A method and system for detecting faults of jet pumps for ships

Abstract

The present invention relates to the technical field of jet pump fault detection, specifically to a method and system for detecting jet pump faults for ships, comprising: collecting working sound data during the test of a jet pump to be tested, classifying time periods according to the use time of the jet pump, and generating a use time label; inputting the audio to be tested into a sound anomaly detection model, identifying abnormal audio samples and performing cluster analysis on them, and classifying the abnormal audio samples into three categories: jet pump shaft abnormality, impeller abnormality, and impeller bracket abnormality; according to the clustering results, disassembling and testing components in a targeted manner to obtain bending detection results, static balance detection results, and deformation detection results; when all detection results are negative, obtaining an unknown abnormal sample, and comprehensively disassembling the components of the unknown abnormality to obtain multiple components; testing the components to obtain comprehensive detection results. The present invention improves the efficiency and reliability of jet pump fault detection through automated sound recognition and targeted multi-dimensional detection.

Description

Technical Field

[0001] The invention relates to the technical field of jet pump fault detection, and in particular to a method and system for detecting jet pump faults for ships. Background Art

[0002] The marine jet pump is a key propulsion system component. The water jet propulsion system is one of the specific application forms of the jet pump. As the core component of the water jet propulsion system, the jet pump sucks water from the bottom of the ship, accelerates the water flow with the impeller, and sprays the water flow at high speed through the nozzle to generate huge thrust to drive the ship to move, aiming to improve the efficiency of the ship's travel.

[0003] In actual scenarios, as the use time of the water jet propulsion device increases, the equipment is susceptible to various factors, and problems gradually accumulate, causing the jet pump to be prone to various failures. Although the current fault detection method for marine jet pumps can identify faulty jet pumps, the location of specific fault points is often not accurate enough. For example, it is difficult to accurately locate faults in key components such as impellers, impeller brackets, and bearings. Although there are some fault detection methods that can achieve fine fault component positioning, these methods usually require maintenance personnel to disassemble the entire jet pump and inspect each component separately, which significantly increases maintenance costs and downtime. Due to the lack of modular design in the overall fault detection process, it is impossible to efficiently separate and identify faults in different components, which ultimately reduces the efficiency of jet pump detection.

[0004] Therefore, a method and system for detecting faults of a jet pump for a ship are proposed. Summary of the invention

[0005] The object of the present invention is to provide a method and system for detecting faults of jet pumps for ships, by collecting the working sound data of the jet pumps, classifying and inputting the sound anomaly detection model according to the usage time, identifying abnormal audio samples and clustering them. The clustering results are composed of jet pump shaft abnormalities, impeller abnormalities and impeller bracket abnormalities. Corresponding disassembly and detection are performed for different fault types, including: bending detection of the jet pump shaft, static balance detection of the impeller, and deformation detection of the impeller bracket to obtain bending detection results, static balance detection results and deformation detection results. When the bending detection results, static balance detection results and deformation detection results are all negative, further disassembly of unknown abnormal samples is performed to obtain multiple components, and comprehensive detection of the components is performed. This method significantly improves the efficiency and reliability of jet pump fault detection.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting faults of a jet pump for a ship, comprising:

[0008] Collect the working sound data of the jet pump to be tested during the test process to obtain the audio to be tested;

[0009] Obtaining the usage time of the spray pump to be inspected, classifying the spray pump to be inspected into time periods according to the usage time, and obtaining a usage time label;

[0010] According to the usage time label, the audio to be inspected is input into the corresponding spray pump sound anomaly detection model to identify abnormal audio samples; the spray pump sound anomaly detection model is obtained based on historical spray pump audio training and is used to detect whether the spray pump to be inspected has a fault;

[0011] Clustering all the abnormal audio samples to obtain a clustering result, wherein the clustering result consists of abnormal audio of the jet pump shaft, abnormal audio of the impeller, and abnormal audio of the impeller bracket; according to the clustering result, dividing all the abnormal audio samples into abnormal samples of the jet pump shaft, abnormal samples of the impeller, and abnormal samples of the impeller bracket;

[0012] Disassembling the abnormal impeller sample to obtain the impeller;

[0013] The impeller is placed on a balancing bracket to perform a free rotation test, and the image of the impeller in a static state is collected by rotating the impeller multiple times to obtain a static image of the impeller; the static image of the impeller is input into a trained impeller balance classification model to obtain a static balance test result; the impeller balance classification model is used to identify the balance state of the impeller according to the static image of the impeller;

[0014] Obtain an unknown abnormal sample according to the static balance detection result; comprehensively disassemble the unknown abnormal sample to obtain a plurality of component parts; and detect the component parts to obtain a comprehensive detection result.

[0015] Furthermore, the construction of the injection pump sound abnormality detection model includes the following steps:

[0016] Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels;

[0017] According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label;

[0018] The historical spray pump audio is labeled to obtain a historical fault label; the historical fault label is composed of faults and non-faults; the historical spray pump audio covers the entire working process of the spray pump, and the working process includes a startup phase, an operation phase, and a shutdown phase;

[0019] Preprocessing the historical pump audio, including denoising and normalization, to obtain historical standard pump audio data;

[0020] The audio data of historical standard jet pumps are processed using the sound activity detection method to obtain audio clips of different working stages;

[0021] Extract features of the audio clips at different working stages respectively to obtain stage audio features; normalize the stage audio features to form a stage audio feature set; the stage audio feature set includes a startup stage audio feature set, a running stage audio feature set and a shutdown stage audio feature set;

[0022] The historical fault label training pump sound anomaly detection model includes a startup phase sound anomaly detection model, an operation phase sound anomaly detection model and a shutdown phase sound anomaly detection model;

[0023] The startup phase audio feature set, the operation phase audio feature set and the shutdown phase audio feature set are respectively used as inputs, and model training is performed according to the historical fault labels to obtain the startup phase sound anomaly detection model, the operation phase sound anomaly detection model and the shutdown phase sound anomaly detection model.

[0024] Furthermore, clustering all the abnormal audio samples includes:

[0025] Performing feature extraction on the abnormal audio sample, including extracting frequency domain features, time domain features and time-frequency features of the abnormal audio sample to obtain an abnormal feature set;

[0026] Normalizing the abnormal feature set to obtain a standard abnormal feature set; the standard abnormal feature set includes a feature vector of each abnormal audio sample;

[0027] The standard abnormal feature set is clustered using the K-Means clustering algorithm, the number of clusters is set to 3, the distance between the feature vector of each abnormal audio sample and the cluster center is iteratively calculated, and the position of the cluster center is continuously adjusted until the specified conditions are met, so as to divide the standard abnormal feature set into three abnormal feature category clusters;

[0028] The cluster center features of each abnormal feature category cluster are analyzed, the cluster center features are compared with typical fault feature patterns in the expert experience library, the corresponding relationship between each abnormal feature category cluster and the jet shaft abnormality, the impeller abnormality and the impeller bracket abnormality is identified, and the clustering result is obtained.

