Industrial data analysis abnormal sound fault judgment method

Abstract

The invention belongs to the technical field of industrial data analysis, and particularly relates to an industrial data analysis abnormal sound fault judgment method. According to the method, equipment reference fault sound source data and interference-containing mixed sound source data are acquired through a multi-dimensional distributed sensor array, and a multi-modal sound source database is constructed; performing interference separation on the mixed sound source data, calculating an energy entropy value, comparing de-noised data with reference data after data quality verification, and screening effective data; constructing a fault positioning model to determine space coordinates of fault points, and eliminating false positioning points; and fusing the space and sound source data by using a double-branch neural network, and outputting a fault category and probability. Compared with the prior art, the limitation of independently considering sound or position information is overcome, interference is effectively removed, data quality is verified, accurate fault positioning and category judgment are achieved, fault detection accuracy and maintenance efficiency are improved, and safe and stable operation of industrial equipment is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial data analysis, and particularly relates to a method for judging abnormal sound faults in industrial data analysis. Background Art

[0002] In the technical field of industrial data analysis, equipment fault detection has always been a key link to ensure the stable operation of industrial production. Taking an elevator as an example, in the past, when dealing with the judgment of equipment abnormal sound faults, most of the existing technical solutions had limitations and often considered sound or position separately. Purely based on sound analysis, it was impossible to accurately determine the location where the fault occurred, and it was difficult to quickly and effectively process and repair the fault. Some other technologies focused on position positioning and used methods such as measuring the time difference of signals arriving at different sensors by a sensor array to determine the fault location. However, this method ignored the rich fault feature information carried by sound data, and it was difficult to accurately judge the specific category of the fault only relying on position information. Summary of the Invention

[0003] The present invention aims at the technical problems existing in the background art and proposes a method for judging abnormal sound faults in industrial data analysis.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows, including the following steps:

[0005] S1. Synchronously collect reference fault sound source data and mixed sound source data with interference under non-interference conditions of the equipment through a multi-dimensional distributed sensor array, and construct a multi-modal sound source database;

[0006] S2. Separate the interference from the mixed sound source data to generate denoised target fault sound source data, and calculate the energy entropy value of the denoised target fault sound source data;

[0007] S3. Compare the denoised target fault data with the reference fault sound source data. When the residual index exceeds the threshold, discard this part of the data. Otherwise, enter the fault location stage. The specific judgment process is as follows:

[0008] S31. Convert the current target fault data and the reference fault sound source data into Mel spectrograms;

[0009] S32. Extract 256-dimensional deep feature vectors through a two-stream network;

[0010] S33. Calculate the weighted cosine similarity difference of five-layer feature maps to generate R ML ;

[0011] S34. Set a threshold T ML ;

[0012] S35. When R ML>T ML , it is determined as denoising failure data, and the operation of discarding the data is performed;

[0013] S36. When R ML ≤T ML , the fault location process will be carried out;

[0014] S4. Construct a fault location model to determine the spatial coordinates of the fault point;

[0015] S5. Combine the spatial coordinates of the fault point and the sound source data information to output the fault category and probability.

[0016] Preferably, the forms of interference described in step S1 include echo, noise, and human voice.

[0017] Preferably, the calculation method for separating the interference of the mixed sound source in step S2 is: where is the fault sound source data after denoising, W ψ is the wavelet packet transform operator, X(t) is the mixed sound source data, A is the mixed sound source matrix, S is the independent sound source component to be solved, and TV(S(t)) is the total variation regularization term.

[0018] Preferably, the calculation method for the energy entropy value of the target fault sound source data in step S2 is: where N is the number of layers of wavelet packet decomposition, satisfying N≥6, represents the data energy of the kth decomposition sub-band.

[0019] Preferably, a data quality verification stage is added before calculating the energy entropy value of the target fault sound source data and comparing the target denoised data with the reference fault sound source data to make a quality decision: When the decision is 1, the next comparison is carried out;

[0020] where Q is the quality evaluation index, and the calculation method is: where E norm is the normalized energy entropy, SNR is the signal-to-noise ratio, represents the data kurtosis index;

[0021] where the threshold Q th =0.7Q base , Q base is the average value under 10 tests as the reference value, and the gradient threshold ΔQ max =0.2·Q th / Δt.

[0022] Preferably, the specific implementation of step S4 is:

[0023] S41. Construct a fault location model, and the calculation formula is: where τ opt is used to calculate the distance difference between the fault point and the sensor, is the frequency-domain signal of the m-th sensor, is the complex conjugate of the reference sensor signal;

[0024] S42. Calculate the time difference for the fault sound source to reach each sensor according to the model;

[0025] S43. Establish a hyperbolic variable equation set according to the sensor spatial coordinates, and solve it to obtain the fault point coordinates.

