load capacity.

[0003] A traction elevator is mainly composed of a traction machine, a car, a steel wire rope, a control cabinet and other key components. The traction machine, as the power core of the elevator, its running state is directly related to the safety and stability of the elevator. In order to timely discover possible faults of the traction machine, a fault early warning technology based on voiceprint recognition has emerged. At present, the existing fault early warning schemes based on voiceprint recognition mostly use traditional machine learning methods to construct prediction models. After collecting the voiceprint data of the traction machine, these schemes first extract features from the data, and then directly establish a mapping relationship between the voiceprint features and the fault types, so as to realize fault early warning.

[0004] However, in the actual elevator operation scene, there are many dynamic variables. For example, different load conditions will cause the load torque of the traction machine motor and the tension of the steel wire rope to change, thereby causing the energy baseline and frequency distribution of the voiceprint to shift. At the same time, the voiceprint features generated by the elevator in different operating conditions, such as starting, accelerating, uniform speed, decelerating and braking, are different. The traditional machine learning method only focuses on learning the corresponding relationship between the voiceprint and the fault when constructing the model, but ignores the key problem that the load change and operating condition difference affect the voiceprint features. SUMMARY

[0005] The embodiments of the present application provide an elevator fault early warning method and system based on voiceprint recognition, which can solve the limitation of traditional machine learning methods that only learn the relationship between voiceprint and fault, and improve the fault recognition accuracy.

[0006] An embodiment of the present application provides an elevator fault early warning method based on voiceprint recognition, comprising:

[0007] obtaining the running voiceprint of the traction machine of the elevator to be detected under different elevator loads in the full cycle operating condition; the full cycle operating condition includes starting, accelerating, uniform speed, decelerating and braking;

[0008] performing wavelet transform on the running voiceprint of the traction machine to extract voiceprint features;

[0009] According to the soundprint features corresponding to the soundprint of the traction machine under different elevator loads and full-cycle operation conditions, a first point cloud space is constructed with the operation condition as the X-axis, the elevator load as the Y-axis, and the soundprint feature as the Z-axis;

[0010] The first point cloud space is input into the trained elevator fault risk assessment model, so that the elevator fault risk assessment model extracts horizontal time sequence features along the X-Z plane of the first point cloud space through the internally set time sequence convolution network, extracts vertical spatial features along the Y-Z plane of the first point cloud space through the internally set perception convolution network, and extracts joint distribution features along the X-Y-Z space of the first point cloud space through the internally set first 3D convolution network. The horizontal time sequence features and the vertical spatial features are multiplied by broadcasting to generate spatiotemporal interaction features. After the spatiotemporal interaction features and the joint distribution features are concatenated along the channel dimension, local feature aggregation is performed through the internally set second 3D convolution network to generate first fusion features. According to the first fusion features, a fault risk value is determined.

[0011] In the case where the fault risk value is greater than a preset fault threshold, a fault warning signal is issued.

[0012] Further, after obtaining the traction machine operation soundprint of the to-be-detected elevator under different elevator loads and full-cycle operation conditions, the following steps are further included:

[0013] According to the elevator load and the operation condition, a probability distribution of corresponding noise parameters is called from a preset noise parameter probability distribution library;

[0014] For each operation condition, a noise parameter combination conforming to the probability distribution is generated through random sampling based on the probability distribution of the noise parameters. The noise parameter combination includes amplitude, frequency, phase, and bandwidth.

[0015] According to the sampled frequency and bandwidth, the spectral shape of the noise signal is determined.

[0016] According to the sampled amplitude, the spectral shape is adjusted in amplitude to obtain a frequency domain signal.

[0017] The frequency domain signal is subjected to inverse Fourier transform to obtain a time domain signal.

[0018] According to the sampled phase, the time domain signal is adjusted in phase to generate a noise segment of the operation condition.

[0019] The noise segments of all operation conditions are spliced to obtain a continuous noise signal.

[0020] The noise signal is subjected to short-time Fourier transform to obtain a noise spectrum.

[0021] The traction machine operation soundprint is subjected to short-time Fourier transform to obtain a soundprint spectrum.

[0022] subtract the noise spectrum from the voiceprint spectrum by using a spectrum subtraction method to obtain a denoised voiceprint spectrum;

[0023] perform inverse short-time Fourier transform on the denoised voiceprint spectrum to obtain a denoised traction machine operation voiceprint.

[0024] Further, wavelet transform is performed on the traction machine operation voiceprint to extract voiceprint features, including:

[0025] The traction machine operation voiceprint is divided according to the operating conditions to obtain corresponding voiceprint segments;

[0026] For each voiceprint segment, a corresponding wavelet basis function is determined according to the operating condition in which the voiceprint segment is located;

[0027] According to the elevator load and the rated load, a load coefficient is calculated;

[0028] The wavelet basis function is corrected based on the load coefficient to obtain a corrected wavelet basis function;

[0029] Wavelet transform is performed on the voiceprint segment based on the corrected wavelet basis function to obtain corresponding voiceprint features.

[0030] Further, a built-in time sequence convolution network is used to extract lateral time sequence features along the X-Z plane of the first point cloud space, including:

[0031] The Y-axis direction of the first point cloud space is subjected to multi-scale data slicing processing to obtain a plurality of first slice data for representing the time sequence variation law during the switching process of different operating conditions of the same elevator load;

[0032] Each first slice data is subjected to time sequence convolution processing to obtain a first trend feature;

[0033] All first trend features are spliced to obtain lateral time sequence features.

[0034] Further, a built-in perception convolution network is used to extract vertical spatial features along the Y-Z plane of the first point cloud space, including:

[0035] The X-axis direction of the first point cloud space is subjected to multi-scale data slicing processing to obtain a plurality of second slice data for representing the voiceprint feature variation law under different elevator loads of the same operating condition;

[0036] Each second slice data is subjected to dilated convolution processing to obtain a second trend feature;

[0037] All second trend features are spliced to obtain vertical spatial features.

[0038] Further, a first 3D convolutional network is internally arranged to extract joint distribution features along the X-Y-Z space of the first point cloud space, including:

[0039] The Z-axis direction of the first point cloud space is subjected to multi-scale data space cutting processing to obtain a plurality of associated space data for representing the distribution of the running conditions and the elevator load in the same voiceprint feature range;

[0040] Each associated space data is subjected to 3D convolution processing to obtain associated features;

[0041] The joint distribution features are obtained by splicing all the associated features.

[0042] Further, the horizontal time sequence features and the vertical space features are subjected to broadcast multiplication to generate space-time interaction features, including:

[0043] The horizontal time sequence features and the vertical space features are aligned according to the feature dimensions to obtain adapted horizontal time sequence features and adapted vertical space features;

[0044] The adapted horizontal time sequence features and the adapted vertical space features are subjected to tensor outer product operation to obtain space-time interaction features.

[0045] Further, the preset failure threshold is determined by the following method:

[0046] The design parameters and the current service life of the elevator to be detected are obtained;

[0047] According to the design parameters, an initial threshold is determined;

[0048] According to the preset life distribution of the elevator to be detected, a degradation coefficient under the current service life is calculated;

[0049] According to the degradation coefficient and the initial threshold, the preset failure threshold is determined.