[0029] Furthermore, the method further comprises:

[0030] Disassemble the abnormal sample of the jet pump shaft to obtain the jet pump shaft; disassemble the abnormal sample of the impeller bracket to obtain the impeller bracket;

[0031] Performing a bending test on the jet pump shaft to obtain a bending test result;

[0032] Performing deformation detection on the impeller bracket to obtain a deformation detection result;

[0033] A comprehensive fault detection result is obtained according to the bending detection result, the static balance detection result, the deformation detection result and the comprehensive detection result; and a fault repair plan is formulated according to the comprehensive fault detection result.

[0034] Further, performing the bending detection on the jet pump shaft includes:

[0035] Acquire an image of the jet pump shaft to obtain a jet pump shaft image; acquire an image acquired during bending detection of a historical jet pump shaft in a jet pump shaft operation and maintenance database to obtain a historical jet pump shaft image; acquire an image of the historical jet pump shaft image with a normal bending detection result to obtain a standard jet pump shaft image;

[0036] The jet pump shaft image is registered and aligned with the standard jet pump shaft image to obtain a jet pump shaft alignment image and a standard jet pump shaft alignment image; for the standard jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a standard jet pump shaft edge contour image; for the jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a jet pump shaft edge contour image;

[0037] According to the standard jet pump shaft edge contour image and the jet pump shaft edge contour image, the maximum edge offset of the jet pump shaft is calculated; if the maximum edge offset of the jet pump shaft exceeds a preset threshold, the bending detection result is yes.

[0038] Furthermore, the construction process of the impeller balance classification model includes:

[0039] Acquire an image of a historical impeller in a stationary state when the historical impeller in the impeller operation and maintenance database performs the free rotation test, and obtain a stationary image of the historical impeller;

[0040] Performing expert annotation on the historical impeller still image, wherein the annotation results include balance and imbalance;

[0041] Preprocessing the annotated historical impeller still image to obtain a historical standard impeller still image; the preprocessing includes: acquiring the historical impeller still image corresponding to the standard balanced impeller to obtain a standard image, annotating the centroid of the standard image to obtain a standard centroid position; registering and aligning the historical impeller still image according to the standard image, and adjusting the historical impeller still image to the same size and proportion as the standard image;

[0042] Extracting features from the historical standard impeller still image to obtain comprehensive features of the impeller; the comprehensive features of the impeller include impeller shape features, structural abnormality features and center of mass features;

[0043] The impeller shape features include impeller curvature, blade edge curvature and blade tip angle;

[0044] The structural abnormality characteristics include impeller defect characteristics, surface contamination characteristics and surface corrosion characteristics;

[0045] The center of mass characteristics include the blade center of mass position, the center of mass position and the center of mass error distance; the blade center of mass position includes the center of mass positions of M blades of the impeller; the center of mass error distance represents the distance between the predicted center of mass position and the standard center of mass position, and the predicted center of mass is the overall center of mass of the impeller.

[0046] Furthermore, the method for obtaining the predicted centroid position includes the following steps:

[0047] Using an image segmentation algorithm to segment each blade in the historical standard impeller still image separately, to obtain M blade segmentation regions;

[0048] For each of the blade segmentation regions, a regional centroid calculation method is used to obtain the centroid position of the blade;

[0049] The average of the M blade mass center positions is calculated to obtain the predicted mass center position.

[0050] Furthermore, the components are tested to obtain the comprehensive test results, including:

[0051] Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels;

[0052] According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label;

[0053] Obtaining operation and maintenance test data corresponding to the historical spray pump audio, wherein the operation and maintenance test data includes a fault label and a specific fault component;

[0054] Calculate component failure frequencies according to the operation and maintenance test data; sort the component failure frequencies from high to low to obtain component failure ranking;

[0055] According to the order in which the component failures are ranked, the component parts are tested in sequence until a failure is identified, thereby obtaining the comprehensive test result.

[0056] A jet pump fault detection system for a ship, comprising:

[0057] The sound data acquisition module is used to collect the working sound data of the test pump to be tested during the test process to obtain the audio to be tested;

[0058] A usage time acquisition and classification module is used to acquire the usage time of the spray pump to be inspected, classify the spray pump to be inspected into time periods according to the usage time, and obtain a usage time label;

[0059] A sound anomaly detection module is used to input the audio to be inspected into a corresponding spray pump sound anomaly detection model according to the usage time label to identify abnormal audio samples; the spray pump sound anomaly detection model is obtained based on historical spray pump audio training and is used to detect whether the spray pump to be inspected has a fault;

[0060] A clustering analysis module is used to cluster all the abnormal audio samples to obtain a clustering result, wherein the clustering result is composed of abnormal audio of the jet shaft, abnormal audio of the impeller and abnormal audio of the impeller bracket; according to the clustering result, all the abnormal audio samples are divided into abnormal samples of the jet shaft, abnormal samples of the impeller and abnormal samples of the impeller bracket;

[0061] The impeller static balance detection module is used to disassemble the impeller abnormal sample to obtain the impeller; place the impeller on a balance bracket to perform a free rotation test, and collect images of the impeller in a static state by rotating the impeller multiple times to obtain an impeller static image; input the impeller static image into a trained impeller balance classification model to obtain a static balance detection result; the impeller balance classification model is used to identify the balance state of the impeller according to the impeller static image;

[0062] The unknown abnormality detection module is used to obtain an unknown abnormality sample according to the static balance detection result, comprehensively disassemble the unknown abnormality sample to obtain multiple components; and detect the components to obtain a comprehensive detection result.

[0063] Furthermore, the system further comprises:

[0064] A jet pump shaft bending detection module is used to disassemble the abnormal jet pump shaft sample to obtain the jet pump shaft; perform bending detection on the jet pump shaft to obtain a bending detection result;

[0065] The impeller bracket deformation detection module is used to disassemble the abnormal sample of the impeller bracket to obtain the impeller bracket; perform deformation detection on the impeller bracket to obtain the deformation detection result.

