CN-121980347-A - Multi-mode data volume imbalance self-adaptive fault diagnosis method

CN121980347ACN 121980347 ACN121980347 ACN 121980347ACN-121980347-A

Abstract

The invention relates to the technical field of intelligent fault diagnosis, in particular to a self-adaptive fault diagnosis method for unbalanced multi-mode data volume. The technical scheme includes that multi-mode data of industrial equipment are synchronously collected, time alignment and standardization processing are carried out on the multi-mode data to obtain standardized signals, and different time sequence image coding methods are respectively adopted for converting the standardized signals in different modes. The method can dynamically balance the utilization of different data volume modes, effectively overcomes information use deviation caused by sampling frequency difference, automatically enhances the perceptibility of fault symptoms in low data volume but key modes, reduces the overfitting of high data volume modes, simultaneously reduces the dependence of a model on manual setting weight, realizes more fair and more robust multi-mode information fusion under the condition of no manual intervention, and improves diagnosis accuracy and generalization performance.

Inventors

LIU SHAOQING
Wang Mengxu
LIU ZIJIE
Guo Heru
XIAO BINGJIA
JI ZHENSHAN

Assignees

合肥综合性国家科学中心能源研究院(安徽省能源实验室)

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (9)

1. The self-adaptive fault diagnosis method for the unbalance of the multi-mode data volume is characterized by comprising the following steps of: s1, synchronously acquiring multi-mode data of industrial equipment, and performing time alignment and standardization processing on the multi-mode data to obtain a standardized signal; S2, respectively converting the standardized signals of different modes by adopting different time sequence image coding methods to generate corresponding single-channel images, and splicing all the single-channel images along the channel dimension to form a multi-channel fusion image; S3, carrying out feature extraction on the multichannel fusion image by using a convolutional neural network to obtain a feature vector for representing the joint state of the equipment; S4, inputting the feature vector to an agent in a deep reinforcement learning framework, and outputting fault diagnosis actions by the agent; the environment module calculates basic rewards according to the fault diagnosis actions and the real fault labels, calculates self-adaptive balanced rewards by combining prior weights of all mode data and real-time mode utilization degrees, and comprehensively generates total rewards to be fed back to the intelligent agent; S5, updating the strategy network parameters of the intelligent agent through a strategy gradient algorithm by utilizing the total rewards to form a feedback learning loop, and performing fault diagnosis on newly input multi-mode data through a trained strategy network after training is completed.
2. The adaptive fault diagnosis method for multi-modal data volume imbalance according to claim 1, wherein in step S1, the multi-modal data includes at least a vibration signal, a temperature signal, and a current signal; the time alignment is to take the time sequence of the modal signal with the highest sampling rate as a reference to interpolate and resample other low-frequency modal signals; the normalization is to independently perform Min-Max normalization according to modes and map signal values to [ -1,1] intervals.
3. The adaptive fault diagnosis method for multi-modal data volume imbalance according to claim 2, wherein the different time-series image encoding methods specifically include: Generating a first single-channel image by adopting a gram angle and a field code on the vibration signal; the temperature signal is encoded by adopting a gram angle difference field to generate a second single-channel image; the current signal is coded by adopting a Markov transfer field to generate a third single-channel image; and splicing the first single-channel image, the second single-channel image and the third single-channel image into a three-channel RGB fusion image.
4. The adaptive fault diagnosis method for multi-modal data volume imbalance according to claim 1, wherein in step S3, the convolutional neural network comprises a multi-scale convolutional module and a channel attention module; the multi-scale convolution module is used for extracting spatial features of different scales in parallel; the channel attention module is used for adaptively weighting the importance of each channel characteristic.
5. The adaptive fault diagnosis method for multi-modal data volume imbalance according to claim 1, wherein in step S4, the total prize The calculation formula of (2) is as follows: ; Wherein, the The false judgment rate at the time t is 0 or 1; Is a balance coefficient; Is the first The prior weight of each mode is calculated based on the reciprocal of the number of training samples of each mode; Is the first The real-time mode utilization degree of each mode is calculated based on the gradient amplitude ratio of the classification loss to the mode characteristics.
6. The adaptive fault diagnosis method for multi-modal data volume imbalance as claimed in claim 5, wherein said a priori weights The calculation formula of (2) is as follows: ; Wherein, the Is the first The number of training samples for each modality, Is the first The number of valid training samples for each modality.
7. The adaptive fault diagnosis method for multi-modal data volume imbalance as claimed in claim 5, wherein said real-time modal utilization degree The actual degree of dependence of the quantization model on the mth modality information in the current decision is a scalar The larger the value, the more dependent the model is on the mode to make decisions, and the calculation formula is: ; Wherein, the Is the first The feature vectors of the individual modalities are used, In order to cross-entropy loss function, Indicating loss For a pair of Is used for the gradient of (a), The L2 norm is represented by the number, Is the total number of modalities.
8. The method for adaptively diagnosing a fault with unbalanced multimodal data amount according to claim 1, wherein in the step S5, the policy gradient algorithm is a near-end policy optimization algorithm, and the feedback learning loop updates policy network parameters by using a trajectory generated by interaction of an agent with an environment and a dominant function calculated by generalized dominant estimation.
9. The method for adaptive fault diagnosis with unbalanced multimodal data amount according to claim 1, wherein in step S5, the training completion flag is that the verification set diagnosis accuracy is continuously preset for a round without improvement or a maximum training round is reached; and in the diagnosis stage, the real-time multi-mode data is processed in the steps S1 to S3 to obtain a feature vector, the feature vector is input into a trained strategy network, and the fault class with the highest probability is output as a diagnosis result.

