CN-121977801-A - Belt conveyor idler intelligent monitoring system based on double-flow characteristic fusion countermeasure self-encoder

CN121977801ACN 121977801 ACN121977801 ACN 121977801ACN-121977801-A

Abstract

The invention discloses an intelligent monitoring system for a belt conveyor idler based on a double-flow feature fusion countermeasure self-encoder, which comprises a data acquisition module, a data segmentation module, a data preprocessing module, a model training module, a score evaluation module, an anomaly labeling module, a condition feedback module and a condition feedback module, wherein the data acquisition module acquires an idler vibration signal by using a distributed optical fiber sensing system (DAS), the data segmentation module carries out space and time division on the acquired vibration signal, the data preprocessing module carries out instantaneous frequency calculation, time-frequency feature extraction and scaling treatment on the segmented vibration signal, the model training module trains a reconstruction score based on the double-flow feature fusion countermeasure self-encoder model, the score evaluation module calculates the reconstruction score according to the vibration signal and a reconstruction signal output by the model, the anomaly labeling module carries out anomaly labeling on the reconstruction score based on an EWMA control chart, the anomaly diagnosis module carries out time statistics according to the anomaly labeling result, and the condition feedback module carries out condition feedback and model update according to the diagnosis result and on-site verification, and the applicability of the DAS in belt conveyor idler fault monitoring is improved.

Inventors

ZHANG XUPING
LI HONGREN
ZHANG YIXIN
WANG XINYU
Zou Ningmu
JIANG YANAN
WANG SHUN
SONG ZHUO

Assignees

南京大学

Dates

Publication Date: 20260505
Application Date: 20251201

Claims (10)