[0066] Compared with the prior art, the present invention has the following beneficial effects:

[0067] 1. The present invention proposes a spray pump sound anomaly detection model. Different from the traditional machine learning model that is trained based on the full life cycle data of the spray pump, the spray pump sound anomaly detection model only selects the audio data of the inspected spray pumps that are in a similar life cycle to the spray pump to be inspected for model training, which significantly improves the training efficiency of the model and avoids the interference of other life cycle audio features. At the same time, the spray pump sound anomaly detection model includes a startup phase sound anomaly detection model, an operation phase sound anomaly detection model and a shutdown phase sound anomaly detection model. Compared with the traditional integrated detection method, this method can identify the acoustic characteristics of the spray pump in different working stages, thereby improving the accuracy and robustness of the initial classification of samples, and laying the foundation for subsequent fault point location. This multi-stage anomaly detection model trained based on similar life cycle historical data helps to improve the efficiency of overall spray pump fault detection and the reliability of fault point location.

[0068] 2. In the impeller static balance detection step of the present invention, an impeller balance classification model based on a still image of the impeller static balance detection is proposed. Combined with the analysis of the impeller shape characteristics, structural abnormality characteristics and center of mass characteristics, this model can more comprehensively and accurately evaluate the balance state of the impeller and cover more potential fault information. In particular, the calculation of the center of mass error distance can accurately detect the imbalance of the rotating impeller. In addition, through image registration and size standardization processing, the consistency of the image is ensured and the error caused by the shooting conditions is reduced. Compared with the traditional static balance detection technology, this impeller balance classification model based on the training of the still image of the impeller static balance detection ensures the accuracy of the static balance detection and speeds up the static balance detection process, further improving the efficiency and reliability of the overall process of fault detection and fault component location.

[0069] 3. The present invention proposes a more efficient comprehensive inspection method for unknown abnormal samples. For unknown abnormal samples that have passed the jet shaft bending detection, impeller static balance detection and impeller bracket deformation detection at the same time, by obtaining operation and maintenance test data of similar life cycles, high-frequency fault components and component detection priorities can be identified according to the component failure frequency. This detection method based on the component failure sorting order makes the detection of unknown abnormal samples more targeted. Compared with traditional fault point detection and positioning methods, this method reduces unnecessary disassembly and detection steps, optimizes the component disassembly and detection process, helps to save overall fault identification time, and can efficiently separate and identify faults of different components, ultimately improving the efficiency of jet pump detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A flow chart of a method for detecting a jet pump failure for a ship proposed in an embodiment of the present invention;

[0071] Figure 2 It is a front view of the static balance diagram detection of the impeller proposed in an embodiment of the present invention;

[0072] Figure 3 A structural diagram of a jet pump fault detection system for a ship is proposed for an embodiment of the present invention.

[0073] In the figure: 1. impeller; 2. balance bracket. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The following two embodiments are used for detailed description.

[0076] Embodiment 1:

[0077] Shipyard A's main business is ship repair and ship manufacturing. In daily operations, Shipyard A often receives small boats and medium-sized ships of different models for repair. The shipyard has a batch of fiberglass boats, which are equipped with a lightweight and flexible jet pump. The jet pump has the characteristics of light weight, flexible installation length, grease-lubricated bearings and adjustable impeller spacing. It is widely used on various types of domestic ships. Since the jet pump is an imported product, Shipyard A has not been involved in the maintenance of such jet pumps before, and lacks relevant technical information and repair cases. All maintenance and repair work involving the jet pump is the responsibility of the jet pump agent, and the jet pump is dispatched for processing. In order to save costs and gradually accumulate the technology of independently repairing such jet pump equipment, Shipyard A applied a method for fault detection of marine jet pumps, including:

[0078] See also Figure 1 S10 in the above example collects working sound data of the jet pump to be inspected during the test process to obtain the audio to be inspected. The jet pump to be inspected is a jet pump installed on a batch of fiberglass reinforced plastic boats entering shipyard A, and currently needs to be fault-checked.

[0079] See also Figure 1 In S20, the usage time of the spray pump to be inspected (a batch of spray pumps) is obtained, and the spray pump to be inspected is classified into time periods according to the usage time to obtain a usage time label; the usage time is the cumulative operating time of the spray pump, and the spray pump to be inspected is classified into time periods according to the cumulative operating time according to prior knowledge, for example, it is divided into early use spray pumps, mid-use spray pumps and late use spray pumps, and the cumulative operating time corresponding to these three types of spray pumps increases in sequence.

[0080] See also Figure 1 In S30, the audio to be inspected is input into a corresponding spray pump sound anomaly detection model according to the usage time label to identify abnormal audio samples; the spray pump sound anomaly detection model is obtained based on historical spray pump audio training and is used to detect whether the spray pump to be inspected has a fault;

[0081] Furthermore, the construction of the injection pump sound abnormality detection model includes the following steps:

[0082] Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels;

[0083] According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label;

[0084] The historical spray pump audio is labeled to obtain a historical fault label; the historical fault label is composed of faults and non-faults; the historical spray pump audio covers the entire working process of the spray pump, and the working process includes a startup phase, an operation phase, and a shutdown phase;

[0085] Preprocessing the historical pump audio, including denoising and normalization, to obtain historical standard pump audio data;

[0086] The historical standard jet pump audio data was processed using the sound activity detection method to obtain audio clips of different working stages;

[0087] Extract features of the audio clips at different working stages respectively to obtain stage audio features; normalize the stage audio features to form a stage audio feature set; the stage audio feature set includes a startup stage audio feature set, a running stage audio feature set and a shutdown stage audio feature set;

[0088] The historical fault label training pump sound anomaly detection model includes a startup phase sound anomaly detection model, an operation phase sound anomaly detection model and a shutdown phase sound anomaly detection model;

[0089] The startup phase audio feature set, the operation phase audio feature set and the shutdown phase audio feature set are respectively used as inputs, and model training is performed according to the historical fault labels to obtain the startup phase sound anomaly detection model, the operation phase sound anomaly detection model and the shutdown phase sound anomaly detection model.

[0090] The sound anomaly detection model for the spray pump proposed in the present invention constructs a sound anomaly detection model for different working stages (start-up stage, operation stage and shutdown stage) by labeling, preprocessing and feature extracting the historical audio in the spray pump operation and maintenance test database, and combining the historical usage time labels. By classifying the historical usage time of the inspected spray pumps, historical data under similar spray pump life cycles can be selected, and then audio features can be extracted for model training. The sound anomaly detection model can reduce the false detection rate and missed detection rate during the initial detection, accurately identify whether the spray pump to be inspected is faulty, and thus provide an efficient solution for the preliminary classification of the spray pumps to be inspected.