[0050] On the basis of the above method embodiment, the present application correspondingly provides a system embodiment, including: a voiceprint acquisition module, a voiceprint feature extraction module, a point cloud space construction module, a risk assessment module and a failure warning module;

[0051] The voiceprint acquisition module is used to acquire the traction machine running voiceprint of the full-cycle running condition of the elevator to be detected under different elevator loads; the full-cycle running condition includes: starting, accelerating, uniform speed, decelerating and braking;

[0052] The voiceprint feature extraction module is used to perform wavelet transform on the traction machine running voiceprint to extract voiceprint features;

[0053] a point cloud space construction module, configured to construct a first point cloud space according to sound print features corresponding to the sound print of the traction machine under different elevator loads and full-cycle operation conditions;

[0054] a risk assessment module, configured to input the first point cloud space into a trained elevator fault risk assessment model, so that the elevator fault risk assessment model extracts horizontal time sequence features along an X-Z plane of the first point cloud space through an internally set time sequence convolution network, extracts vertical space features along a Y-Z plane of the first point cloud space through an internally set perception convolution network, extracts joint distribution features along an X-Y-Z space of the first point cloud space through an internally set first 3D convolution network, performs broadcast multiplication on the horizontal time sequence features and the vertical space features to generate space-time interaction features, concatenates the space-time interaction features and the joint distribution features along a channel dimension, and then performs local feature aggregation through an internally set second 3D convolution network to generate first fusion features, and determines a fault risk value according to the first fusion features;

[0055] a fault warning module, configured to issue a fault warning signal when the fault risk value is greater than a preset fault threshold.

[0056] Further, the elevator fault warning system based on sound print recognition further comprises a sound print noise reduction module.

[0057] The sound print noise reduction module comprises a noise signal generation unit, a noise frequency domain conversion unit, a sound print frequency domain conversion unit, a spectral subtraction noise reduction unit, and a sound print time domain reconstruction unit.

[0058] The noise signal generation unit is configured to call a probability distribution of corresponding noise parameters from a preset noise parameter probability distribution library according to elevator loads and operation conditions, and perform random sampling to generate a noise signal.

[0059] The noise frequency domain conversion unit is configured to perform short-time Fourier transform on the noise signal to obtain a noise spectrum.

[0060] The sound print frequency domain conversion unit is configured to perform short-time Fourier transform on the traction machine operation sound print to obtain a sound print spectrum.

[0061] The spectral subtraction noise reduction unit is configured to subtract the noise spectrum from the sound print spectrum through spectral subtraction processing to obtain a noise-reduced sound print spectrum.

[0062] The sound print time domain reconstruction unit is configured to perform inverse short-time Fourier transform on the noise-reduced sound print spectrum to obtain a noise-reduced traction machine operation sound print.

[0063] Compared with the prior art, the beneficial effects of the scheme embodiment are as follows:

[0064] The application obtains the running sound print of the traction machine of the elevator to be detected under different elevator loads and full-cycle operating conditions, wherein the full-cycle operating conditions include starting, acceleration, uniform speed, deceleration and braking. Under different loads and conditions, the running state of the traction machine is different, and the sound print generated will also be different. Then, the running sound print of the traction machine is wavelet transformed to extract the sound print features. According to the sound print features corresponding to the running sound print of the traction machine under different elevator loads and full-cycle operating conditions, a first point cloud space is constructed with the operating condition as the X-axis, the elevator load as the Y-axis and the sound print feature as the Z-axis, so that the load, the condition and the sound print feature are structurally integrated. The first point cloud space is input into the trained elevator fault risk assessment model. In the feature extraction stage, through the internally set time sequence convolution network, the transverse time sequence features are extracted along the X-Z plane of the first point cloud space. Since the full-cycle operating conditions are arranged in time sequence, the time sequence convolution network can capture the time sequence change rule of the sound print features in the switching process of different conditions. Through the internally set perception convolution network, the vertical space features are extracted along the Y-Z plane of the first point cloud space. Different elevator loads correspond to different positions on the Y-axis, and the perception convolution network can mine the change rule of the sound print features under different loads. Through the internally set first 3D convolution network, the joint distribution features are extracted along the X-Y-Z space of the first point cloud space. The 3D convolution network can comprehensively consider the joint distribution of the condition, the load and the sound print feature in the three-dimensional space, and capture the complex interaction among the three. In the feature fusion stage, the transverse time sequence features and the vertical space features are multiplied by broadcasting to generate the space-time interaction features. After the space-time interaction features and the joint distribution features are spliced along the channel dimension, the second 3D convolution network is internally set to perform local feature aggregation to generate the first fusion features, which comprehensively consider time, space and their synergies, and can more comprehensively describe the running state of the elevator. According to the first fusion features, the fault risk value is determined. If the fault risk value is greater than the preset fault threshold, a fault warning signal is sent.

[0065] In summary, the application integrates the condition, the load and the sound print feature by constructing a three-dimensional point cloud space, and extracts and couples multi-dimensional features through the model, so that the model can learn the dynamic influence of the load and the condition on the sound print feature, thereby solving the limitation of the traditional machine learning method that only learns the relationship between the sound print and the fault, and improving the fault recognition accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 is a flowchart of an elevator fault warning method based on sound print recognition provided by an embodiment of the application;

[0067] Figure 2 is a structural schematic diagram of an elevator fault warning system based on sound print recognition provided by an embodiment of the application. DETAILED DESCRIPTION

[0068] 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.

[0069] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated.

[0070] Traditional machine learning methods have significant limitations in elevator fault early warning. They focus solely on the relationship between sound patterns and faults, neglecting the impact of key physical factors during elevator operation on the acoustic characteristics of faults. Specifically, the output torque of the traction machine increases with load. If the traction machine has hidden problems such as slight gear wear or excessive bearing clearance, there may be no obvious abnormal noise under low loads, but under high loads, the increased stress will significantly reveal the acoustic characteristics of the fault. Furthermore, the speed and force patterns of the elevator traction machine differ during startup, acceleration, constant speed, deceleration, and braking. For example, during startup, the traction machine needs to overcome static friction, causing a sudden increase in speed from 0; during braking, the brake contacts the traction sheave, generating instantaneous friction. Different fault types will exhibit different acoustic characteristics under these different operating conditions. Therefore, traditional methods cannot comprehensively capture these complex situations, limiting the accuracy and reliability of fault early warning.

[0071] like Figure 1 As shown, in order to overcome the limitation of traditional machine learning methods that only learn the relationship between voiceprints and faults, an embodiment of the present invention provides an elevator fault early warning method based on voiceprint recognition, which includes at least the following steps:

[0072] Step S1: Obtain the traction machine operating sound pattern of the elevator under different elevator loads during the entire cycle of operation; the entire cycle of operation includes: starting, acceleration, constant speed, deceleration and braking;

[0073] For step S1, in order to accurately capture the acoustic characteristics of the elevator traction machine in actual operation, the traction machine's operating sound patterns are collected by microphones at the periphery of the traction machine under different elevator loads throughout the entire cycle of operation.

[0074] In this embodiment, considering the diversity of loads during actual elevator operation, two or more different elevator loads are set. This is because the output characteristics of the traction machine change with the load; the stress and vibration of its internal components differ under different loads, resulting in different sound characteristics. For example, the interaction between components is relatively mild under low loads, while abnormal noises may occur due to excessive force under high loads. By setting multiple different loads, the acoustic performance of the traction machine under various load scenarios can be captured more comprehensively.

[0075] The full-cycle operating conditions include the starting phase, acceleration phase, constant speed phase, deceleration phase, and braking phase. It should be noted that for some elevators with shorter operating times, the full-cycle operating conditions can be simplified to include only the starting phase, acceleration phase, deceleration phase, and braking phase to suit actual conditions. In this case, the duration of the constant speed phase is 0.

[0076] In a preferred embodiment, after obtaining the traction machine operating sound signature of the elevator under different elevator loads throughout its entire operating cycle, the method further includes:

[0077] Based on the elevator load and operating conditions, the probability distribution of the corresponding noise parameter is retrieved from the preset noise parameter probability distribution library;

[0078] For each operating condition, based on the probability distribution of noise parameters, a combination of noise parameters conforming to the probability distribution is generated through random sampling; the combination of noise parameters includes: amplitude, frequency, phase, and bandwidth;

[0079] The spectral shape of the noise signal is determined based on the frequency and bandwidth obtained from the sampling.

