CN-121977844-A - Bearing fault diagnosis method and system based on simulated physical neural network

CN121977844ACN 121977844 ACN121977844 ACN 121977844ACN-121977844-A

Abstract

The invention discloses a bearing fault diagnosis method and system based on an analog physical neural network, which comprises the steps of S1, preprocessing acquired vibration signals to obtain signals to be diagnosed, S2, executing a hybrid optimization algorithm to obtain optimal passband parameters and optimal weight parameters, S3, configuring a feature extraction module according to the optimal passband parameters to write the optimal weight parameters into a classification module, S4, inputting the signals to be diagnosed into the feature extraction module to obtain analog feature vectors, S5, inputting the analog feature vectors into the classification module to output classification voltage signals, S6, determining fault types and outputting results. The invention calculates the ratio of the inter-class dispersion and intra-class convergence of the feature vector by adopting a hybrid optimization algorithm, realizes the optimal adaptation of the feature extraction and classification tasks, improves the diagnosis precision and generalization capability, and realizes the real-time response of the micro watt power consumption and the microsecond level by constructing a full analog domain signal processing architecture without analog-to-digital conversion and digital calculation.

Inventors

LU SILIANG
Tian ao
HU ZHIYONG
WANG XIAOXIAN
SONG JUNCAI
GUI YONG
XU ANNING
CHEN WENYUE
LI XIANG

Assignees

安徽大学

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (9)

1. A bearing fault diagnosis method based on a simulated physical neural network is characterized by comprising the following steps: s1, collecting vibration signals and preprocessing the vibration signals to obtain signals to be diagnosed; S2, executing a hybrid optimization algorithm, performing cooperative optimization on passband parameters of a feature extraction module and weight parameters of a classification module serving as variables to be optimized to obtain optimal passband parameters and optimal weight parameters, wherein the feature extraction module consists of a plurality of analog bandpass filter channels which are arranged in parallel, and the classification module consists of an analog neural network circuit; S3, configuring the feature extraction module according to the optimal passband parameters, and writing the optimal weight parameters into the classification module; S4, inputting the signal to be diagnosed into a configured feature extraction module, and extracting to obtain a simulation feature vector; s5, inputting the analog feature vector into a classification module after writing weights, executing on-line gradient descent training by a closed feedback loop, and outputting classification voltage signals; The analog neural network circuit consists of an analog multiplier array, a weighted aggregation circuit and a lossy integrator; S5, inputting the analog feature vector and the weight parameter in the classification module written with the weight into an analog multiplier array in parallel, obtaining a weighted result through analog multiplication operation, inputting the weighted result into a weighted aggregation circuit, completing calculation of neuron weighted sum through a summing topology constructed by an operational amplifier, outputting a predicted voltage signal, carrying out differential processing on the predicted voltage signal and a target voltage signal, calculating to obtain an error signal, carrying out multiplication processing on the error signal and the analog feature vector to obtain a weight gradient, inputting the weight gradient into a lossy integrator, updating the weight parameter stored in the classification module through integral operation, forming a closed loop feedback loop and outputting the classified voltage signal; s6, determining the fault type of the bearing according to the classified voltage signals and outputting a diagnosis result.
2. The method for diagnosing the bearing fault based on the simulated physical neural network is characterized in that S1 specifically comprises the steps of performing high-pass filtering on an acquired vibration signal, removing a direct-current component to obtain an alternating-current vibration signal, performing amplitude normalization processing on the alternating-current vibration signal to obtain a standardized vibration signal, and dividing the standardized vibration signal according to a fixed time window to obtain a signal to be diagnosed.
3. A bearing fault diagnosis method based on a simulated physical neural network is characterized by comprising the steps of S2, initializing a population based on a bearing fault characteristic frequency distribution range, introducing a nonlinear convergence factor to adjust a search step length, fusing an individual history optimal guiding mechanism and a global optimal guiding mechanism in position updating, iteratively updating passband parameters and weight parameters of the individuals, S23, calculating fitness function values corresponding to the individuals in the current population, wherein the fitness function is defined as a ratio of inter-class dispersion to intra-class dispersion of simulated characteristic vectors obtained after different fault class samples are processed by a characteristic extraction module, S24, iteratively executing S22-S23 until the fitness function value converges, and outputting passband parameters and weight parameters corresponding to the individuals with the largest fitness function values as optimal passband parameters and optimal weight parameters respectively.
4. The bearing fault diagnosis method based on the simulated physical neural network according to claim 3 is characterized in that the fitness function value convergence is specifically that the fitness function value of a globally optimal individual in each iteration is recorded, the variation amplitude of the fitness function value of the globally optimal individual in a plurality of iterations is calculated, and when the variation amplitude is smaller than a preset threshold value, the convergence of the fitness function value is judged.
5. The method for diagnosing bearing faults based on the simulated physical neural network according to claim 1 is characterized in that the configuration of the feature extraction module according to the optimal passband parameters is specifically that the optimal passband parameters are converted into digital control words and sent to programmable analog filters in the feature extraction module so as to set passband of each filter channel; The writing of the optimal weight parameters into the classification module is specifically that the optimal weight parameters are written into a lossy integrator in the classification module in an analog voltage mode, so that the weight parameters are stored at two ends of a feedback capacitor of the lossy integrator in a charge mode.
6. The method for diagnosing the bearing fault based on the simulated physical neural network according to claim 1, wherein S4 is specifically characterized by comprising the steps of inputting the signal to be diagnosed into a simulated band-pass filter channel in a configured feature extraction module, outputting a filtering signal in a corresponding frequency band range by each channel based on the optimal passband parameters, calculating root mean square values of the filtering signals output by each channel respectively to obtain a plurality of simulated feature values, and combining the simulated feature values into simulated feature vectors.
7. The bearing fault diagnosis method based on the simulated physical neural network, which is disclosed in claim 1, is characterized in that the classification module is provided with a feedforward aggregation interface and a gradient return interface, wherein the feedforward aggregation interface is used for receiving an intermediate aggregation voltage signal output by a previous stage circuit board and summing the weighting result output by a weighting aggregation circuit in the classification module to realize expansion of characteristic dimensions, the gradient return interface is used for receiving a gradient component signal output by a next stage circuit board and coupling the gradient component signal in the classification module to realize expansion of classification quantity, and a plurality of physical neural network units are connected in a cascading manner through the feedforward aggregation interface and the gradient return interface to form an expandable modularized simulated neural network architecture.
8. The method for diagnosing the bearing faults based on the simulated physical neural network is characterized by comprising the following steps of S6, calculating error energy between the classified voltage signals and target voltages corresponding to preset faults to obtain a plurality of error energy values, taking a fault type corresponding to the smallest error energy value in the error energy values as an initial diagnosis result, calculating the ratio of the smallest error energy value to the next smallest error energy value to obtain a relative confidence index, and taking the initial diagnosis result as a final fault type and outputting the final fault type when the relative confidence index is larger than a preset threshold.
9. A bearing fault diagnosis system based on a simulated physical neural network for implementing the bearing fault diagnosis method based on the simulated physical neural network according to any one of claims 1-8, characterized by comprising a feature extraction module, a detection module and a detection module, wherein the feature extraction module is composed of a plurality of simulated bandpass filter channels arranged in parallel and is used for receiving signals to be diagnosed and extracting simulated feature vectors; The classification module is used for receiving the analog feature vector and executing on-line gradient descent training and outputting a classification voltage signal; and the result output module is used for determining the fault type of the bearing according to the classified voltage signals and outputting the fault type.