[0091] Specifically, the audio to be inspected is segmented to obtain the audio to be inspected in the startup phase, the audio to be inspected in the operation phase, and the audio to be inspected in the shutdown phase; the audio to be inspected in the startup phase is input into the corresponding startup phase sound anomaly detection model to obtain the startup phase anomaly detection result; the audio to be inspected in the operation phase is input into the corresponding operation phase sound anomaly detection model to obtain the operation phase anomaly detection result; the audio to be inspected in the shutdown phase is input into the corresponding shutdown phase sound anomaly detection model to obtain the shutdown phase anomaly detection result, and the startup phase anomaly detection result, the operation phase anomaly detection result, and the shutdown phase anomaly detection result are combined to obtain the anomaly detection result. If one of the startup phase anomaly detection result, the operation phase anomaly detection result, and the shutdown phase anomaly detection result is yes (abnormal), then the anomaly detection result is yes.

[0092] Among them, in the training stage of the corresponding spray pump sound anomaly detection model, historical spray pump audio data with the same usage time label (historical usage time label) is extracted from the historical spray pump operation and maintenance database to ensure that the data input during model training corresponds to the usage time label of the audio to be detected. For example, the usage time of the spray pump to be inspected corresponding to the audio to be inspected is 80 hours, and the spray pump is in the initial stage of spray pump use. Then the corresponding spray pump sound anomaly detection model is also trained by the historical spray pump audio data in the initial stage of spray pump use. If the abnormal detection result is yes, the corresponding spray pump to be inspected is marked as an abnormal sample, and the corresponding audio to be inspected is placed in the abnormal sample set; the spray pump sound anomaly detection model is obtained based on the historical spray pump audio training, and is used to detect whether the spray pump to be inspected has a fault;

[0093] See also Figure 1 In S40, clustering all the abnormal audio samples to obtain a clustering result, wherein the clustering result is composed of abnormal audio of the jet shaft, abnormal audio of the impeller, and abnormal audio of the impeller bracket; according to the clustering result, dividing all the abnormal audio samples into abnormal samples of the jet shaft, abnormal samples of the impeller, and abnormal samples of the impeller bracket;

[0094] Furthermore, clustering all the abnormal audio samples includes:

[0095] Performing feature extraction on the abnormal audio sample, including extracting frequency domain features, time domain features and time-frequency features of the abnormal audio sample to obtain an abnormal feature set;

[0096] Normalizing the abnormal feature set to obtain a standard abnormal feature set; the standard abnormal feature set includes a feature vector of each abnormal audio sample;

[0097] The standard abnormal feature set is clustered using the K-Means clustering algorithm, the number of clusters is set to 3, the distance between the feature vector of each abnormal audio sample and the cluster center is iteratively calculated, and the position of the cluster center is continuously adjusted until the specified condition is met, and the standard abnormal feature set is divided into three abnormal feature category clusters; in this embodiment, the specified condition is that the change in the cluster center position is less than a set threshold (0.01) or the maximum number of iterations (10 times) is reached.

[0098] The cluster center features of each abnormal feature category cluster are analyzed, the cluster center features are compared with typical fault feature patterns in the expert experience library, the corresponding relationship between each abnormal feature category cluster and the jet shaft abnormality, the impeller abnormality and the impeller bracket abnormality is identified, and the clustering result is obtained.

[0099] After different parts of the spray pump fail, the audio of the spray pump when it is working has obvious differences in sound characteristics. By classifying the abnormal audio, different fault types can be effectively distinguished and targeted detection can be performed. This clustering analysis based on the K-Means clustering algorithm can reduce the mixing of sample features, make it easier to identify specified fault modes, and reduce false positives and missed detections. At the same time, this clustering method simplifies the subsequent spray pump disassembly and component fault analysis process, effectively classifies abnormal audio samples into different types of fault feature clusters, and the classified clustering results can quickly point to the fault type of a specific component, shortening the positioning time. By comparing with the typical fault modes in the expert experience library, the specific faults corresponding to each cluster (such as abnormal spray pump shaft, abnormal impeller, and abnormal impeller bracket) can be accurately identified, laying the foundation for subsequent spray pump shaft disassembly and detailed inspection, and helping to improve the overall fault detection efficiency.

[0100] See also Figure 1 In S50, the abnormal sample of the jet pump shaft is disassembled to obtain the jet pump shaft; the abnormal sample of the impeller is disassembled to obtain the impeller; the abnormal sample of the impeller bracket is disassembled to obtain the impeller bracket; in this embodiment, the impeller of the imported jet pump is a three-piece impeller.

[0101] See also Figure 1 In S60, a bending test is performed on the jet pump shaft to obtain a bending test result;

[0102] Further, performing the bending detection on the jet pump shaft includes:

[0103] Acquire an image of the jet pump shaft to obtain a jet pump shaft image; acquire an image acquired during bending detection of a historical jet pump shaft in a jet pump shaft operation and maintenance database to obtain a historical jet pump shaft image; acquire an image of the historical jet pump shaft image with a normal bending detection result to obtain a standard jet pump shaft image;

[0104] The jet pump shaft image is registered and aligned with the standard jet pump shaft image to obtain a jet pump shaft alignment image and a standard jet pump shaft alignment image; for the standard jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a standard jet pump shaft edge contour image; for the jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a jet pump shaft edge contour image;

[0105] According to the standard jet pump shaft edge contour image and the jet pump shaft edge contour image, the maximum edge offset of the jet pump shaft is calculated; if the maximum edge offset of the jet pump shaft exceeds a preset threshold, the bending detection result is yes.

[0106] This bending detection method based on image analysis realizes non-contact automatic detection, simplifies the operation process of traditional bending detection, and greatly reduces the impact of operator operation errors on the results, which can improve the efficiency of bending detection while ensuring accuracy.

[0107] See also Figure 1 In S70, a static balance test is performed on the impeller to obtain a static balance test result;

[0108] Further, performing the static balance detection on the impeller includes:

[0109] See also Figure 2 , Figure 2 The main view of the impeller during the static balance test is shown in FIG. 1 . The disassembled impeller 1 is placed on the balance bracket 2 for a free rotation test. The impeller 1 is rotated multiple times to collect images of the impeller 1 in a static state to obtain a static image of the impeller. The image collection angle is the same as that of the impeller 1. Figure 2 As shown, it is necessary to ensure that the blades of the impeller 1 are clearly visible and in the same plane, laying the foundation for the subsequent image-based static balance detection; the impeller static image is input into the trained impeller balance classification model to obtain the static balance detection result; the static balance detection result includes balance and imbalance. The impeller balance classification model is used to identify the balance state of the impeller according to the impeller static image.