Description

Multi-mode data volume imbalance self-adaptive fault diagnosis method Technical Field The invention relates to the technical field of intelligent fault diagnosis, in particular to a self-adaptive fault diagnosis method for unbalanced multi-mode data volume. Background Along with the rapid development of intelligent manufacturing and industrial Internet of things technology, on-line state monitoring and intelligent fault diagnosis of key industrial equipment are realized, and the method has become a core means for guaranteeing production safety and improving operation and maintenance efficiency. In a practical industrial scenario, single sensor information is often insufficient to fully, robustly reflect complex operating conditions and early signs of failure of the device. Therefore, the multimode sensing data of various physical quantities such as vibration, temperature, current and the like are fused, and the accuracy and the reliability of diagnosis are improved through information complementation, so that the multimode sensing data has become an important research direction in the field. However, multimodal data fusion presents a common and serious challenge in that the amount of data between modalities is severely unbalanced. This is mainly due to the fact that the natural sampling frequencies of the sensors of different physical principles are different by orders of magnitude. For example, vibration acceleration sensors typically acquire at frequencies of thousands of Hz or even higher, which can produce thousands of data points per second, whereas sensors of temperature, pressure, etc. typically sample at a lower frequency, possibly only a few Hz or lower, resulting in a significant difference in the effective sample numbers of different modalities within the same time window. This imbalance is not random noise, but rather a systematic deviation. In the existing multi-mode fault diagnosis method based on deep learning, an end-to-end characteristic fusion architecture is mostly adopted, such as a multi-channel Convolutional Neural Network (CNN) or a fusion network based on an attention mechanism. These methods typically splice or weight aggregate the multimodal data at the data entry layer or feature layer. However, its optimization objective (such as cross entropy loss) is often for overall classification accuracy, and does not explicitly account for the large differences in the number of samples between modalities. During training, massive amounts of high frequency modal data (e.g., vibrations) dominate the gradient update direction, so that the model tends to learn and rely on characteristic modes in these modes, while low frequency modes (e.g., temperature ramp indicating overheat trend) where samples are sparse but critical fault information may be contained are ignored. This essentially results in "overfitting" of the model to the high data volume mode and "under learning" to the low data volume mode, and during the test phase, diagnostic performance is significantly degraded once the high frequency mode signal is disturbed or a particular fault is more pronounced in the low frequency mode. To alleviate the data imbalance problem, there have been studies attempting to introduce cost-sensitive learning or oversampling of a few modality samples. However, these methods belong to static or heuristic strategies, and cannot be dynamically adjusted according to the real-time learning state in the model training process. In recent years, deep Reinforcement Learning (DRL) is introduced into the field of fault diagnosis, and an intelligent body thereof learns an optimal diagnosis strategy through interactive trial and error with the environment and shows good self-adaptive potential. However, the design of the reward function in the existing DRL diagnostic framework is still limited to whether the final classification is correct (such as sparse +1/-1 rewards), and the single accuracy reward signal cannot sense the difference of the utilization degree of the information of each mode in the decision process of the agent. Therefore, the traditional DRL method is also difficult to automatically balance the contributions of different data volume modes, and real self-adaptive fusion learning cannot be realized under the condition of unbalanced data. In summary, in the multi-mode fault diagnosis of the prior art, the problem of model deviation caused by unbalanced data volume between modes cannot be effectively overcome, so that the robustness, generalization capability and sensitivity to critical but sparse fault symptoms of a diagnosis model are limited. Therefore, there is a need for an intelligent fault diagnosis method capable of dynamically sensing and adaptively compensating for unbalance of modal data volume, and a guided model fairly and fully utilizes all modal information, so as to realize more reliable and more accurate equipment health state assessment and fault early warning in a complex industrial