[0026] Preferably, after obtaining the coordinates of the fault point, it is necessary to judge the coordinate points and eliminate the pseudo-positioning points that exceed the physical boundary of the device; the calculation method for judgment is: ||P(x,y,z)-P0||≤D max , where P(x,y,z) is the estimated fault point coordinates, P0 is the geometric center of the device, and D max is the maximum allowable offset, taking 20% of the characteristic size of the device.

[0027] Preferably, the implementation method of outputting the fault category probability in step S5 is to first input the spatial coordinates and sound source data information, use a dual-branch neural network to fuse the spatial and sound source data, and output the fault category and probability.

[0028] Compared with the prior art, the advantages and positive effects of the present invention are that multi-dimensional distributed sensor arrays are used to collect the sound source data of the device without interference and with interference to construct a multi-modal sound source database, a unique interference separation algorithm is used to remove interferences such as echoes, noises, and human voices and calculate the energy entropy value of the target fault sound source data, a data quality verification stage is added to comprehensively normalize the energy entropy, signal-to-noise ratio, and data kurtosis indicators for quality decision-making to avoid the influence of low-quality data, the fault point spatial coordinates are calculated by constructing a fault location model and pseudo-positioning points are eliminated, and a dual-branch neural network is used to fuse the spatial and sound source data to output the fault category and probability, so as to accurately locate the fault position, judge the fault category, improve the accuracy of fault detection and the maintenance efficiency, and ensure the safe and stable operation of the device. Description of the Drawings

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0030] Figure 1 It is the overall structure flow chart of the present invention. Detailed implementation mode

[0031] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0032] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the limitations of the specific embodiments disclosed in the following specification.

[0033] Embodiment. In order to accurately judge the location and category of abnormal noise faults of industrial equipment and solve the problem of judgment limitations caused by separately considering sound or position information in traditional technologies, the present invention proposes a method for judging abnormal noise faults in industrial data analysis. Taking an elevator as an example, during the operation of the elevator, the traditional fault judgment method either only relies on listening to the sound emitted during the operation of the elevator, and maintenance personnel rely on experience to distinguish the sound characteristics to roughly judge the fault. However, this method is difficult to accurately point out which component of the elevator the fault occurs in, whether it is the mechanical structure inside the car or a certain device in the hoistway. In a large building, when multiple elevators are running simultaneously, the limitations of this judgment method are more obvious; either simply rely on position detection technology, such as determining the fault position by measuring the time difference of signals reaching different position sensors through sensors, but ignore the rich fault clues contained in the sound. For example, when a fault occurs in the elevator traction machine, the difference in sound characteristics generated by different fault types can reflect the severity of the fault and the specific problem, but the position detection technology cannot utilize this information, resulting in the inability to accurately judge the fault category, affecting the maintenance efficiency and the safe operation of the elevator. The specific implementation process of the present invention is as Figure 1 shown.

[0034] When the present invention is applied to an elevator, first, the reference fault sound source data of the elevator under non-interference conditions (such as simulating normal operating conditions) and the mixed sound source data containing various interferences (such as noisy background noise in the machine room, well shaft echoes, etc.) during actual operation are collected, and a multi-modal sound source database is constructed. For example, when the elevator is operating under no-load and full-load conditions and stopping at different floors, the collected sound data is classified and stored in the database.

[0035] Next, considering that the technology often lacks effective interference separation means when processing mixed sound source data. The common practice is to simply perform filtering, but this method can only remove some interferences with obvious frequency characteristics, and it is impossible to accurately remove complex and diverse interferences such as echoes, noise, and human voices. The present invention performs interference separation on the mixed sound source data, and the calculation formula is:

[0036] Among them is the fault sound source data after denoising, W ψ is the wavelet packet transform operator, X(t) is the mixed sound source data, A is the mixed sound source matrix, S is the independent sound source component to be solved, and TV(S(t)) is the total variation regularization term. At the same time, the calculation method of the energy entropy value of the target fault sound source data is as follows: where N is the number of layers of wavelet packet decomposition, satisfying N≥6, represents the data energy of the kth decomposition subband. By improving the interference separation algorithm and introducing a scientific energy entropy value calculation method, the efficient removal of various interferences in the mixed sound source data and the comprehensive and accurate extraction of the characteristics of the target fault sound source data are realized. This process can effectively extract the sound characteristics truly related to elevator faults and exclude the influence of external interference factors. For example, in a noisy machine room environment, the sound data after interference separation can clearly present the subtle abnormal friction sounds of the traction machine, which may be masked in the original mixed data, providing key information for accurate fault judgment.