[0110] The balancing bracket 2 can allow the impeller 1 to rotate freely without resistance, so as to collect the balancing condition of the impeller 1 in a stationary state, and detect the static balancing state of the impeller 1 through multiple rotation processes.

[0111] By placing the impeller on a resistance-free balance bracket and performing multiple free rotation tests, a static balance detection method for the impeller based on image analysis is proposed, which can efficiently and accurately determine whether the impeller is in a balanced state. This detection method based on image analysis can avoid complex physical measurements. It only needs to collect images of the impeller when it is stationary to quickly determine its balance state through a trained classification model. This non-destructive and non-contact image detection can significantly improve the detection efficiency.

[0112] Furthermore, the construction process of the impeller balance classification model includes:

[0113] Acquire an image of a historical impeller in a stationary state when the historical impeller in the impeller operation and maintenance database performs the free rotation test, and obtain a stationary image of the historical impeller;

[0114] Performing expert annotation on the historical impeller still image, wherein the annotation results include balance and imbalance;

[0115] Preprocessing the annotated historical impeller still image to obtain a historical standard impeller still image; the preprocessing includes: acquiring the historical impeller still image corresponding to the standard balanced impeller to obtain a standard image, annotating the centroid of the standard image to obtain a standard centroid position; registering and aligning the historical impeller still image according to the standard image, and adjusting the historical impeller still image to the same size and proportion as the standard image;

[0116] Extracting features from the historical standard impeller still image to obtain comprehensive features of the impeller; the comprehensive features of the impeller include impeller shape features, structural abnormality features and center of mass features;

[0117] The impeller shape features include impeller curvature, blade edge curvature and blade tip angle;

[0118] The structural abnormality characteristics include impeller defect characteristics, surface contamination characteristics and surface corrosion characteristics;

[0119] The centroid feature includes the blade centroid position, the centroid position and the centroid error distance; the blade centroid position includes the centroid positions of the M blades of the impeller; the centroid error distance represents the distance between the predicted centroid position and the standard centroid position, and the predicted centroid is the overall centroid of the impeller. In this embodiment, since the impeller of the jet pump is a three-blade impeller, M is 3.

[0120] The present invention constructs an impeller balance classification model based on multi-dimensional features. The impeller balance classification model can extract the shape features, structural abnormality features and center of mass features of the impeller, and through the accurate error calculation of the center of mass position, the model can achieve higher accuracy in judging the static balance state of the impeller. Among them, the historical data is registered and aligned with the standard image to ensure the consistency of image processing, make the historical impeller static images comparable, and reduce the error of the center of mass error distance. By comprehensively analyzing key parameters such as impeller curvature, edge curvature, surface defects and center of mass error distance, the impeller balance classification model can accurately classify the balance and imbalance states of the impeller. This method is based on automatic image analysis, which not only improves the detection efficiency and reduces human operation errors, but also provides important technical guarantees for the continuous operation and maintenance of the equipment.

[0121] Furthermore, the method for obtaining the predicted centroid position includes the following steps:

[0122] Using an image segmentation algorithm to segment each blade in the historical standard impeller still image separately, obtaining three blade segmentation regions;

[0123] For each of the blade segmentation regions, a regional centroid calculation method is used to obtain the centroid position of the blade;

[0124] Calculate the centroid of each leaf area separately and get the centroid coordinates of the three leaves:

[0125] The centroid coordinates of blade 1: (X1, Y1);

[0126]

[0127] The centroid coordinates of blade 2: (X2, Y2);

[0128]

[0129] The centroid coordinates of blade 3: (X3, Y3);

[0130]

[0131] The average of the centroid positions of the three blades is calculated to obtain the predicted centroid position (X pred ,Y pred ), then:

[0132]

[0133] A method of calculating indicators based on automated image detection is used to make the acquisition of the predicted center of mass position simpler and more efficient. Through automated image segmentation and blade center of mass calculation, the center of mass of each blade can be calculated independently and accurately, which helps to improve the accuracy of the overall impeller center of mass calculation.

[0134] Where N1, N2 and N3 are the total number of pixels in the first leaf segmentation area, the second leaf segmentation area and the third leaf segmentation area respectively; x i and i Represents the horizontal and vertical coordinates of each pixel point in the corresponding leaf segmentation area.

[0135] See also Figure 1 In S80, deformation detection is performed on the impeller bracket to obtain a deformation detection result;

[0136] Optionally, performing the deformation detection on the impeller support includes:

[0137] Capture an image of the impeller bracket to obtain an impeller bracket image; obtain an image captured when a historical impeller bracket is subjected to deformation detection in an impeller bracket operation and maintenance database to obtain a historical impeller bracket image; obtain an image with a normal deformation detection result in the historical impeller bracket image to obtain a standard impeller bracket image; preprocess the impeller bracket image and the standard impeller bracket image, wherein the preprocessing includes registration and alignment to obtain an impeller bracket alignment image and a standard impeller bracket alignment image; perform contour extraction on the impeller bracket alignment image and the standard impeller bracket alignment image respectively to obtain an impeller bracket contour and a standard impeller bracket contour, and use the standard impeller bracket contour as a standard template; perform contour comparison on the impeller bracket contour and the standard template to obtain contour similarity; if the contour similarity is greater than a preset similarity threshold, the deformation detection result is no (impeller bracket deformation detection is passed).

[0138] According to the static balance test result, unknown abnormal samples are obtained (samples with a static balance test result of No are unknown abnormal samples); meanwhile, the unknown abnormal samples also include samples with a bending test result of No and samples with a deformation test result of No.

[0139] Specifically, see Figure 1 In S90, samples of which the bending detection result, the static balance detection result and the deformation detection result are negative are obtained to obtain unknown abnormal samples; the unknown abnormal samples are fully disassembled to obtain multiple component parts; and the component parts are tested to obtain comprehensive test results.

[0140] Furthermore, the components are tested to obtain the comprehensive test results, including:

[0141] Refer to Table 1 to obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels; the method of classifying the inspected spray pumps into time periods according to the historical usage time (historical usage time labels) is not limited to the form in Table 1, and those skilled in the art may adjust or subdivide the classification time periods according to specific circumstances.

[0142] According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label;

[0143] Obtain operation and maintenance test data corresponding to the historical spray pump audio, the operation and maintenance test data including a fault label and a specific faulty component; calculate component failure frequency according to the operation and maintenance test data; sort the component failure frequencies from high to low to obtain a component failure ranking; perform corresponding inspections on the components one by one in the order of the component failure ranking until the fault is identified to obtain the comprehensive inspection result.