[0080] Based on the amplitude obtained from sampling, the shape of the spectrum is adjusted to obtain the frequency domain signal;

[0081] Perform an inverse Fourier transform on the frequency domain signal to obtain the time domain signal;

[0082] Based on the sampled phase, the time-domain signal is phase-adjusted to generate a noise segment for this operating condition;

[0083] Noise segments from all operating conditions are spliced ​​together to obtain a continuous noise signal;

[0084] The noise signal is subjected to a short-time Fourier transform to obtain the noise spectrum;

[0085] The acoustic waveform of the traction machine is subjected to a short-time Fourier transform to obtain the acoustic waveform spectrum.

[0086] The noise spectrum is subtracted from the speaker spectrum using spectral subtraction to obtain the noise-reduced speaker spectrum;

[0087] The denoised acoustic signature spectrum is subjected to inverse short-time Fourier transform to obtain the denoised traction machine operating acoustic signature.

[0088] In one embodiment of the present invention, after obtaining the traction machine operating sound pattern of the elevator under different elevator loads throughout its entire operating cycle, the method further includes noise reduction processing on the inherent noise contained in the traction machine operating sound pattern. Specifically:

[0089] During actual operation, elevator traction machines inevitably contain various inherent noises, such as airflow noise within the elevator shaft and electromagnetic interference noise from surrounding equipment. These inherent noises mix with the actual operating sound of the traction machine, interfering with subsequent accurate analysis of the traction machine's operating status and fault characteristics.

[0090] To effectively eliminate these inherent noises, it is first necessary to retrieve the probability distribution of the corresponding noise parameters from a pre-set noise parameter probability distribution library based on the elevator's load and operating conditions. This is because the characteristics of inherent noise will differ under different loads and operating conditions. For example, when the elevator is running under high load, the increased output power of the traction machine may cause stronger electromagnetic interference noise, and the probability distribution of its noise parameters will change accordingly.

[0091] The pre-defined noise parameter probability distribution library was established through extensive preliminary experiments and actual data collection. Long-term acoustic signature data was collected from elevators under various loads and operating conditions in diverse elevator operating environments. Simultaneously, specialized noise analysis equipment was used to separate and extract features from the inherent noise in these collected acoustic signature data. For example, a spectrum analyzer can decompose the acoustic signature signal into different frequency components, presenting the frequency distribution of noise; an acoustic vibration analyzer can analyze the relationship between noise and mechanical vibration, obtaining more comprehensive noise characteristic information. For each data collection, the elevator load, operating conditions, and corresponding noise characteristic parameters, including amplitude, frequency, phase, and bandwidth, were recorded in detail. After accumulating a large amount of data, statistical methods were used to analyze and model this data, thereby obtaining the probability distribution of noise parameters under different loads and operating conditions, and storing it to form the pre-defined noise parameter probability distribution library.

[0092] It should be noted that professional noise analysis equipment is typically expensive, bulky, and complex to operate. Equipping and continuously using such equipment for every elevator in actual elevator operation monitoring would significantly increase monitoring costs. Furthermore, in real-world elevator operating environments, space constraints and equipment compatibility issues may exist, making it difficult to directly use professional equipment for real-time, long-term noise monitoring. This invention, by establishing a pre-defined noise parameter probability distribution library and performing noise simulation and reduction processing based on this library, can efficiently remove inherent noise from the acoustic signature of elevator traction machines without relying on professional equipment. This allows for more convenient and economical accurate monitoring and analysis of elevator operating status.

[0093] For each specific operating condition, based on the probability distribution of the noise parameters obtained from the call, a Monte Carlo random sampling method is used to generate a combination of noise parameters that conforms to these probability distributions. Noise parameters include amplitude, frequency, phase, and bandwidth parameters. Taking the frequency parameter as an example, if its probability distribution shows that the frequency is mainly concentrated in a certain range under a certain operating condition, random sampling can obtain the specific frequency values ​​within that range, and simultaneously determine the values ​​of amplitude, phase, and bandwidth, forming a complete set of noise parameter combinations.

[0094] Based on the frequency and bandwidth parameters obtained from sampling, the spectral shape of the noise signal is constructed. A suitable window function is typically used to simulate the spectral characteristics, with the frequency as the center frequency and the bandwidth used to define the width of the spectrum, thus constructing a preliminary spectral model. Then, the amplitude of the spectrum is adjusted according to the amplitude parameters obtained from sampling, so that the overall amplitude of the spectrum matches the sampled values, thereby obtaining the frequency domain signal.

[0095] The adjusted frequency domain signal is subjected to an inverse Fourier transform to convert it to the time domain, resulting in a time domain signal. Then, the time domain signal is phase-adjusted based on the sampled phase parameters. For complex signals, this can be achieved by multiplying by the corresponding phase factor; for real signals, Euler's formula is used for conversion. Finally, a noise segment under this operating condition is generated.

[0096] Noise segments generated under all operating conditions are spliced ​​together according to the actual time sequence of elevator operation to form a continuous noise signal. Then, short-time Fourier transforms are performed on the spliced ​​noise signal and the collected traction machine operation acoustic waveforms to obtain the noise spectrum and acoustic waveform spectrum, respectively.

[0097] The noise spectrum is subtracted from the acoustic signature spectrum using spectral subtraction. It should be noted that during the spectral subtraction process, the estimated value of the noise spectrum amplitude is subtracted from the amplitude of the noisy acoustic signature spectrum while keeping the phase of the acoustic signature spectrum unchanged.

[0098] The noise-reduced ripple spectrum after spectral subtraction is subjected to inverse short-time Fourier transform to convert the frequency domain signal back to the time domain signal, ultimately obtaining the noise-reduced traction machine operating sound pattern. This series of processes removes the interference of inherent noise, allowing the traction machine operating sound pattern to more accurately reflect its true operating status.

[0099] Step S2: Perform wavelet transform on the traction machine's operating acoustic fingerprint to extract acoustic fingerprint features;

[0100] In a preferred embodiment, wavelet transform is performed on the traction machine's operating acoustic signature to extract acoustic signature features, including:

[0101] The traction machine's operating acoustic signature is divided according to the operating conditions to obtain corresponding acoustic signature segments;

[0102] For each voiceprint segment, the corresponding wavelet basis function is determined according to the operating condition of the voiceprint segment;

[0103] Calculate the load factor based on the elevator load and rated load;

[0104] Based on the modified wavelet basis function using the loading coefficient, the modified wavelet basis function is obtained;

[0105] Wavelet transform is performed on the voiceprint segments based on the modified wavelet basis function to obtain the corresponding voiceprint features.

[0106] Regarding step S2, it should be noted that the essence of elevator operation soundprints is a time-series signal, the length of which is directly determined by the operation cycle. For example, the operation time from the 1st floor to the 3rd floor is short, and the soundprint signal duration may be only 5 seconds; the operation time from the 1st floor to the 10th floor is long, and the soundprint signal duration may reach 15 seconds. Before performing wavelet transform on the traction machine operation soundprints to extract soundprint features, it is necessary to align the time scales of all traction machine operation soundprints to ensure that the soundprint features extracted by wavelet transform have consistency and reliability.

[0107] Taking an office building elevator as an example, with a rated load of 1250kg, the traction machine's operating sound patterns under different loads throughout the entire operating cycle have been obtained through step S1. Now, these sound patterns will be divided according to the operating conditions. For example, the sound pattern data of the starting phase is extracted as one sound pattern segment, and the corresponding sound pattern data of the acceleration phase, constant speed phase, deceleration phase, and braking phase are also extracted to form their respective sound pattern segments. The originally continuous sound pattern signal is now divided into multiple segments corresponding to different operating phases.

