CN-121994480-A - Mining shovel robot gear box cross-working condition fault diagnosis method and system

CN121994480ACN 121994480 ACN121994480 ACN 121994480ACN-121994480-A

Abstract

The invention provides a method and a system for diagnosing a cross-working condition fault of a mining shovel robot gearbox, wherein the method comprises the following steps of S1, collecting vibration signals of the mining shovel robot lifting mechanism gearbox under different working conditions; the method comprises the steps of S2, constructing a diagnosis network and carrying out self-adaptive feature extraction on vibration signals, wherein the diagnosis network comprises a feature extractor and a fault classifier of a one-dimensional deformable convolution network, S3, training the diagnosis network by adopting a double-stage domain self-adaptive strategy fused with triplet loss, and outputting a diagnosis model, and S4, carrying out fault diagnosis on the vibration signals under a target working condition according to the diagnosis model, and outputting a diagnosis result. According to the invention, a dual-stage domain self-adaptive strategy is adopted, the first stage performs feature optimization on the source domain, the second stage performs data enhancement and domain invariant feature learning through mixing the source domain and the target domain labeling samples, so that the domain deviation problem caused by working condition change is effectively relieved, and the cross-working condition high-efficiency fault diagnosis is realized.

Inventors

QIN CHENGJIN
CAI RUIJIE
LIU CHENGLIANG
TAO JIANFENG
XIA PENGCHENG
JIANG JIACHENG

Assignees

上海交通大学

Dates

Publication Date: 20260508
Application Date: 20260120

Claims (10)

1. A mining shovel robot gear box cross-working condition fault diagnosis method is characterized by comprising the following steps: step S1, collecting vibration signals of a lifting mechanism gearbox of a mining shovel robot under different working conditions; S2, constructing a diagnosis network and carrying out self-adaptive feature extraction on the vibration signals; The diagnosis network comprises a feature extractor and a fault classifier of a one-dimensional deformable convolution network; Step S3, training the diagnosis network by adopting a double-stage domain self-adaptive strategy of fusion triplet loss based on the extracted characteristics, and outputting a diagnosis model; and S4, performing fault diagnosis on the vibration signal under the target working condition according to the diagnosis model and outputting a diagnosis result.
2. The mining shovel robot gear box cross-working condition fault diagnosis method according to claim 1 is characterized in that a one-dimensional deformable convolution network is adopted as a core module of a feature extractor, and the feature extraction of the mining shovel robot lifting mechanism gear box vibration signals under different working conditions is completed by dynamically adjusting the sampling position and the weight of a convolution kernel and self-adaptively capturing local time sequence features in the vibration signals.
3. The mining shovel robot gearbox cross-condition fault diagnosis method according to claim 1, wherein the double-stage domain adaptive strategy for fusing triplet losses comprises: Training a diagnosis model on source domain data, performing active learning by using the diagnosis model, selecting a sample with the lowest or highest prediction confidence from target domain samples for expert labeling, and obtaining a label of the part of samples; and learning domain invariant features by using a triplet loss guide diagnosis model to finish domain adaptation fault diagnosis of the cross-working condition.
4. The mining shovel robot gearbox cross-working-condition fault diagnosis method according to claim 3, wherein in the double-stage-domain self-adaptive strategy for fusing triplet loss: The loss function used in the first stage is: Wherein the method comprises the steps of A source domain sample is represented and, Representing the cross-entropy loss, Representing the use of cross entropy loss on source domain samples, Representing the loss of a triplet, Indicating the use of triplet loss on source domain samples, Representing a loss weight; The loss function used in the second stage is: Wherein, the Representing a target domain sample, which, in an initial state, The samples in the test are unlabeled; Representing a set of labeled target domain samples obtained by expert labeling after an active learning strategy is adopted for unlabeled target domain samples; Representation pair And After the union operation, the data enhancement operation is performed The specific operation mode of (a) is as follows Shown therein, wherein Representing the random selection of two different samples from the operated set, the pair Obtaining a new sample by linear interpolation Combining all generated new sample sets and original sets into a union operation, and enhancing data Namely, the completion is completed; Representing samples which are not marked after the active learning strategy is adopted, adopting pseudo-label processing, and obtaining corresponding cross entropy loss through calculating the pseudo-labels; All represent loss weights.
5. The mining shovel robot gear box cross-working condition fault diagnosis method according to claim 1, wherein an nn packet under Pytorch frames is used for constructing and training a mining shovel robot lifting mechanism gear box fault diagnosis model, and model performance is evaluated according to diagnosis results and actual comparison.
6. A mining shovel robot gear box cross-working condition fault diagnosis system is characterized by comprising: The module M1 is used for collecting vibration signals of a lifting mechanism gearbox of the mining shovel robot under different working conditions; A module M2, constructing a diagnosis network and carrying out self-adaptive feature extraction on the vibration signals; The diagnosis network comprises a feature extractor and a fault classifier of a one-dimensional deformable convolution network; A module M3, based on the extracted characteristics, training the diagnosis network by adopting a double-stage domain self-adaptive strategy of fusion triplet loss, and outputting a diagnosis model; And the module M4 is used for carrying out fault diagnosis on the vibration signal under the target working condition according to the diagnosis model and outputting a diagnosis result.
7. The mining shovel robot gear box cross-working-condition fault diagnosis system according to claim 6 is characterized in that a one-dimensional deformable convolution network is adopted as a core module of a feature extractor, and the feature extraction of the mining shovel robot lifting mechanism gear box vibration signals under different working conditions is completed by dynamically adjusting the sampling position and the weight of a convolution kernel and self-adaptively capturing local time sequence features in the vibration signals.
8. The mining shovel robot gearbox cross-condition fault diagnosis system according to claim 6, wherein the dual-stage domain adaptive strategy of fusion triplet loss comprises: Training a diagnosis model on source domain data, performing active learning by using the diagnosis model, selecting a sample with the lowest or highest prediction confidence from target domain samples for expert labeling, and obtaining a label of the part of samples; and learning domain invariant features by using a triplet loss guide diagnosis model to finish domain adaptation fault diagnosis of the cross-working condition.
9. The mining shovel robot gearbox cross-condition fault diagnosis system according to claim 8, wherein in the double-stage domain adaptive strategy of fusion triplet loss: The loss function used in the first stage is: Wherein the method comprises the steps of A source domain sample is represented and, Representing the cross-entropy loss, Representing the use of cross entropy loss on source domain samples, Representing the loss of a triplet, Indicating the use of triplet loss on source domain samples, Representing a loss weight; The loss function used in the second stage is: Wherein, the Representing a target domain sample, which, in an initial state, The samples in the test are unlabeled; Representing a set of labeled target domain samples obtained by expert labeling after an active learning strategy is adopted for unlabeled target domain samples; Representation pair And After the union operation, the data enhancement operation is performed The specific operation mode of (a) is as follows Shown therein, wherein Representing the random selection of two different samples from the operated set, the pair Obtaining a new sample by linear interpolation Combining all generated new sample sets and original sets into a union operation, and enhancing data Namely, the completion is completed; Representing samples which are not marked after the active learning strategy is adopted, adopting pseudo-label processing, and obtaining corresponding cross entropy loss through calculating the pseudo-labels; All represent loss weights.
10. The mining shovel robot gear box cross-working-condition fault diagnosis system according to claim 6, wherein an nn package under Pytorch frames is used for constructing and training a mining shovel robot lifting mechanism gear box fault diagnosis model, and model performance is evaluated according to diagnosis results and actual comparison.