[0144] Table 1. Example table of data in the spray pump operation and maintenance test database

[0145]

[0146] Specifically, when the components are tested in sequence according to the order of the component failure ranking, if the component is one of the jet pump shaft, impeller and impeller bracket, and the unknown abnormal sample has been tested accordingly, the repeated component detection is skipped. In this embodiment, the jet pump audio data of a jet pump to be tested in the early stage of use is classified into two categories by the corresponding jet pump sound abnormality detection model and is identified as an abnormal audio sample. After cluster analysis, it is classified as a jet pump shaft abnormal sample. After disassembly, the jet pump shaft component is obtained for bending detection, but the bending detection result is no (passed the bending test). When further comprehensive testing is performed on it, the historical jet pump audio corresponding to the inspected jet pump that is also in the early stage of use is first obtained; the jet pump operation and maintenance test data corresponding to the historical jet pump audio is obtained; the jet pump operation and maintenance test data has undergone pre-processing steps such as data cleaning; the component failure frequency is calculated according to the operation and maintenance test data; the component failure frequency is sorted from high to low to obtain the component failure ranking, as shown in Table 2, which gives the ranking of the top 5 high-frequency fault components. The unknown abnormal sample is completely disassembled to obtain multiple components, which are all sub-components of the jet pump to be inspected; each component is inspected in sequence according to the sorting order of each component in Table 2; since the unknown abnormal sample has passed the jet pump shaft bending test, this test is skipped and the sealing device test is directly carried out. If the sealing device test result is no (passed the sealing device test), the impeller test is continued, and so on.

[0147] Table 2. Component failure ranking table

[0148]

[0149] This comprehensive detection method, which sorts fault frequencies based on historical data and prioritizes detection of high-frequency fault components, significantly improves the efficiency of fault detection, reduces unnecessary full-coverage detection processes, and enables comprehensive detection to focus on high-frequency fault components, making detection more targeted and reducing detection time and maintenance costs.

[0150] Furthermore, the method also includes: obtaining a comprehensive fault detection result based on the bending detection result, the static balance detection result, the deformation detection result and the comprehensive detection result; formulating a fault maintenance plan based on the comprehensive fault detection result, refer to Table 3, which is the subsequent maintenance data of some samples that have obtained comprehensive fault detection results.

[0151] Table 3. Fault analysis and maintenance plan

[0152]

[0153] In this embodiment, the batch of spray bottles were repaired according to the fault repair plan and then reinstalled and tested for navigation. The repair of this batch of spray pumps achieved the expected effect, was recognized by the boat owner, and successfully passed the acceptance.

[0154] By combining the bending detection results, the static balance detection results, the deformation detection results and the comprehensive detection results, the comprehensive fault detection results of the batch of spray pumps can be obtained. The overall repair cost of the spray pumps can be reduced by uniformly formulating a maintenance plan based on the comprehensive fault detection results.

[0155] In order to evaluate the effectiveness of this method and obtain the key indicators of jet pump fault detection, the relevant comparative indicators before and after the application of this method are shown in Table 4.

[0156] Table 4. Comparison of fault detection methods

[0157]

[0158] Referring to Table 4, the fault detection method proposed in the present invention is superior to the dispatched repair method before the method is applied in terms of detection efficiency, detection accuracy and cost control. The detection time of each spray pump is shortened from 4 hours of dispatched repair to 2.5 hours; the fault detection accuracy is increased by 4 percentage points, from 92% to 96%; at the same time, the total cost of detection and repair is significantly reduced, and the total detection and repair cost is saved by 56,000 yuan. Therefore, the fault detection method of this embodiment not only improves the detection efficiency and accuracy, but also greatly reduces the cost, with better economic benefits and operating efficiency.

[0159] The marine jet fault detection method proposed in the present invention can quickly classify the jets to be inspected based on sound data. This method based on preliminary sound detection does not require downtime, and can perform real-time detection on a large scale when the jet is working normally. The construction process of the jet sound anomaly detection model also fully considers the differences in sound characteristics of the jet at different use stages. By classifying the use time of the jet and using the sound anomaly detection model of the historical jet audio sequence in a similar life cycle, it is possible to accurately analyze the jets under different use conditions, avoid introducing sound data from other use stages to interfere with model training, achieve accurate identification of the jet, and also improve the training efficiency of the model. After the subsequent cluster analysis of the abnormal audio samples, it can be classified according to different fault types (abnormal jet shaft, abnormal impeller, and abnormal impeller bracket), ensuring the pertinence and effectiveness of subsequent fault detection, reducing unnecessary comprehensive disassembly and detection processes, improving the overall fault detection efficiency and reducing the maintenance cost of the jet. At the same time, the method also includes a comprehensive detection of unknown faults, making the overall jet detection method more reliable.

[0160] Embodiment 2:

[0161] Shipyard B owns many ships and boats. Shipyard B is also responsible for the repair and maintenance of the jet pumps of each ship and boat. Shipyard B has applied a marine jet pump fault detection system, such as Figure 3 As shown, including:

[0162] The sound data acquisition module is used to collect the working sound data of the test pump to be tested during the test process to obtain the audio to be tested;

[0163] A usage time acquisition and classification module is used to acquire the usage time of the spray pump to be inspected, classify the spray pump to be inspected into time periods according to the usage time, and obtain a usage time label;

[0164] A sound anomaly detection module is used to input the audio to be inspected into a corresponding spray pump sound anomaly detection model according to the usage time label to identify abnormal audio samples; the spray pump sound anomaly detection model is obtained based on historical spray pump audio training and is used to detect whether the spray pump to be inspected has a fault;

[0165] A clustering analysis module is used to cluster all the abnormal audio samples to obtain a clustering result, wherein the clustering result is composed of abnormal audio of the jet shaft, abnormal audio of the impeller and abnormal audio of the impeller bracket; according to the clustering result, all the abnormal audio samples are divided into abnormal samples of the jet shaft, abnormal samples of the impeller and abnormal samples of the impeller bracket;

[0166] Specifically, the abnormal audio of the jet pump shaft, the abnormal audio of the impeller and the abnormal audio of the impeller bracket all have their corresponding jet pump IDs. According to the jet pump ID, the abnormal jet pump sample can be corresponded to obtain the abnormal sample of the jet pump shaft, the abnormal sample of the impeller and the abnormal sample of the impeller bracket.