[0108] For each segmented acoustic signature, the corresponding wavelet basis function is determined based on the operating condition of the segment. For example, during the startup phase, the traction machine accelerates from a standstill, and its acoustic signature signal exhibits sudden changes and rapid increases. In this case, a wavelet basis function with good temporal localization and abrupt change detection capabilities is selected, such as the db4 wavelet from the Daubechies wavelet (dbN wavelet family). This is because the db4 wavelet has a tight support characteristic in the time domain, which can effectively capture abrupt changes in the signal and is suitable for analyzing acoustic signature signals with large dynamic changes during the startup phase. During the acceleration phase, the traction machine's rotational speed continuously increases, and its acoustic signature signal exhibits a gradual increase and a constantly changing frequency. At this point, the coif3 wavelet from the Coiflets wavelet family (coifN wavelet family) can be selected. Coiflets wavelets have a high vanishing moment, meaning they have good localization capabilities in the frequency domain while maintaining certain characteristics in the time domain. During the constant-speed phase, the traction machine runs relatively smoothly, and the acoustic signature signal is relatively stable. Wavelet basis functions with good frequency domain localization characteristics can be selected, such as the sym8 wavelet from the Symlets wavelet family (symN wavelet family). It can provide more accurate analysis in the frequency domain, helping to extract the frequency characteristics of the acoustic signature signal during the constant-speed phase. During the deceleration phase, the traction machine speed gradually decreases, and the trend of the acoustic signature signal is opposite to that of the acceleration phase; the energy gradually weakens, and the frequency also gradually decreases. At this time, the bi... The Bior3.7 wavelet, with its linear phase characteristics, is suitable for the traction machine during deceleration. As the interaction between components changes, the phase relationship of the acoustic signal becomes complex. The Bior3.7 wavelet effectively preserves the signal's phase information, and its biorthogonal properties give it an advantage in signal reconstruction. During braking, the brake contacts the traction sheave, generating momentary friction that causes brief impacts and noise in the acoustic signal. In this case, the Meyer wavelet is chosen. Defined in the frequency domain, the Meyer wavelet possesses excellent regularity and symmetry. During braking, the impact component in the acoustic signal contains rich fault information. The symmetry of the Meyer wavelet ensures that phase distortion is not introduced during signal analysis, and its good regularity helps accurately capture abrupt changes and impact characteristics in the signal.

[0109] The load factor is calculated based on the elevator's actual load and rated load. Assuming the elevator load corresponding to the currently collected voiceprint segment is 800kg and the rated load is 1250kg, then the load factor... The load factor reflects the ratio of the elevator's current load to its rated load. The operating state and acoustic characteristics of the traction machine will differ under different loads. Therefore, the wavelet basis function for each operating stage is corrected using the load factor to better suit the acoustic signature segment analysis under the current load. The specific correction process is as follows:

[0110] According to the elevator's load factor The load is divided into different levels. Taking the rated load of the office building elevator as 1250kg as an example, the rated load elevator load level classification table is set as shown in Table 1:

[0111] Table 1 Rated Load Classification of Elevator Load Ratings

[0112]

[0113] When the elevator load factor This corresponds to the medium load level. To ensure that subsequent wavelet transform analysis can more accurately adapt to the voiceprint segments under the current medium load level, the wavelet basis functions are corrected using preset correction strategies for each load level. The following explanation uses the startup and acceleration phases as examples:

[0114] When an elevator traction machine starts under medium load, the amplitude of abrupt changes in the acoustic signature signal is larger than under low load, but not as drastic as under high load. According to the preset correction strategy, the db4 wavelet, with its tight support in the time domain, is originally suitable for capturing signal abrupt changes of moderate magnitude. However, to better adapt to acoustic signature signals under medium load, the parameters of the db4 wavelet need to be adjusted. Specifically, its scaling parameter is multiplied by an adjustment factor of approximately 0.9. This is done to slightly narrow the support range of the db4 wavelet in the time domain, enhancing its sensitivity to medium-amplitude abrupt changes. Because the abrupt changes in the acoustic signature signal under medium load have a certain intensity, narrowing the support range allows the wavelet to focus more precisely on the abrupt change region, more effectively capturing the abrupt change information of the acoustic signature signal.

[0115] During the acceleration process of a medium-load elevator, the traction machine speed continuously increases, and the acoustic signature signal exhibits a gradual increase in intensity and a constantly changing frequency, with the amplitude and complexity of the frequency changes being moderate. The coif3 wavelet has a high vanishing moment and performs well in both the frequency and time domains. Following a preset correction strategy, the scale parameter of the coif3 wavelet is appropriately reduced, by approximately 0.85 of the original scale parameter. Reducing the scale parameter improves the resolution of the coif3 wavelet in the frequency domain, enabling it to more sensitively capture subtle changes in the acoustic signature signal frequency during medium-load acceleration. Because frequency changes at medium loads are not as drastic and complex as at high loads, adjusting the scale parameter allows the wavelet to be analyzed within a suitable frequency range.

[0116] It should be noted that the acoustic characteristics of the traction machine of different brands and models of elevators may vary under different loads. In addition, the elevator's operating environment (such as temperature, humidity, noise level, etc.) will also affect the acoustic signal. Therefore, the preset correction strategy can be set and adjusted according to the actual situation.

[0117] Finally, wavelet transforms are performed on the corresponding stage voiceprint segments using these modified wavelet basis functions. After the wavelet transform is completed, representative voiceprint features are extracted from the transform results, mainly including wavelet energy entropy and wavelet correlation coefficient.

[0118] This invention fully considers the differences in elevator operating conditions and tailors wavelet basis functions for each stage, making them more closely aligned with signal characteristics and accurately capturing key features. This avoids feature omissions or misjudgments caused by using a uniform wavelet basis function to process acoustic fingerprint signals with significant differences under different operating conditions. Simultaneously, a load coefficient is introduced to classify load levels, and the wavelet basis function is modified accordingly. Since the actual load on the elevator is constantly changing, and the acoustic characteristics of the traction machine differ under different load levels, the modified wavelet basis function can better adapt to the acoustic fingerprint signal under the current load, enabling the wavelet transform to more accurately decompose the signal. The extracted wavelet energy entropy and wavelet correlation coefficient, among other features, more realistically reflect the operating status.

[0119] Step S3: Based on the acoustic characteristics of the traction machine's operating acoustics under different elevator loads throughout the entire operating cycle, construct the first point cloud space with the operating conditions as the X-axis, the elevator load as the Y-axis, and the acoustic characteristics as the Z-axis.

[0120] For step S3, the operating conditions are used as the X-axis to locate and distinguish the different operating stages of the elevator, such as starting, accelerating, constant speed, deceleration, and braking. The elevator load is used as the Y-axis to represent the actual weight carried by the elevator on the vertical axis. The extracted voiceprint features are used as the Z-axis. Each operating condition, elevator load, and corresponding voiceprint feature is mapped to a three-dimensional coordinate system to form data points. All data points are aggregated to form the first point cloud space.

[0121] This point cloud space provides a clear visual representation of the correlation between different operating conditions, elevator loads, and acoustic signature characteristics. For example, the constructed first point cloud space clearly shows the distribution of features such as wavelet energy entropy and wavelet correlation coefficient of the acoustic signature under high-load acceleration conditions. This provides spatial data support for subsequent analysis of the traction machine's operating status under different conditions and loads, as well as for fault diagnosis.

[0122] Step S4: Input the first point cloud space into the trained elevator fault risk assessment model, so that the elevator fault risk assessment model can extract lateral temporal features along the XZ plane of the first point cloud space through the built-in temporal convolutional network; extract vertical spatial features along the YZ plane of the first point cloud space through the built-in perceptual convolutional network; extract joint distribution features along the XYZ space of the first point cloud space through the built-in first 3D convolutional network; perform broadcast multiplication on the lateral temporal features and vertical spatial features to generate spatiotemporal interactive features; after concatenating the spatiotemporal interactive features and joint distribution features along the channel dimension, perform local feature aggregation through the built-in second 3D convolutional network to generate the first fusion feature; determine the fault risk value based on the first fusion feature.