Description

Mining shovel robot gear box cross-working condition fault diagnosis method and system Technical Field The invention relates to the field of fault diagnosis, in particular to a mining shovel robot gearbox cross-working-condition fault diagnosis method and system, and especially relates to a mining shovel robot lifting mechanism gearbox cross-working-condition fault diagnosis method and system. Background In mining operation, the shovel robot is used as key equipment for large-scale surface mining, a lifting mechanism of the shovel robot bears the tasks of digging and loading heavy materials, the working strength is high, the working condition is complex, and the stability requirement of a transmission system is extremely high. The gear box in the lifting mechanism of the electric shovel robot is used as a core component for power transmission, is in severe environments such as high load, frequent start and stop, dust, moisture and the like for a long time, and is extremely easy to cause faults such as gear abrasion, pitting, gear breakage, bearing damage and the like, so that the transmission efficiency is reduced, the vibration is aggravated, and equipment shutdown and even safety accidents can be caused when the equipment is serious. The gear box of the lifting mechanism of the mining shovel robot often works under complex and changeable working conditions such as variable rotation speed, variable load, variable rotation speed, high noise and the like. In this case, the conventional fault diagnosis method such as vibration analysis, acoustic detection, etc. is susceptible to interference of environmental noise, and the diagnosis accuracy is limited. Therefore, how to realize the high-precision diagnosis of the cross-working condition of the lifting mechanism gear box fault of the shovel robot becomes a key challenge for guaranteeing the safe operation of equipment and promoting the development of intelligent operation and maintenance of mines. Patent document CN115409110a discloses a fault diagnosis method of a cross-working condition gearbox, which comprises the steps of obtaining a large number of labeled source domain fault data and a few unlabeled target domain fault data, carrying out wavelet packet transformation on the collected data, obtaining wavelet packet coefficients, using the wavelet packet coefficients as model input, constructing a diagnosis model, including a feature extractor, a domain discriminator and a classifier, respectively extracting features of the source domain fault data and the target domain fault data by the feature extractor, carrying out countermeasure training on the source domain fault data through the domain discriminator and the target domain fault data after weighting treatment, calculating countermeasure loss, calculating classification loss through classification results of the classifier, optimizing parameters of the diagnosis model through the anti-loss and the classification loss, carrying out wavelet packet transformation on the data to be detected, inputting the wavelet coefficients into the model, and outputting fault diagnosis results by the model. The core of the patent document relies on wavelet packet transformation to extract fixed frequency band characteristics and realizes domain alignment through countermeasure training, and the method has two key limitations that firstly, the characteristic extraction of the wavelet packet transformation relies on manual selection of decomposition layer number and frequency band division, and is difficult to adapt to non-stationarity and frequency offset of vibration signals under complex working conditions of mines, and secondly, the countermeasure training is easy to cause mode collapse or unstable training, and in-class compactness and inter-class separability of a characteristic space cannot be ensured, and the interpretability and diagnosis robustness of a model are affected. In contrast, the invention adopts a one-dimensional deformable convolution network, can adaptively and dynamically adjust the convolution kernel sampling position and weight according to the local waveform characteristics of the input signal, realizes the self-adaptive characteristic extraction of the vibration signal, and is more suitable for the non-stationary characteristics of the signal of the mining equipment under the working conditions of impact load, frequent start and stop and the like. Meanwhile, the invention actively builds a compact intra-class structure and a clear inter-class boundary in the first stage through the triple loss, thereby avoiding the problem of feature confusion possibly occurring in the countermeasure training and providing a structured feature space foundation for subsequent active learning. The patent document CN119719862A discloses a method and a system for diagnosing a cross-working condition fault of a planetary gear box driven by digital-analog fusion, and the method comprises the steps of collecting vibr