1. The intelligent belt conveyor idler monitoring system based on the double-flow characteristic fusion countermeasure self-encoder is characterized by comprising the following modules: the data acquisition module is used for deploying a single sensing optical fiber along a conveyor through a distributed optical fiber sensing system DAS and acquiring a continuous space-time vibration signal V of a carrier roller with a specified space length and a time length of H; the data segmentation module divides the vibration signals V acquired by the DAS according to the monitoring period T to obtain M vibration signals with the monitoring period T; the data preprocessing module is used for carrying out instantaneous frequency calculation, time-frequency characteristic extraction and scaling processing on the segmented vibration signals to obtain characteristic tensors And (3) with ; Model training module for training tensors corresponding to abnormal-free historical carrier roller vibration signals And (3) with Training based on double-flow characteristic fusion antagonism self-encoder model, obtaining tensor through double-branch decoder And (3) with Is a reconstruction result of (2) And (3) with And output; score evaluation module for evaluating the tensor of vibration signal And (3) with And reconstructing the result And (3) with Calculating reconstruction loss to obtain reconstruction score of each vibration signal; the abnormal labeling and diagnosing module is used for carrying out abnormal labeling on the reconstruction score based on the EWMA control chart and carrying out time statistics according to the abnormal labeling result to obtain an abnormal diagnosing result; And the state feedback module is used for carrying out state feedback according to the abnormal diagnosis result and the field verification and determining whether the double-current feature fusion countermeasure self-encoder model needs to be trained and updated.
2. The intelligent monitoring system for a carrier roller of a belt conveyor according to claim 1, wherein the data dividing module divides the vibration signal V with a monitoring period T to obtain M vibration signals with a monitoring period T ; The data dividing module selects N spatial positions from the specified spatial length, and each vibration signal is divided into N spatial positions Separating out the sequence vibration signals The process is expressed as: ; Wherein the function is The vibration signal V is divided to obtain a serial vibration signal 。
3. The intelligent monitoring system of a belt conveyor idler of claim 2, wherein the data preprocessing module comprises: Instantaneous frequency calculation sub-module for vibration signal Is a vibration signal of each sequence of (a) Sequentially performing inverse cosine transform, forward differential processing and one-dimensional third-order median filtering to obtain a filtered signal, and copying and splicing the tail value of the filtered signal to the tail of the filtered signal to form an instantaneous frequency signal Ensure that Data points The same; Time-frequency characteristic extraction submodule taking time window as time window The overlapping time length is Will be Divided into n segments, denoted as Will (i) be Divided into n segments, denoted as Obtaining And (3) with Frequency signal of (2) And (3) with The sizes are all Will (i) be And (3) with Are all spliced along the k direction to obtain And (3) with Corresponding tensor And (3) with The sizes are all ; Scaling processing submodule for tensor And (3) with Scaling to obtain the final product with the same size Tensors of (a) And For vibration signals The method comprises the following steps: ; ; Wherein the function is Representation pair And performing scaling processing.
4. The intelligent monitoring system of belt conveyor idlers according to claim 3, wherein the model training module processes as follows: S41, acquiring a history carrier roller vibration signal judged to be abnormal-free The corresponding tensor is obtained through the processing of the data segmentation module and the data preprocessing module And (3) with ; S42, tensor And (3) with Feeding into a dual-stream feature fusion-based antagonistic self-encoder model; S43, training the double-flow characteristic fusion countermeasure self-encoder model through an alternate optimization strategy until the model total loss function Convergence, model total loss function Expressed as: ; Wherein, the And As the weight coefficient of the light-emitting diode, In order to reconstruct the loss of the device, In order to combat the loss of this, Is a loss of KL divergence.
5. The intelligent belt conveyor idler monitoring system of claim 4, wherein the dual-stream feature fusion countermeasure self-encoder model comprises: Double-stream encoder sub-module comprising two independent encoders And (3) with Two tensors for processing input respectively Each outputting a latent variable And ; Attention fusion submodule for opposite And Weighting and fusing to generate a combined latent variable : ; Wherein, the And Is the attention weight; dual-branch decoder submodule based on Reconstructing two output tensors The decoder adopts a weight sharing mechanism; A generator submodule for sampling from normal distribution to generate false latent variable With combined latent variables Inputting the two models together into a discriminator to form countermeasure training, and enhancing the capacity of the model for coding the hidden space; A discriminator submodule for judging whether the combined latent variable is from the real prior distribution 。
6. The intelligent monitoring system of belt conveyor idlers according to claim 5, wherein said reconstruction loss is expressed as: ; Wherein, the To calculate And (3) with An average value in a time direction after the Manhattan distance sequence in a frequency axis direction; the countering loss is expressed as: ; Wherein, the And (3) with Representing the output probability of the arbiter for the real sample and the output probability of the arbiter for the generated sample respectively, Representing the log-likelihood of the real sample, Representing the log-likelihood of the generated samples; the KL divergence loss is expressed as: ; Wherein, the In order to calculate the KL distances of the two distributions, For the posterior distribution of the encoder output, To follow the prior distribution of the normal distribution.
7. The intelligent monitoring system of belt conveyor idlers according to claim 5, wherein said alternate optimization strategy is processed as follows: S431, fixed generator Distinguishing device Training the encoder and decoder, forward propagating, calculating reconstruction loss And KL divergence loss Back propagation, updating encoder and decoder parameters; S432, fixed encoder and decoder, training generator and discriminator, calculating the countering loss of the discriminator Counter-propagating, updating generator and arbiter parameters; s433, repeating steps S431 and S432 to the model total loss function Convergence, each loss function convergence condition is as follows: the reconstruction loss convergence condition is that the reconstruction loss The curve gradually drops and tends to be stationary; the loss-countermeasure convergence condition is a discriminator The true and the generated samples cannot be distinguished, the output of the discriminator is completely random, and effective information cannot be provided; The convergence condition of the KL divergence loss is that the KL divergence loss The curve gradually drops and tends to be stationary, and the generator output distribution approaches the true sample distribution.
8. The intelligent monitoring system of a belt conveyor idler of claim 6, wherein the score evaluation module is configured to compare the sequence vibration signal to the self-encoder model by loading a trained dual-stream feature fusion countermeasure According to its correspondence with And And an encoder And (3) with Reconstructed signal And (3) with Calculating a reconstruction loss as a reconstruction score The formula is: + 。
9. the intelligent monitoring system of a belt conveyor idler of claim 1, wherein the anomaly labeling and diagnostic module processes as follows: S61, calculating normal signal reconstruction fraction, namely, vibrating signals The reconstruction score is obtained through the processing of the data segmentation module, the data preprocessing module and the score evaluation module ; S62, drawing the EWMA, namely solving an EWMA control chart corresponding to each position And upper and lower thresholds of EWMA 、 ; S63, calculating the reconstruction fraction of the signal to be detected, namely inputting a carrier roller vibration signal to be diagnosed The reconstruction score is obtained through the processing of the data segmentation module, the data preprocessing module and the score evaluation module ; S64, labeling abnormal scores, namely reconstructing the scores Sending the data into EWMA according to the position, marking the state, when Falls into Recording the interval as normal state, recording as abnormal state when the interval does not fall into the interval, obtaining the reconstruction score Corresponding state distribution ; S65, outputting diagnosis information, namely distributing the state Along the time axis, record Monitoring cycle number duty ratio of position occurrence abnormality Setting normal and abnormal threshold values as 、 When (when) When the corresponding position diagnosis information is recorded as fault abnormality Recording the corresponding position diagnosis information as normal when When the corresponding position diagnosis information is recorded as alarm, all position diagnosis information is counted And sending the data to a state feedback module.
10. The intelligent monitoring system of a belt conveyor idler of claim 9, wherein the status feedback module is configured to diagnose information based on location Checking by monitoring personnel to the abnormal position of the fault, if judging that the abnormal carrier roller is positioned at the position, outputting a vibration signal corresponding to the position If it is judged that there is no abnormal carrier roller at the position, waiting for the next position diagnosis information If the position is still judged to be abnormal and the site check is normal, performing erroneous judgment feedback, namely adding the carrier roller vibration signal in the latest period of time And (3) retraining the dual-stream feature fusion countermeasure self-encoder model until the model total loss function converges.