[0123] For step S4, after inputting the first point cloud space into the trained elevator fault risk assessment model, within the model, in order to capture the dynamic patterns of voiceprint features (Z-axis) in the time dimension as the operating conditions (X-axis) change during elevator operation, a built-in temporal convolutional network is used to extract lateral temporal features along the XZ plane of the first point cloud space. Specifically:

[0124] In a preferred embodiment, a built-in temporal convolutional network is used to extract lateral temporal features along the XZ plane of the first point cloud space, including:

[0125] Multi-scale data slicing was performed on the Y-axis direction of the first point cloud space to obtain several first slice data that characterize the temporal variation law during the switching of different operating conditions of the same elevator load.

[0126] Perform temporal convolution processing on each first slice of data to obtain the first trend feature;

[0127] By concatenating all primary trend features, we obtain the lateral time series features.

[0128] In one embodiment of the present invention, multi-scale data slicing is performed along the Y-axis of the first point cloud space. Slicing along the Y-axis essentially involves extracting the acoustic signature features (Z-axis) corresponding to all operating conditions (X-axis) under a fixed load value. Multi-scale slicing here means segmenting the time-series curve under the same load according to different numbers of operating conditions, thereby obtaining subsequences of different lengths. This allows for capturing both the subtle changes in operating condition switching within short spans, such as the minute changes in acoustic signature features during the elevator's start-up and acceleration processes, and the overall trend over long spans, such as the general direction of the elevator's acoustic signature features throughout its entire operating cycle.

[0129] For example, in small-scale slicing, data from two consecutive operating conditions are extracted each time, such as the elevator transitioning from a starting state to an acceleration state, or from an acceleration state to a constant speed state. In this case, the focus is on the detailed characteristics during the transition between adjacent operating conditions. By analyzing the abrupt change pattern of the voiceprint during the transition from acceleration to constant speed, we can gain a deeper understanding of the dynamic characteristics of the elevator during the local operating state transition. In medium-scale slicing, data from three to four consecutive operating conditions are extracted each time, such as the elevator transitioning from a starting state, through an acceleration state, and then to a constant speed state. During this process, we focus on the coordinated change trend of multiple consecutive operating conditions. In large-scale slicing, data from the entire cycle of operating conditions is extracted to capture the overall temporal pattern of the complete operating process.

[0130] These slices, spanning different operating conditions, constitute a multi-scale slice. Smaller scale slices cover fewer operating conditions, allowing for a focus on localized, detailed changes, such as the precise capture of transitions between adjacent operating conditions. Larger scale slices cover a wider range of operating conditions, reflecting overall trends, such as the grasp of the entire operational cycle's patterns. This multi-scale slicing process enables the comprehensive extraction of acoustic signature information during elevator operation from multiple levels.

[0131] Temporal convolution is performed on each slice of data obtained after slicing. Temporal convolution can extract the first trend feature from these subsequence data. This feature can reflect the changing trend of voiceprint features within the range of specific load and local working conditions.

[0132] The first trend features obtained after temporal convolution are concatenated. Through this concatenation operation, the features extracted from slices at different scales are integrated into a comprehensive lateral temporal feature. This lateral temporal feature incorporates the changes in voiceprint features during various operating condition switching processes under different loads of the elevator.

[0133] To capture the dynamic patterns of voiceprint features (Z-axis) spatial distribution as the elevator load (Y-axis) changes during elevator operation, a built-in perceptual convolutional network extracts vertical spatial features along the YZ plane of the first point cloud space. Specifically:

[0134] In a preferred embodiment, vertical spatial features are extracted along the YZ plane of the first point cloud space using a built-in perceptual convolutional network, including:

[0135] Multi-scale data slicing was performed on the X-axis direction of the first point cloud space to obtain several second slice data for characterizing the variation law of voiceprint features under different elevator loads under the same operating conditions.

[0136] Dilated convolution is performed on each second slice of data to obtain the second trend feature;

[0137] By concatenating all the second-trend features, the vertical spatial features are obtained.

[0138] In one embodiment of the present invention, multi-scale data slicing processing is performed along the X-axis direction of the first point cloud space. Slicing along the X-axis essentially involves extracting the voiceprint features (Z-axis) corresponding to all elevator loads (Y-axis) under a fixed operating condition. Here, multi-scale slicing means cutting relevant data under the same operating condition according to different load variation ranges or intervals, thereby obtaining subsets of data with different ranges, i.e., several second slices of data used to characterize the variation patterns of voiceprint features under different elevator loads under the same operating condition.

[0139] For example, in small-scale slicing, data is extracted from a relatively small range of load variation each time. For instance, when the elevator is running at a constant speed, the focus is on the acoustic signature characteristics corresponding to load variations within a small interval. In this case, the focus is on the changes in acoustic signature characteristics caused by subtle load changes. By analyzing these subtle changes, we can gain a deeper understanding of the impact of local load fluctuations on acoustic signature characteristics under specific operating conditions. In medium-scale slicing, data from a medium range of load variation is extracted each time. For example, when the elevator is accelerating, data is extracted from load variations within a certain interval, focusing on the co-change relationship between load variations within this range and acoustic signature characteristics. In large-scale slicing, data from the entire load range under this operating condition is extracted to grasp the overall change pattern of acoustic signature characteristics across the entire load variation range.

[0140] These slices, representing different load variation ranges, constitute a multi-scale slice. Smaller scale slices cover a narrower range of load variations, focusing more intently on subtle local load changes; larger scale slices cover a wider load range, reflecting the overall load variation trend. This multi-scale slicing process allows for the comprehensive acquisition of acoustic signature characteristics corresponding to different loads under the same operating condition from multiple perspectives.

[0141] Dilated convolution is performed on each second slice of data obtained after slicing. Dilated convolution can expand the receptive field without increasing the number of parameters, thus better capturing long-term dependencies and trends in the data. Through dilated convolution, a second trend feature can be extracted from the second slice data, which can reflect the changing trend of acoustic signature features under specific operating conditions and within the range of local load variations.

[0142] All the secondary trend features obtained after dilated convolution are concatenated. Through this concatenation operation, the features extracted from slices at different scales are integrated into a comprehensive vertical spatial feature. This vertical spatial feature incorporates the changes in voiceprint features during different load variations under the same operating conditions of the elevator.

[0143] To capture the complex and dynamic relationship between elevator operating conditions (X-axis), elevator load (Y-axis), and voiceprint features (Z-axis) during elevator operation, a built-in first 3D convolutional network extracts joint distribution features along the XYZ space of the first point cloud space. Specifically:

[0144] In a preferred embodiment, joint distribution features are extracted along the XYZ space of the first point cloud space using a built-in first 3D convolutional network, including:

[0145] Multi-scale data space cutting processing is performed on the Z-axis direction of the first point cloud space to obtain several associated spatial data that characterize the operating conditions and elevator load distribution within the same voiceprint feature range.

[0146] Perform 3D convolution on each associated spatial data to obtain associated features;

[0147] By concatenating all associated features, a joint distribution feature is obtained.

[0148] In one embodiment of the present invention, multi-scale data spatial segmentation processing is performed along the Z-axis direction of the first point cloud space. The segmentation operation along the Z-axis essentially involves analyzing the distribution of operating conditions (X-axis) and elevator load (Y-axis) within a fixed range of acoustic signature features. Here, multi-scale data spatial segmentation means segmenting according to different acoustic signature feature intervals or intervals to obtain subspace data of different ranges, i.e., several associated spatial data used to characterize the distribution of operating conditions and elevator load within the same acoustic signature feature range.

