CN-121990136-A - Fault diagnosis method for mechanical system of water jet propulsion device based on time-frequency prototype network

CN121990136ACN 121990136 ACN121990136 ACN 121990136ACN-121990136-A

Abstract

The invention provides a fault diagnosis method of a mechanical system of a water jet propulsion device based on a time-frequency prototype network, which comprises the steps of carrying out data acquisition and time-frequency conversion through sensor measuring points to obtain a single-channel time-frequency matrix, obtaining a fused RGB time-frequency diagram through RGB mapping rules, carrying out image processing on the fused RGB time-frequency diagram to obtain a standard input sample, carrying out staged training on a built initial model to obtain a diagnosis model, carrying out reasoning diagnosis and verification on the standard input sample through the diagnosis model to obtain a fault diagnosis result, a credibility basis and high-resolution evidence. The invention can realize accurate, efficient and interpretable fault diagnosis, can realize visual monitoring of the fault diagnosis process and case-based reasoning type postattribution of the diagnosis result, and provides a fault diagnosis scheme with high precision and high reliability for intelligent operation and maintenance of ships.

Inventors

GAO TIANYU
CHEN QIGANG
YANG JINGLI
JIANG SHOUDA

Assignees

哈尔滨工业大学

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (8)

1. A mechanical system fault diagnosis method of a water jet propulsion device based on a time-frequency prototype network is characterized by comprising the following steps: Data acquisition and time-frequency conversion are carried out through sensor measuring points arranged on a mechanical system of the water jet propulsion device, a single-channel time-frequency matrix is obtained, and a space coordinate system is defined for physical mapping; converting the single-channel time-frequency matrix into a fused RGB time-frequency diagram by using a defined RGB mapping rule, and then carrying out image processing on the fused RGB time-frequency diagram to obtain a standard input sample; collecting state data sets comprising five health states on the sensor measuring points, and training the constructed initial model in stages by utilizing the state data sets to obtain a diagnosis model; and carrying out reasoning diagnosis on the standard input sample by using the diagnosis model to obtain a fault diagnosis result and a credibility basis, and verifying the fault diagnosis result to obtain high-resolution evidence.
2. The method for diagnosing faults of a mechanical system of a water jet propulsion device based on a time-frequency prototype network as claimed in claim 1, wherein the steps of acquiring a single-channel time-frequency matrix by data acquisition and time-frequency transformation through sensor measuring points arranged on the mechanical system of the water jet propulsion device, defining a space coordinate system for physical mapping include: the mechanical system based on the water jet propulsion device is characterized in that sensor measuring points are respectively arranged on a primary impeller shell, a pump bearing shell, a secondary impeller shell and a main shaft bearing shell, and data acquisition is carried out on the sensor measuring points to obtain multichannel original data; Based on a mechanical system of the water jet propulsion device, a direction parallel to a main shaft central line of the propulsion pump and pointing to water jet is defined as an X axis, a direction of a sensor mounting point hairline and penetrating through a geometric center of the pump body is defined as a Y axis, a tangential direction along a shell is defined as a Z axis, and a space coordinate system is obtained for physical mapping; Based on the multichannel original data, continuous wavelet transformation based on complex Morlet wavelet is utilized, microsecond impact position capturing is carried out in a high-frequency-band automatic narrowing time window through telescopic translation operation, and frequency component refinement is carried out in a low-frequency-band automatic widening time window, so that a single-channel time-frequency matrix is obtained.
3. The method for diagnosing mechanical system failure of water jet propulsion device based on time-frequency prototype network as set forth in claim 2, wherein the converting the single channel time-frequency matrix into a fused RGB time-frequency diagram by using defined RGB mapping rules, and then performing image processing on the fused RGB time-frequency diagram to obtain a standard input sample, comprises: setting a defined Z axis as an R channel, setting a Y axis as a G channel and setting an X axis as a B channel, and completing definition of RGB mapping rules; Based on the single-channel time-frequency matrix, capturing rotation centrifugal force and transverse sweeping characteristics by utilizing the R channel, capturing radial impact and rigidity change of a supporting structure by utilizing the G channel, and capturing axial thrust fluctuation caused by fluid-solid coupling by utilizing the B channel to obtain a fused RGB time-frequency diagram; And based on the fusion RGB time-frequency diagram, performing size unification, and generating a group of multidimensional image tensors containing space and time-frequency characteristics for each acquisition time period to obtain a standard input sample.
4. A method for diagnosing a mechanical system failure of a water jet propulsion system based on a time-frequency prototype network as recited in claim 3, wherein collecting a state data set including five health states at said sensor measurement points, and using said state data set to train the initial model constructed in stages to obtain a diagnostic model, comprising: Combining a depth feature extractor and a designed prototype decision layer to obtain an initial model, mapping the standard input sample into a high-dimensional feature map by using the depth feature extractor, calculating Euclidean distance between an input feature block and each prototype vector by using the prototype decision layer based on the high-dimensional feature map, and completing initial reasoning; Based on the sensor measuring points, collecting state data sets of five health states including a normal state, too small blade tip clearance, too large blade tip clearance, large impeller damage and small impeller damage under different working conditions, and dividing the state data sets into a training set and a testing set according to the proportion of 7:3 after randomly disturbing the state data sets; Designing a mixed loss function comprising a classification accuracy loss function, a clustering loss function and a separation loss function, and training the initial model in stages by utilizing a training set, a testing set and the mixed loss function to obtain a diagnosis model.
5. The method for diagnosing a mechanical system fault of a water jet propulsion system based on a time-frequency prototype network as recited in claim 4, wherein said prototype decision layer is configured to preset a plurality of learnable prototype vectors, and assign a plurality of said learnable prototype vectors to each type of fault, wherein each of said learnable prototype vectors corresponds to a potential image block in said standard input sample.
6. A method for diagnosing a mechanical system failure of a water jet propulsion system based on a time-frequency prototype network as recited in claim 5, wherein the step of training comprises: Freezing the weight of the depth feature extractor, training the prototype decision layer and the full connection layer, initializing prototype vectors by using an Adam optimizer and a preset learning rate, and completing warm-up stage training; Performing joint training on the unfrozen depth feature extractor and the prototype decision layer, and introducing a learning rate attenuation strategy until the mixed loss function converges to complete joint fine tuning stage training; and replacing the leachable prototype vector with the true image block closest to the training set to complete the prototype projection stage training.
7. The method for diagnosing faults in a mechanical system of a water jet propulsion system based on a time-frequency prototype network as claimed in claim 6, wherein the step of performing a reasoning diagnosis on the standard input sample by using the diagnosis model to obtain a fault diagnosis result and a credibility basis, and verifying the fault diagnosis result to obtain high-resolution evidence comprises the steps of: Inputting the standard input sample into the diagnosis model, performing feature mapping, euclidean distance measurement and distance conversion to obtain category conditional probability, and outputting a fault diagnosis result and a credibility basis according to the category conditional probability; And carrying out pixel-level weighting of layering gradients by using layering activation mapping based on the fault diagnosis result and the credibility basis, generating activation graphs of each layer, interpolating and aligning the activation graphs with different resolutions to an input size, fusing and superposing the activation graphs to obtain a multi-scale fused high-resolution thermodynamic diagram, and carrying out physical consistency verification on the high-resolution thermodynamic diagram to obtain high-resolution evidence.
8. A method for diagnosing mechanical system failure of a water jet propulsion system based on a time-frequency prototype network as recited in claim 7, wherein said physical consistency verification includes a fishback frequency verification and an RGB channel coupling verification.