Description

Belt conveyor idler intelligent monitoring system based on double-flow characteristic fusion countermeasure self-encoder Technical Field The invention relates to the technical field of belt conveyor roller fault monitoring, in particular to an intelligent belt conveyor roller monitoring system based on a double-flow characteristic fusion countermeasure self-encoder. Background In recent years, the market demand of belt conveyors has been expanding due to the green environmental advantages. The carrier roller is used as a key bearing part to bear more than 70% of resistance, and monitoring is important to guaranteeing the safety of the unit. However, the belt conveyor has long length and a large number of carrier rollers, and the existing monitoring means such as manual inspection, ultrasonic detection, high-density electric sensors and the like have obvious limitations in the aspects of cost, instantaneity, environmental adaptability and the like. The distributed optical fiber sensing technology DAS has the natural advantages of convenient deployment, no networking power supply and the like on carrier roller monitoring, but has obvious limitations in the application of carrier roller fault monitoring of the belt conveyor. The Chinese patent publication No. CN119911628B discloses a fault diagnosis method of a belt conveyor based on sound signals, vibration data are compared with a preset threshold range, and when the vibration data are not in the threshold range, the abnormality of a corresponding carrier roller is indicated. In addition, the normal carrier roller can cause different frequency characteristics due to different working conditions of working conditions caused by different positions, so that the traditional threshold judgment method based on characteristic extraction is invalid. The Chinese patent application with publication number of CN120440541A discloses a distributed optical fiber monitoring system for carrier roller faults of a belt conveyor, a dual-light source DAS system is used for inhibiting polarization fading noise, and then a CNN-LSTM fusion model is used for fault classification after self-adaptive filtering and multi-modal feature fusion. However, in practical situations, it is difficult to accumulate all types of faults, severity of each fault and voiceprints under various complex working conditions (environment and load), so that the fault identification capability of the scheme is limited by fault sample labels. In addition, coherent fading noise generated by uneven refractive index distribution of the optical fiber can also cause deterioration of signal-to-noise ratio of vibration signals, thereby interfering with carrier roller fault identification. In summary, although the existing DAS can conveniently collect all-line carrier roller signals by using one optical fiber, the problem of noise interference of the DAS system still needs to be solved, and carrier roller fault signal feature mining and abnormal monitoring are carried out to realize carrier roller fault monitoring. Aiming at the technical bottleneck, a new solution is needed to be developed, frequency characteristic decoupling and characterization are carried out on all-line carrier roller vibration signals acquired by the DAS, so that carrier roller working condition information is effectively mined, the anti-interference capability of carrier roller monitoring on fading noise of a DAS system and external environment noise is further improved, and the practical application value of the DAS in carrier roller fault monitoring of a belt conveyor is fully exerted. Disclosure of Invention The intelligent belt conveyor idler monitoring system based on the double-flow feature fusion anti-self-encoder solves the problem that a high-density electric sensor is difficult to monitor on a belt conveyor network by utilizing a distributed optical fiber sensing technology, accurately captures frequency constitution and change rules of faults by utilizing a FRE and FOIF feature extraction algorithm according to all-line vibration signals acquired by a DAS, utilizes a double-flow feature fusion anti-self-encoder strategy, builds an intelligent diagnosis model with the capability of identifying negative samples under the condition that only a large number of positive samples are used for training, judges a fruiting long period statistical strategy based on the model, improves the anti-interference capability of an idler monitoring system, utilizes the characteristic that an exponential weighted moving average control diagram (EWMA) is more sensitive to small drift change, discovers potential faults earlier, and finally realizes DAS-AI collaborative idler monitoring. The intelligent monitoring system for the belt conveyor idler roller based on the double-flow feature fusion countermeasure self-encoder comprises a data acquisition module, a data segmentation module, a data preprocessing module, a model training module, a score evaluati