[0149] For example, in small-scale segmentation, data is extracted from a relatively narrow range of voiceprint features each time. For instance, when the voiceprint features are within a small numerical range, the focus is on the specific distribution of different elevator operating conditions and loads within that range. The focus is on the combination of operating conditions and loads corresponding to subtle changes in the voiceprint features. By analyzing these local distributions, the correlation between elevator operating conditions and loads within a specific voiceprint feature range can be understood. In medium-scale segmentation, data from a medium range of voiceprint features is extracted each time. For example, when the voiceprint features change within a certain range, the distribution patterns of operating conditions and loads within that range are analyzed, focusing on the coordinated changes between the two. In large-scale segmentation, data from the entire voiceprint feature range is extracted to grasp the overall distribution patterns of operating conditions and loads across the entire range of voiceprint feature variations.

[0150] These segmentations of different acoustic signature ranges constitute multi-scale segmentation. Smaller scales cover a narrower range of acoustic signature features, allowing for a more focused analysis of subtle local changes in these features and their corresponding operating conditions and load distributions. Larger scales cover a wider range of acoustic signature features, better reflecting the overall distribution trends of operating conditions and loads under global changes in acoustic signature features. This multi-scale segmentation process enables the comprehensive acquisition of operating condition and elevator load distribution information within the same acoustic signature range from multiple perspectives.

[0151] Each segmented, associated spatial data point is processed using 3D convolution. 3D convolution considers information from three dimensions of space simultaneously. By performing convolution on the associated spatial data, the complex relationships between operating conditions, elevator load, and acoustic signature features can be fully explored, thus yielding associated features. These features can reflect the inherent connections and distribution characteristics of operating conditions and loads within a specific acoustic signature feature range.

[0152] All associated features obtained after 3D convolution are concatenated. Through this concatenation operation, features extracted at different scales are integrated into a comprehensive joint distribution feature. This joint distribution feature incorporates the operating conditions and load distribution information corresponding to different acoustic signature ranges throughout the elevator's operation.

[0153] After extracting horizontal temporal features, vertical spatial features, and joint distribution features, in order to more comprehensively and deeply explore the information during elevator operation and further grasp the complex relationship between elevator operating status and voiceprint features, these features are fused. Specifically:

[0154] In a preferred embodiment, broadcast multiplication is performed on the horizontal temporal features and the vertical spatial features to generate spatiotemporal interactive features, including:

[0155] Align the horizontal temporal features and vertical spatial features according to the feature dimensions to obtain the adapted horizontal temporal features and adapted vertical spatial features;

[0156] Tensor outer product operation is performed on the adapted horizontal temporal features and the adapted vertical spatial features to obtain spatiotemporal interaction features.

[0157] In one embodiment of the present invention, firstly, the horizontal temporal features reflect the temporal evolution of the acoustic signature features as the operating conditions continuously change under a fixed load; the vertical spatial features reflect the correlation between load changes and acoustic signature features under specific operating conditions. In other words, the information such as the range of operating condition changes and time span covered by the horizontal temporal features differs in dimension from the load change information covered by the vertical spatial features. Therefore, it is necessary to perform dimensional analysis and mapping on the horizontal temporal features and vertical spatial features according to their feature dimensions to achieve alignment and ensure that the horizontal temporal features and vertical spatial features can be subsequently calculated and analyzed within the same dimensional framework. Specifically, by constructing a shared dimensional mapping matrix, the horizontal temporal features and vertical spatial features are projected onto the same high-dimensional latent space. During the mapping process, the matrix parameters are not fixed but dynamically adjusted through extensive data training and optimization algorithms. During training, the goal is to minimize the information loss of the two features in the new dimensional space and maximize their matching degree in the new space, continuously optimizing the matrix parameters. After such optimization, the two transformed features can fully retain their core information in the new dimensional space and achieve complete dimensional matching. After alignment, the adapted horizontal temporal features and adapted vertical spatial features were obtained.

[0158] Tensor outer product operations are performed on the adapted lateral temporal features and the adapted vertical spatial features. Tensor outer product operations enable comprehensive combination and interaction of elements from both features. Through this operation, the potential connections between the lateral temporal features regarding the audioprint temporal information under elevator operating condition switching and load changes, and the vertical spatial features regarding the spatial distribution of elevator load and audioprint features, and the correlation distribution between operating conditions and audioprint features, are fully explored. The resulting spatiotemporal interactive features integrate key information from both time and space dimensions, providing a more comprehensive and detailed reflection of the complex characteristics of elevator operation.

[0159] The spatiotemporal interaction features contain the evolution law of elevator voiceprint features under dynamic combinations of load and operating conditions. It includes the temporal changes of voiceprint when different operating conditions are continuously switched under the same load, as well as the voiceprint differences under different loads for the same operating condition. It fully presents the temporal changes of operating conditions in the time dimension, and the voiceprint characteristics under the cross-effect of load differences in the spatial dimension. The joint distribution features focus on the global correlation between load, operating condition and voiceprint. The spatiotemporal interaction features and joint distribution features are spliced ​​along the channel dimension. This can not only retain the dynamic details of voiceprint under specific load and operating condition combinations, but also integrate the global distribution information of the three correlations, providing a richer data foundation for subsequent feature processing.

[0160] After the splicing is completed, the built-in second 3D convolutional network performs local feature aggregation. The 3D convolutional network has powerful feature extraction and integration capabilities, which can perform in-depth analysis and aggregation of local information in the spliced ​​features, extract more representative and key features, and generate the first fused feature. This feature further integrates the effective information in the spatiotemporal interaction features and joint distribution features, and expresses the relevant features of the elevator's operating status more concisely.

[0161] Finally, through the decoding layer built into the model, based on the knowledge and patterns learned in the previous training, the first fusion feature is analyzed to determine the fault risk value, providing an intuitive quantitative indicator for assessing the operational safety of the elevator.

[0162] The training process of the elevator failure risk assessment model is explained in detail below:

[0163] First, acquire elevator operation data under normal operation and different fault scenarios. This data covers the acoustic characteristics of the elevator under different elevator loads throughout the entire cycle of operation.

[0164] Professional elevator maintenance personnel and experts annotate the collected operating data, and assign a corresponding fault risk value label to each data sample based on the actual operating status and fault conditions of the elevator.

[0165] Based on the acoustic characteristics of the elevator under different elevator loads throughout the entire cycle of operation, a second point cloud space is constructed with the operation condition as the X-axis, the elevator load as the Y-axis, and the acoustic characteristics as the Z-axis.

[0166] Initialize the model parameters of the elevator failure risk assessment model to be trained;

[0167] The second point cloud space is input into the elevator fault risk assessment model to be trained for iterative training. The elevator fault risk assessment model extracts horizontal temporal sample features along the XZ plane of the second point cloud space through the built-in temporal convolutional network; extracts vertical spatial sample features along the YZ plane of the second point cloud space through the built-in perceptual convolutional network; extracts joint distributed sample features along the XYZ space of the second point cloud space through the built-in first 3D convolutional network; broadcast multiplication is performed on the horizontal temporal sample features and the vertical spatial sample features to generate spatiotemporal interactive sample features; the spatiotemporal interactive sample features and the joint distributed sample features are concatenated along the channel dimension, and then local feature aggregation is performed through the built-in second 3D convolutional network to generate a second fusion feature; the fault prediction risk value is determined based on the second fusion feature.

[0168] During each training iteration, a loss function is calculated based on the predicted fault risk value and its label to measure the difference between the predicted and labeled true fault risk values. A smaller loss value indicates that the model's prediction is closer to reality. Based on the calculated loss value, the error is propagated back from the output layer to the input layer using a backpropagation algorithm, calculating the gradient of each parameter. Then, an optimization algorithm (such as the Adam optimizer) is used to update the model's parameters based on the gradient information to reduce the loss value.