[0037] Before entering the fault location and category judgment, the present invention adds a data quality verification stage. Make a quality decision: When the decision is 1, the following comparison is carried out; among them, Q is the quality evaluation index, and the calculation method is: where E norm is the normalized energy entropy, SNR is the signal-to-noise ratio, represents the data kurtosis index; among them, the threshold Q th = 0.7Q base , Q base is the average value under 10 tests as the reference value, and the gradient threshold ΔQ max = 0.2·Q th / Δt. This link ensures that the data used for fault judgment has high reliability and avoids misjudgment caused by data quality problems. For example, if the data collected in a certain acquisition is affected by a sudden strong interference and is found to not meet the requirements after quality verification, then the data is discarded and does not participate in subsequent analysis, thus ensuring the accuracy of the fault judgment result.

[0038] Next, compare the target fault data after denoising with the reference fault sound source data. Discard this part of the data when the residual index exceeds the threshold. The comparison method is to first convert the current target fault data after denoising and the reference fault sound source data into Mel spectrograms. The Mel spectrogram is a spectral representation method based on the auditory characteristics of the human ear. It can better reflect the energy distribution of the sound signal at different frequencies and is closer to the human ear's perception of sound. Through this conversion, the originally complex sound data is transformed into a spectral form that is easier to analyze and understand. Then, use a two-stream network to extract 256-dimensional deep feature vectors. The two-stream network is a model structure widely used in the field of deep learning. It can extract features from data from different perspectives simultaneously, improving the comprehensiveness and accuracy of feature extraction. Here, the two-stream network fully mines the deep features in the Mel spectrogram and condenses them into 256-dimensional feature vectors. These 256-dimensional deep feature vectors contain rich information about the sound data, covering features in multiple aspects such as frequency, amplitude, and phase. After that, calculate the weighted cosine similarity difference of the five-layer feature maps and generate a specific index R ML . In a deep learning model, feature maps are the results obtained after the network extracts features from the input data at different layers. By calculating the weighted cosine similarity difference of the five-layer feature maps, the similarity between the target fault data after denoising and the reference fault sound source data can be compared more carefully. The weighted cosine similarity takes into account the differences in the importance of different feature dimensions, making the calculation result more in line with the actual situation. To accurately judge the reliability of the data, a reasonable threshold T ML also needs to be set. When the generated index R ML is greater than the threshold T ML , it is determined as denoising failure data. This indicates that the difference between the target fault data after denoising and the reference fault sound source data is too large, and the reliability of the data is in doubt. It may be affected by unknown interference or there may be problems in the denoising process. At this time, to ensure the accuracy of subsequent fault judgment, an operation to discard the data needs to be performed to avoid the negative impact of this unreliable data on the final result. When the index R ML is less than or equal to the threshold T ML , it means that the target fault data after denoising and the reference fault sound source data have a high similarity, and the reliability of the data is high, and it can enter the subsequent fault location process.

[0039] In the fault location process, first construct a fault location model. The calculation formula is: where τ opt is used to calculate the distance difference between the fault point and the sensor, is the frequency-domain signal of the m-th sensor, is the complex conjugate of the reference sensor signal; then calculate the time difference between the fault sound source reaching each sensor according to the model; establish a hyperbolic variable equation set based on the sensor spatial coordinates where the sensor spatial coordinates are (x m , y m , z m ), c is the speed of light, and Δσ m1 = σ m -σ1 is the time difference for calculating the sound source reaching each sensor. During the solution process, it is necessary to judge the coordinate points and eliminate the pseudo-positioning points beyond the physical boundary of the device; the calculation method for judgment is: ||P(x, y, z)-P0||≤D max , where P(x, y, z) is the estimated fault point coordinate, P0 is the geometric center of the device, and D max is the maximum allowable offset, taking 20% of the characteristic size of the device. Then, perform clustering analysis on the solution set that meets the constraints, and take the weighted centroid as the final fault location.

[0040] Finally, take the spatial coordinates of the fault point and the sound source data information as inputs. The spatial coordinate information reflects the position distribution of the fault in the device, and the sound source data contains the characteristics of the fault sound. Then, the dual-branch neural network starts to work, which has a spatial data branch and a sound source data branch. The spatial data branch receives the spatial coordinate information, extracts features through a series of convolutional layers, pooling layers, etc., and mines the spatial position features, such as the relative position relationship between the fault point and the key components of the device. The sound source data branch processes the sound source data, and also uses convolutional and pooling operations to obtain features such as the frequency and amplitude of the sound. Then, fuse the features extracted by the two branches. Generally, use the splicing method to connect the feature vectors of different branches in sequence to form a new feature vector containing both spatial and sound source information. Finally, input the fused feature vector into the fully connected layer for classification and probability calculation. The fully connected layer further integrates the features through the operations of the weight matrix and the bias term, and finally outputs the fault category and the probability corresponding to each category, realizing accurate fault category judgment and probability prediction, providing clear repair directions and fault possibility references for maintenance personnel. Applied to the elevator industry, it realizes the accurate positioning and category judgment of elevator abnormal noise faults, greatly improves the accuracy of fault detection and the repair efficiency, and effectively guarantees the safe and stable operation of the elevator.