Description

Bearing fault diagnosis method and system based on simulated physical neural network Technical Field The invention relates to the technical field of bearing fault diagnosis, in particular to a bearing fault diagnosis method and system based on a simulated physical neural network. Background With the rapid development of industrial automation and intelligent manufacturing, state monitoring and fault diagnosis technologies of rotating machinery equipment are increasingly paid attention to. The rolling bearing is used as a core component of the rotary machine, and the running state of the rolling bearing directly influences the safety and reliability of the whole machine. At present, a fault diagnosis method based on vibration signal analysis becomes a mainstream technical means, and a framework of a sensor, a data acquisition card and a digital processor is widely adopted. The architecture firstly converts an analog vibration signal into a digital signal through an analog-to-digital converter, and then a digital signal processor, a field programmable gate array or a computer is used for running a machine learning or deep learning algorithm to complete fault identification. However, the above-described digital processing approach based on von neumann architecture has many limitations in industrial field applications for edge-oriented computing. Firstly, bearing fault characteristics often contain high-frequency components, and according to the nyquist sampling theorem, extremely high sampling rate is needed to effectively retain fault information, so that not only are severe requirements on bandwidth and precision of an analog-to-digital converter set, but also massive original data can be generated, and huge pressure is brought to data transmission bandwidth and storage space. Secondly, the digital calculation is discrete and serial in nature, data are required to be stored and processed, a complex neural network model depends on a large number of multiply-add operation instruction periods, unavoidable calculation delay is caused, and the requirement of high-speed rotating equipment on millisecond-level or even microsecond-level real-time response of sudden faults is difficult to meet. Again, to maintain high frequency sampling and complex floating point operations, the digital processor needs a high frequency clock drive, the power consumption is usually in the watt level, and for wireless sensor nodes relying on battery-powered or energy harvesting techniques, this power consumption level severely limits the deployment life of the monitoring node and the popularization and application of the edge devices. In order to overcome the power consumption and delay bottleneck of digital computation, part of researches try to migrate signal processing and computation tasks to an analog domain, and directly complete computation by utilizing the physical properties of voltage and current, and theoretically, the method has the advantages of extremely low power consumption and infinite time resolution. For example, patent CN120508813B discloses a method and a system for diagnosing bearing faults based on a multi-channel analog filtering characteristic network, in which a plurality of parallel analog bandpass filter channels are used to extract root mean square values of vibration signals as characteristics, fault classification is realized by an analog neural network classifier, and meanwhile, a particle swarm algorithm is used to jointly optimize the passband parameters of the filter and the classifier weight. The scheme realizes signal processing and classification reasoning of an analog domain to a certain extent, and has the characteristics of low power consumption and high instantaneity. However, the existing analog circuit scheme still has obvious defects that firstly, the scheme disclosed by CN120508813B adopts an offline training and online reasoning working mode, namely, filter parameters and classifier weights are required to be configured to hardware after the optimization is finished in advance at a computer end, online self-adaptive updating cannot be realized according to working condition changes or performance degradation in the running process of equipment, life learning capability is lacking, secondly, the classifier only carries out forward reasoning, a hardware loop with error back propagation is not constructed, online gradient descent training of the weights cannot be finished in an analog domain, thirdly, the hardware scale of the analog neural network classifier of the scheme is fixed, a modularized expansion interface is lacked, the requirements of different feature dimensions or classification tasks are difficult to flexibly adapt, fourthly, the feature extraction module and the classification module in the existing scheme are jointly optimized, but once the optimized parameters are configured, the parameters are solidified, and signal distribution change caused by equipment aging or environmen