[0169] The process of forward propagation, loss calculation, backpropagation, and parameter update is repeated, continuously adjusting the model parameters until the loss function converges, resulting in a well-trained elevator failure risk assessment model. At this point, the elevator failure risk assessment model has learned the inherent patterns and correlations between elevator operating characteristics and failure risks, enabling it to predict elevator failure risks relatively accurately.

[0170] Step S5: If the fault risk value is greater than the preset fault threshold, issue a fault warning signal.

[0171] For step S5, after determining the fault risk value, it is compared with the preset fault threshold. The preset fault threshold is set based on a large amount of historical data, professional knowledge of elevator operation, and actual operation and maintenance experience.

[0172] When the calculated fault risk value exceeds the preset fault threshold, it indicates that the elevator's current operating status is abnormal. At this time, the system will immediately issue a fault warning signal to promptly notify relevant operation and maintenance personnel and managers to conduct further inspections and diagnoses of the elevator, so as to take corresponding maintenance and repair measures before the fault actually occurs, thereby avoiding the occurrence of the fault or mitigating its potential impact.

[0173] In a preferred embodiment, the preset fault threshold is determined in the following manner:

[0174] Obtain the design parameters and current service life of the elevator to be tested;

[0175] Determine the initial threshold based on the design parameters;

[0176] Calculate the degradation coefficient under the current service life based on the preset life distribution of the elevator to be tested;

[0177] Determine the preset fault threshold based on the degradation coefficient and the initial threshold.

[0178] In one embodiment of the present invention, the design parameters and current service life of the elevator to be tested are first obtained. The design parameters include key parameters such as the elevator's rated load, operating speed, motor power, and control system configuration. These parameters reflect the performance and capability range of the elevator at the time of its initial design. The current service life clarifies the length of time the elevator has been in use since it was put into service.

[0179] Based on the obtained elevator design parameters and combined with the technical standards of the elevator industry, an initial fault threshold is determined. This initial threshold is a reference value based on the ideal design state of the elevator and represents the critical point of fault risk for the elevator under new and normal operating conditions.

[0180] Considering that elevator performance degrades over time during use, the degradation coefficient for the current service life is calculated based on the elevator's preset lifespan distribution. This preset lifespan distribution is established based on the elevator's design lifespan, statistical data on the actual service life of similar elevators, and theoretical frameworks, describing the probability distribution of elevator malfunctions at different stages of use. By analyzing the preset lifespan distribution and combining it with the elevator's current service life, the degree of performance degradation relative to a brand-new state can be quantified, and the degradation coefficient can be calculated. For example, if the elevator's preset lifespan is 20 years and it has been in use for 10 years, according to the preset lifespan distribution model, the degradation coefficient might be 0.6, meaning the elevator's performance has degraded to 60% of its brand-new level.

[0181] Finally, the preset fault threshold is determined based on the calculated degradation coefficient and the initial threshold. Since elevator performance gradually degrades with age, its operating pressure tolerance and fault risk threshold also decrease accordingly. Therefore, the initial threshold is multiplied by the degradation coefficient to obtain a preset fault threshold that adapts to the elevator's current actual condition. For example, if the initial threshold is set to 80 and the degradation coefficient is 0.6, then the preset fault threshold is 48.

[0182] This invention calculates a degradation coefficient based on a preset lifespan distribution and the current service life, and uses the degradation coefficient to adjust the initial threshold, thereby scientifically and rationally determining the preset fault threshold. It fully considers the performance degradation of the elevator during use, making the preset fault threshold more closely match the current actual state of the elevator. This helps to detect potential elevator fault risks more promptly and accurately, and ensures the safe and stable operation of the elevator.

[0183] like Figure 2 As shown, based on the above method embodiments, corresponding system embodiments are provided;

[0184] An embodiment of the present invention provides an elevator fault early warning system based on voiceprint recognition, including: a voiceprint acquisition module, a voiceprint feature extraction module, a point cloud spatial construction module, a risk assessment module, and a fault early warning module;

[0185] The acoustic signature acquisition module is used to acquire the traction machine operating acoustic signature of the elevator under different elevator loads throughout the entire operating cycle. The entire operating cycle includes: starting, acceleration, constant speed, deceleration and braking.

[0186] The voiceprint feature extraction module is used to perform wavelet transform on the voiceprint of the traction machine to extract voiceprint features;

[0187] The point cloud space construction module is used to construct the first point cloud space based on the acoustic features of the traction machine's operating acoustic features under different elevator loads and the full-cycle operating conditions, with the operating conditions as the X-axis, the elevator load as the Y-axis, and the acoustic features as the Z-axis.

[0188] The risk assessment module is used to input the first point cloud space into the trained elevator fault risk assessment model. This allows the elevator fault risk assessment model to extract lateral temporal features along the XZ plane of the first point cloud space using a built-in temporal convolutional network; extract vertical spatial features along the YZ plane of the first point cloud space using a built-in perceptual convolutional network; extract joint distribution features along the XYZ space of the first point cloud space using a built-in first 3D convolutional network; perform broadcast multiplication on the lateral temporal features and vertical spatial features to generate spatiotemporal interactive features; concatenate the spatiotemporal interactive features and joint distribution features along the channel dimension; and then perform local feature aggregation using a built-in second 3D convolutional network to generate a first fusion feature; finally, determine the fault risk value based on the first fusion feature.

[0189] The fault warning module is used to issue a fault warning signal when the fault risk value exceeds the preset fault threshold.

[0190] In a preferred embodiment, the elevator fault early warning system based on voiceprint recognition further includes: a voiceprint noise reduction module;

[0191] The voiceprint noise reduction module includes: a noise signal generation unit, a noise frequency domain transformation unit, a voiceprint frequency domain transformation unit, a spectral subtraction noise reduction unit, and a voiceprint time domain reconstruction unit;

[0192] The noise signal generation unit is used to call the probability distribution of the corresponding noise parameter from the preset noise parameter probability distribution library according to the elevator load and operating conditions, and to generate a noise signal by random sampling.

[0193] The noise frequency domain transformation unit is used to perform a short-time Fourier transform on the noise signal to obtain the noise spectrum;

[0194] The acoustic signature frequency domain transformation unit is used to perform short-time Fourier transform on the acoustic signature of the traction machine to obtain the acoustic signature spectrum;

[0195] The spectral subtraction noise reduction unit is used to subtract the noise spectrum from the acoustic spectrum through spectral subtraction processing to obtain the noise-reduced acoustic spectrum.

[0196] The acoustic signature time-domain reconstruction unit is used to perform inverse short-time Fourier transform on the denoised acoustic signature spectrum to obtain the denoised traction machine operating acoustic signature.

[0197] It is understood that the above system embodiments correspond to the method embodiments of the present invention, and can implement the elevator fault early warning method based on voiceprint recognition provided by any of the above method embodiments of the present invention.

[0198] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0199] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for early warning of elevator malfunctions based on voiceprint recognition, characterized in that, include: Acquire the traction machine operating sound signature of the elevator under test under different elevator loads throughout its entire operating cycle. The full-cycle operating conditions include: start-up, acceleration, constant speed, deceleration, and braking; Wavelet transform is performed on the acoustic fingerprint of the traction machine to extract its features; Based on the acoustic characteristics of the traction machine's operating acoustics under different elevator loads throughout the entire operating cycle, the first point cloud space is constructed with the operating conditions as the X-axis, the elevator load as the Y-axis, and the acoustic characteristics as the Z-axis. The first point cloud space is input into the trained elevator fault risk assessment model, so that the elevator fault risk assessment model extracts lateral temporal features along the XZ plane of the first point cloud space through the built-in temporal convolutional network; extracts vertical spatial features along the YZ plane of the first point cloud space through the built-in perceptual convolutional network; extracts joint distribution features along the XYZ space of the first point cloud space through the built-in first 3D convolutional network; broadcast multiplication is performed on the lateral temporal features and vertical spatial features to generate spatiotemporal interaction features; the spatiotemporal interaction features and joint distribution features are concatenated along the channel dimension, and then local feature aggregation is performed through the built-in second 3D convolutional network to generate a first fusion feature; the fault risk value is determined based on the first fusion feature. When the fault risk value exceeds the preset fault threshold, a fault warning signal is issued.