[0167] A jet pump shaft bending detection module is used to disassemble the jet pump shaft abnormal sample to obtain the jet pump shaft (disassemble the jet pump shaft of each jet pump in the jet pump shaft abnormal sample one by one); perform bending detection on the jet pump shaft to obtain a bending detection result;

[0168] The impeller static balance detection module is used to disassemble the impeller abnormal sample to obtain the impeller (disassemble the impeller of each jet pump in the impeller abnormal sample one by one); place the impeller on a balance bracket, perform a free rotation test, and collect images of the impeller in a static state by rotating the impeller multiple times to obtain an impeller static image; input the impeller static image into a trained impeller balance classification model to obtain a static balance detection result; the impeller balance classification model is used to identify the balance state of the impeller according to the impeller static image;

[0169] An impeller bracket deformation detection module is used to disassemble the impeller bracket abnormal sample to obtain the impeller bracket (disassemble the impeller bracket of each jet pump in the impeller bracket abnormal sample one by one); perform deformation detection on the impeller bracket to obtain the deformation detection result;

[0170] The unknown abnormality detection module is used to obtain samples of the bending detection results, the static balance detection results and the deformation detection results that are negative, and obtain unknown abnormal samples; comprehensively disassemble the unknown abnormal samples to obtain multiple components; and detect the components to obtain comprehensive detection results.

[0171] Furthermore, the construction of the injection pump sound abnormality detection model includes the following steps:

[0172] Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels;

[0173] According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label;

[0174] The historical spray pump audio is labeled to obtain a historical fault label; the historical fault label is composed of faults and non-faults; the historical spray pump audio covers the entire working process of the spray pump, and the working process includes a startup phase, an operation phase, and a shutdown phase;

[0175] Preprocessing the historical pump audio, including denoising and normalization, to obtain historical standard pump audio data;

[0176] The audio data of historical standard jet pumps are processed using the sound activity detection method to obtain audio clips of different working stages;

[0177] Extract features of the audio clips at different working stages respectively to obtain stage audio features; normalize the stage audio features to form a stage audio feature set; the stage audio feature set includes a startup stage audio feature set, a running stage audio feature set and a shutdown stage audio feature set;

[0178] The historical fault label training pump sound anomaly detection model includes a startup phase sound anomaly detection model, an operation phase sound anomaly detection model and a shutdown phase sound anomaly detection model;

[0179] The startup phase audio feature set, the operation phase audio feature set and the shutdown phase audio feature set are respectively used as inputs, and model training is performed according to the historical fault labels to obtain the startup phase sound anomaly detection model, the operation phase sound anomaly detection model and the shutdown phase sound anomaly detection model.

[0180] Furthermore, clustering all the abnormal audio samples includes:

[0181] Performing feature extraction on the abnormal audio sample, including extracting frequency domain features, time domain features and time-frequency features of the abnormal audio sample to obtain an abnormal feature set;

[0182] Normalizing the abnormal feature set to obtain a standard abnormal feature set; the standard abnormal feature set includes a feature vector of each abnormal audio sample;

[0183] The standard abnormal feature set is clustered using the K-Means clustering algorithm, the number of clusters is set to 3, the distance between the feature vector of each abnormal audio sample and the cluster center is iteratively calculated, and the position of the cluster center is continuously adjusted until the specified conditions are met, so as to divide the standard abnormal feature set into three abnormal feature category clusters;

[0184] The cluster center features of each abnormal feature category cluster are analyzed, the cluster center features are compared with typical fault feature patterns in the expert experience library, the corresponding relationship between each abnormal feature category cluster and the jet shaft abnormality, the impeller abnormality and the impeller bracket abnormality is identified, and the clustering result is obtained.

[0185] Further, performing the bending detection on the jet pump shaft includes:

[0186] Acquire an image of the jet pump shaft to obtain a jet pump shaft image; acquire an image acquired during bending detection of a historical jet pump shaft in a jet pump shaft operation and maintenance database to obtain a historical jet pump shaft image; acquire an image of the historical jet pump shaft image with a normal bending detection result to obtain a standard jet pump shaft image;

[0187] The jet pump shaft image is registered and aligned with the standard jet pump shaft image to obtain a jet pump shaft alignment image and a standard jet pump shaft alignment image; for the standard jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a standard jet pump shaft edge contour image; for the jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a jet pump shaft edge contour image;

[0188] According to the standard jet pump shaft edge contour image and the jet pump shaft edge contour image, the maximum edge offset of the jet pump shaft is calculated; if the maximum edge offset of the jet pump shaft exceeds a preset threshold, the bending detection result is yes.

[0189] Further, performing the static balance detection on the impeller includes:

[0190] The disassembled impeller is placed on a balancing bracket and subjected to a free rotation test. By rotating the impeller multiple times, images of the impeller in a stationary state are collected to obtain a static image of the impeller. The static image of the impeller is input into a trained impeller balance classification model to obtain a static balance detection result. The static balance detection result includes balance and imbalance.

[0191] Furthermore, the construction process of the impeller balance classification model includes:

[0192] Acquire an image of a historical impeller in a stationary state when the historical impeller in the impeller operation and maintenance database performs the free rotation test, and obtain a stationary image of the historical impeller;

[0193] Performing expert annotation on the historical impeller still image, wherein the annotation results include balance and imbalance;

[0194] Preprocessing the annotated historical impeller still image to obtain a historical standard impeller still image; the preprocessing includes: acquiring the historical impeller still image corresponding to the standard balanced impeller to obtain a standard image, annotating the centroid of the standard image to obtain a standard centroid position; registering and aligning the historical impeller still image according to the standard image, and adjusting the historical impeller still image to the same size and proportion as the standard image;

[0195] Extracting features from the historical standard impeller still image to obtain comprehensive features of the impeller; the comprehensive features of the impeller include impeller shape features, structural abnormality features and center of mass features;

[0196] The impeller shape features include impeller curvature, blade edge curvature and blade tip angle;

[0197] The structural abnormality characteristics include impeller defect characteristics, surface contamination characteristics and surface corrosion characteristics;

[0198] The center of mass characteristics include the blade center of mass position, the center of mass position and the center of mass error distance; the blade center of mass position includes the center of mass positions of M blades of the impeller; the center of mass error distance represents the distance between the predicted center of mass position and the standard center of mass position, and the predicted center of mass is the overall center of mass of the impeller.

[0199] Furthermore, the method for obtaining the predicted centroid position includes the following steps:

[0200] Using an image segmentation algorithm to segment each blade in the historical standard impeller still image separately, to obtain M blade segmentation regions;

[0201] For each of the blade segmentation regions, a regional centroid calculation method is used to obtain the centroid position of the blade;

[0202] The average of the M blade mass center positions is calculated to obtain the predicted mass center position.