Description

Fault diagnosis method for mechanical system of water jet propulsion device based on time-frequency prototype network Technical Field The invention relates to the technical field of intelligent operation and maintenance and fault diagnosis of ships, in particular to a fault diagnosis method of a mechanical system of a water jet propulsion device based on a time-frequency prototype network. Background The water jet propulsion device is a complex power and control system which relates to multi-physical field coupling, and mainly comprises a control system, a hydraulic system and a mechanical system according to the functional structure and signal characteristics of the water jet propulsion device. The control system comprises a main control circuit, a signal conditioning circuit and a power driving circuit, is responsible for command calculation and signal amplification, consists of a hydraulic pump, a proportional reversing valve and a hydraulic cylinder, and is responsible for driving the action execution of a steering and reversing mechanism, and the mechanical system is a core power transmission unit and mainly comprises an impeller, a guide vane body and a shafting, and is used for bearing fluid excitation and mechanical friction for a long time. Because the ship is in a complex marine environment for a long time, the mechanical system of the water jet propulsion device is subjected to the coupling effects of fluid excitation, unsteady cavitation load and mechanical friction at all times in the running process. Common failure modes include impeller blade damage, blade tip clearance anomalies, bearing wear, and the like. If the faults cannot be found and handled in time, not only the propulsion efficiency is greatly reduced, but also catastrophic mechanical damage accidents are caused even when the faults are serious. Therefore, accurate, efficient and interpretable fault diagnosis is a key to ensuring safe operation of the water jet propulsion device. At present, the traditional fault diagnosis method mainly depends on a signal processing technology, and a professional vibration analyst is required to manually read spectrum characteristics, so that the efficiency is low, and automatic real-time monitoring is difficult to realize. In recent years, with the development of deep learning technology, convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are widely used in mechanical fault diagnosis, and a significant breakthrough is made in fault classification accuracy. However, existing deep learning diagnostic methods still face significant challenges in practical marine engineering applications, mainly in terms of (1) lack of process interpretability, i.e., the "black box" problem is severe. The traditional deep neural network maps vibration signals into abstract feature vectors, and an operator cannot intuitively see whether the model is judged according to which part of the signals. In a water jet propulsion system, fluid noise tends to mask weak mechanical failure features, which if the model is fitted with background noise only, rather than failure impact features, would result in serious false positives or false negatives, and are not noticeable to the user during the diagnostic process. (2) lack of post-accountability and case support. When the model outputs an alarm of 'impeller big damage', only one confidence probability is usually given, and no evidence support of a physical layer can be provided. For a decision maker, such a lack of basis decision is difficult to use as a basis for decision for downtime maintenance. Disclosure of Invention In order to overcome the defects of the prior art, the invention aims to provide a fault diagnosis method of a mechanical system of a water jet propulsion device based on a time-frequency prototype network, which can realize accurate, efficient and interpretable fault diagnosis, can also realize the visual monitoring of a fault diagnosis process and the case-based reasoning type postattribution of a diagnosis result, and provides a fault diagnosis scheme with high precision and high credibility for intelligent operation and maintenance of ships. In order to achieve the purpose, the invention provides a fault diagnosis method of a mechanical system of a water jet propulsion device based on a time-frequency prototype network, which comprises the following steps: Data acquisition and time-frequency conversion are carried out through sensor measuring points arranged on a mechanical system of the water jet propulsion device, a single-channel time-frequency matrix is obtained, and a space coordinate system is defined for physical mapping; converting the single-channel time-frequency matrix into a fused RGB time-frequency diagram by using a defined RGB mapping rule, and then carrying out image processing on the fused RGB time-frequency diagram to obtain a standard input sample; collecting state data sets comprising five health states on the senso