2. The elevator fault early warning method based on voiceprint recognition according to claim 1, characterized in that, After obtaining the traction machine operating sound fingerprints of the elevator under different elevator loads throughout its entire operating cycle, the process also includes: Based on the elevator load and operating conditions, the probability distribution of the corresponding noise parameter is retrieved from the preset noise parameter probability distribution library; For each operating condition, based on the probability distribution of noise parameters, a combination of noise parameters conforming to the probability distribution is generated by random sampling; the combination of noise parameters includes: amplitude, frequency, phase, and bandwidth; The spectral shape of the noise signal is determined based on the frequency and bandwidth obtained from the sampling. Based on the amplitude obtained from sampling, the shape of the spectrum is adjusted to obtain the frequency domain signal; Perform an inverse Fourier transform on the frequency domain signal to obtain the time domain signal; Based on the sampled phase, the time-domain signal is phase-adjusted to generate a noise segment for the operating condition; Noise segments from all operating conditions are spliced ​​together to obtain a continuous noise signal; The noise signal is subjected to a short-time Fourier transform to obtain the noise spectrum; The acoustic waveform of the traction machine is subjected to a short-time Fourier transform to obtain the acoustic waveform spectrum. The noise spectrum is subtracted from the acoustic signature spectrum using spectral subtraction to obtain the noise-reduced acoustic signature spectrum. The inverse short-time Fourier transform of the noise-reduced acoustic spectrum is used to obtain the noise-reduced traction machine operating acoustics.

3. The elevator fault early warning method based on voiceprint recognition according to claim 1, characterized in that, Wavelet transform is performed on the acoustic fingerprint of the traction machine to extract acoustic fingerprint features, including: The operating acoustic fingerprint of the traction machine is divided according to the operating conditions to obtain corresponding acoustic fingerprint segments; For each voiceprint segment, the corresponding wavelet basis function is determined according to the operating condition of the voiceprint segment; Calculate the load factor based on the elevator load and rated load; Based on the load coefficients, the wavelet basis function is modified to obtain the modified wavelet basis function; Wavelet transform is performed on the voiceprint segments based on the modified wavelet basis function to obtain the corresponding voiceprint features.

4. The elevator fault early warning method based on voiceprint recognition according to claim 1, characterized in that, The built-in temporal convolutional network extracts lateral temporal features along the XZ plane of the first point cloud space, including: Multi-scale data slicing was performed on the Y-axis direction of the first point cloud space to obtain several first slice data that characterize the temporal variation law during the switching of different operating conditions of the same elevator load. Perform temporal convolution processing on each first slice of data to obtain the first trend feature; By concatenating all primary trend features, we obtain the lateral time series features.

5. The elevator fault early warning method based on voiceprint recognition according to claim 4, characterized in that, The built-in perceptual convolutional network extracts vertical spatial features along the YZ plane of the first point cloud space, including: Multi-scale data slicing was performed on the X-axis direction of the first point cloud space to obtain several second slice data for characterizing the variation law of voiceprint features under different elevator loads under the same operating conditions. Dilated convolution is performed on each second slice of data to obtain the second trend feature; By concatenating all the second-trend features, the vertical spatial features are obtained.

6. The elevator fault early warning method based on voiceprint recognition according to claim 5, characterized in that, The built-in first 3D convolutional network extracts joint distribution features along the XYZ space of the first point cloud space, including: Multi-scale data space cutting processing is performed on the Z-axis direction of the first point cloud space to obtain several associated spatial data that characterize the operating conditions and elevator load distribution within the same voiceprint feature range. Perform 3D convolution on each associated spatial data to obtain associated features; By concatenating all associated features, a joint distribution feature is obtained.

7. The elevator fault early warning method based on voiceprint recognition according to claim 6, characterized in that, Broadcast multiplication is performed on horizontal temporal features and vertical spatial features to generate spatiotemporal interactive features, including: Align the horizontal temporal features and vertical spatial features according to the feature dimensions to obtain the adapted horizontal temporal features and adapted vertical spatial features; Tensor outer product operation is performed on the adapted horizontal temporal features and the adapted vertical spatial features to obtain spatiotemporal interaction features.

8. The elevator fault early warning method based on voiceprint recognition according to claim 1, characterized in that, The preset fault threshold is determined in the following way: Obtain the design parameters and current service life of the elevator to be tested; Determine the initial threshold based on the design parameters; Calculate the degradation coefficient under the current service life based on the preset life distribution of the elevator to be tested; The preset fault threshold is determined based on the degradation coefficient and the initial threshold.

9. An elevator fault early warning system based on voiceprint recognition, characterized in that, include: The system includes a voiceprint acquisition module, a voiceprint feature extraction module, a point cloud spatial construction module, a risk assessment module, and a fault early warning module. The soundprint acquisition module is used to acquire the traction machine operating soundprint of the elevator under different elevator loads throughout the entire cycle of operation. The full-cycle operating conditions include: start-up, acceleration, constant speed, deceleration, and braking; The voiceprint feature extraction module is used to perform wavelet transform on the operating voiceprint of the traction machine to extract voiceprint features; The point cloud space construction module is used to construct the first point cloud space based on the acoustic features corresponding to the traction machine's operating acoustic features under different elevator loads and full-cycle operating conditions, with the operating conditions as the X-axis, the elevator load as the Y-axis, and the acoustic features as the Z-axis. The risk assessment module is used to input the first point cloud space into the trained elevator fault risk assessment model, so that the elevator fault risk assessment model extracts horizontal temporal features along the XZ plane of the first point cloud space through the built-in temporal convolutional network; extracts vertical spatial features along the YZ plane of the first point cloud space through the built-in perceptual convolutional network; extracts joint distribution features along the XYZ space of the first point cloud space through the built-in first 3D convolutional network; performs broadcast multiplication on the horizontal temporal features and vertical spatial features to generate spatiotemporal interactive features; concatenates the spatiotemporal interactive features and joint distribution features along the channel dimension, and then performs local feature aggregation through the built-in second 3D convolutional network to generate a first fusion feature; and determines the fault risk value based on the first fusion feature. The fault warning module is used to issue a fault warning signal when the fault risk value is greater than a preset fault threshold.

10. The elevator fault early warning system based on voiceprint recognition according to claim 9, characterized in that, It also includes: a voiceprint noise reduction module; The voiceprint noise reduction module includes: a noise signal generation unit, a noise frequency domain transformation unit, a voiceprint frequency domain transformation unit, a spectral subtraction noise reduction unit, and a voiceprint time domain reconstruction unit. The noise signal generation unit is used to call the probability distribution of the corresponding noise parameter from the preset noise parameter probability distribution library according to the elevator load and operating conditions, and perform random sampling to generate a noise signal. The noise frequency domain transformation unit is used to perform a short-time Fourier transform on the noise signal to obtain the noise spectrum; The acoustic frequency domain transformation unit is used to perform a short-time Fourier transform on the acoustic signature of the traction machine to obtain the acoustic signature spectrum. The spectral subtraction noise reduction unit is used to subtract the noise spectrum from the acoustic spectrum through spectral subtraction processing to obtain the noise-reduced acoustic spectrum. The acoustic signature time-domain reconstruction unit is used to perform inverse short-time Fourier transform on the denoised acoustic signature spectrum to obtain the denoised traction machine operation acoustic signature.

Images