[0041] The above is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for judging abnormal noise faults in industrial data analysis, characterized in that, It includes the following steps: S1. Synchronously collect the benchmark fault sound source data and the mixed sound source data with interference under the condition of no interference of the device through a multi-dimensional distributed sensor array, and construct a multi-modal sound source database; S2. Separate the interference from the mixed sound source data to generate the target fault sound source data after denoising, and calculate the energy entropy value of the target fault sound source data after denoising; S3. Compare the target fault data after denoising with the benchmark fault sound source data. When the residual index exceeds the threshold, discard this part of the data. Otherwise, enter the fault location stage. The specific judgment process is as follows: S31. Convert the current target fault data and the benchmark fault sound source data into Mel spectrograms; S32. Extract 256-dimensional deep feature vectors through a two-stream network; S33. Calculate the weighted cosine similarity difference of the five-layer feature maps to generate R ML ; S34. Set the threshold value T ML ; S35. When R ML >T ML , it is determined as denoising failure data, and the operation of discarding data is performed; S36. When R ML ≤ T ML , the fault location process will be carried out; S4. Construct a fault location model to determine the spatial coordinates of the fault point; S5. Combine the spatial coordinates of the fault point and the sound source data information to output the fault category and probability.

2. The method for judging abnormal noise faults in industrial data analysis according to claim 1, wherein, The forms of interference described in step S1 include echo, noise, and human voice.

3. The method for judging abnormal noise faults in industrial data analysis according to claim 1, wherein, The calculation method for interfering separation of the mixed sound source in step S2 is as follows: where is the fault sound source data after denoising, W ψ is the wavelet packet transform operator, X(t) is the mixed sound source data, A is the mixed sound source matrix, S is the independent sound source component to be solved, and TV(S(t)) is the total variation regularization term.

4. The method for judging abnormal noise faults in industrial data analysis according to claim 1, characterized in that The calculation method of the energy entropy value of the target fault sound source data in step S2 is as follows: where N is the number of layers of wavelet packet decomposition, satisfying N≥6, represents the data energy of the k-th decomposed sub-band.

5. The method for judging abnormal noise faults in industrial data analysis according to claim 1, characterized in that, A data quality verification stage is added before calculating the energy entropy value of the target fault sound source data and before comparing the target denoised data and the reference fault sound source data to make a quality decision: When the decision is 1, the next comparison is carried out; Among them, Q is the quality evaluation index, and the calculation method is as follows: where E norm is the normalized energy entropy, SNR is the signal-to-noise ratio, represents the data kurtosis index; Among them, the threshold Q th = 0.7Q base , Q base is the average value under 10 tests as the reference value, and the gradient threshold ΔQ max = 0.2·Q th / Δt.

6. The method for judging abnormal noise faults in industrial data analysis according to claim 1, characterized in that, The specific implementation of step S4 is as follows: S41. Construct a fault location model, and the calculation formula is: where τ opt is used to calculate the distance difference between the fault point and the sensor, is the frequency domain signal of the m-th sensor, is the complex conjugate of the reference sensor signal; S42. Calculate the time difference between the fault sound source reaching each sensor according to the model; S43. Establish a hyperbolic variable equation set according to the sensor spatial coordinates, and solve it to obtain the fault point coordinates.

7. The method for judging abnormal noise faults in industrial data analysis according to claim 6, wherein After obtaining the coordinates of the fault point, it is necessary to judge the coordinate points and eliminate the pseudo-location points that exceed the physical boundaries of the equipment. The calculation method for the judgment is: ||P(x,y,z)-P0||≤D max , where P(x,y,z) is the estimated coordinate of the fault point, P0 is the geometric center of the equipment, and D max is the maximum allowable offset, which is taken as 20% of the characteristic size of the equipment.

8. A method for judging abnormal noise faults in industrial data analysis according to claim 1, characterized in that, The implementation method of outputting the fault category probability in step S5 is to first input the spatial coordinates and the sound source data information, and use a two-branch neural network to fuse the spatial and sound source data to output the fault category and probability.

Images