[0203] Furthermore, the components are tested to obtain the comprehensive test results, including:

[0204] Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels;

[0205] According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label;

[0206] Obtaining operation and maintenance test data corresponding to the historical spray pump audio, wherein the operation and maintenance test data includes a fault label and a specific fault component;

[0207] Calculate component failure frequencies according to the operation and maintenance test data; sort the component failure frequencies from high to low to obtain component failure ranking;

[0208] According to the order in which the component failures are ranked, the component parts are tested in sequence until a failure is identified, thereby obtaining the comprehensive test result.

[0209] Furthermore, the system also includes: obtaining a comprehensive fault detection result according to the bending detection result, the static balance detection result, the deformation detection result and the comprehensive detection result; and formulating a fault repair plan according to the comprehensive fault detection result.

[0210] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting faults of a jet pump for a ship, characterized in that: include: Collect the working sound data of the jet pump to be tested during the test process to obtain the audio to be tested; Obtaining the usage time of the spray pump to be inspected, classifying the spray pump to be inspected into time periods according to the usage time, and obtaining a usage time label; According to the usage time label, the audio to be inspected is input into the corresponding spray pump sound anomaly detection model to identify abnormal audio samples; the spray pump sound anomaly detection model is obtained based on historical spray pump audio training and is used to detect whether the spray pump to be inspected has a fault; Clustering all the abnormal audio samples to obtain a clustering result, wherein the clustering result consists of abnormal audio of the jet pump shaft, abnormal audio of the impeller, and abnormal audio of the impeller bracket; according to the clustering result, dividing all the abnormal audio samples into abnormal samples of the jet pump shaft, abnormal samples of the impeller, and abnormal samples of the impeller bracket; Disassembling the abnormal impeller sample to obtain the impeller; The impeller is placed on a balancing bracket to perform a free rotation test, and the image of the impeller in a static state is collected by rotating the impeller multiple times to obtain a static image of the impeller; the static image of the impeller is input into a trained impeller balance classification model to obtain a static balance test result; the impeller balance classification model is used to identify the balance state of the impeller according to the static image of the impeller; Obtain an unknown abnormal sample according to the static balance detection result; comprehensively disassemble the unknown abnormal sample to obtain a plurality of component parts; and detect the component parts to obtain a comprehensive detection result.

2. A method for detecting faults of a marine jet pump according to claim 1, characterized in that: The construction of the pump sound anomaly detection model includes the following steps: Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels; According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label; The historical spray pump audio is labeled to obtain a historical fault label; the historical fault label is composed of faults and non-faults; the historical spray pump audio covers the entire working process of the spray pump, and the working process includes a startup phase, an operation phase, and a shutdown phase; Preprocessing the historical pump audio, including denoising and normalization, to obtain historical standard pump audio data; The historical standard jet pump audio data was processed using the sound activity detection method to obtain audio clips of different working stages; Extract features of the audio clips at different working stages respectively to obtain stage audio features; normalize the stage audio features to form a stage audio feature set; the stage audio feature set includes a startup stage audio feature set, a running stage audio feature set and a shutdown stage audio feature set; The historical fault label training pump sound anomaly detection model includes a startup phase sound anomaly detection model, an operation phase sound anomaly detection model and a shutdown phase sound anomaly detection model; The startup phase audio feature set, the operation phase audio feature set and the shutdown phase audio feature set are respectively used as inputs, and model training is performed according to the historical fault labels to obtain the startup phase sound anomaly detection model, the operation phase sound anomaly detection model and the shutdown phase sound anomaly detection model.

3. A method for detecting a jet pump failure for a ship according to claim 1, characterized in that: Clustering all the abnormal audio samples includes: Performing feature extraction on the abnormal audio sample, including extracting frequency domain features, time domain features and time-frequency features of the abnormal audio sample to obtain an abnormal feature set; Normalizing the abnormal feature set to obtain a standard abnormal feature set; the standard abnormal feature set includes a feature vector of each abnormal audio sample; The standard abnormal feature set is clustered using the K-Means clustering algorithm, the number of clusters is set to 3, the distance between the feature vector of each abnormal audio sample and the cluster center is iteratively calculated, and the position of the cluster center is continuously adjusted until the specified conditions are met, so as to divide the standard abnormal feature set into three abnormal feature category clusters; The cluster center features of each of the abnormal feature category clusters are analyzed, the cluster center features are compared with the typical fault feature patterns in the expert experience library, the correspondence between each of the abnormal feature category clusters and the abnormal audio of the jet pump shaft, the abnormal audio of the impeller and the abnormal audio of the impeller bracket is identified, and the clustering result is obtained.

4. A method for detecting a jet pump failure for a ship according to claim 1, characterized in that: The method further comprises: Disassemble the abnormal sample of the jet pump shaft to obtain the jet pump shaft; disassemble the abnormal sample of the impeller bracket to obtain the impeller bracket; Performing a bending test on the jet pump shaft to obtain a bending test result; Performing deformation detection on the impeller bracket to obtain a deformation detection result; A comprehensive fault detection result is obtained according to the bending detection result, the static balance detection result, the deformation detection result and the comprehensive detection result; and a fault repair plan is formulated according to the comprehensive fault detection result.

5. A method for detecting faults of a jet pump for a ship according to claim 4, characterized in that: The bending detection of the jet pump shaft includes: Acquire an image of the jet pump shaft to obtain a jet pump shaft image; acquire an image acquired during bending detection of a historical jet pump shaft in a jet pump shaft operation and maintenance database to obtain a historical jet pump shaft image; acquire an image of the historical jet pump shaft image with a normal bending detection result to obtain a standard jet pump shaft image; The jet pump shaft image is registered and aligned with the standard jet pump shaft image to obtain a jet pump shaft alignment image and a standard jet pump shaft alignment image; for the standard jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a standard jet pump shaft edge contour image; for the jet pump shaft alignment image, the outer edge of the jet pump shaft is extracted using the Canny edge detection method to obtain a jet pump shaft edge contour image; According to the standard jet pump shaft edge contour image and the jet pump shaft edge contour image, the maximum edge offset of the jet pump shaft is calculated; if the maximum edge offset of the jet pump shaft exceeds a preset threshold, the bending detection result is yes.

6. A method for detecting faults of a jet pump for a ship according to claim 1, characterized in that: The construction process of the impeller balance classification model includes: Acquire an image of a historical impeller in a stationary state when the historical impeller in the impeller operation and maintenance database performs the free rotation test, and obtain a stationary image of the historical impeller; Performing expert annotation on the historical impeller still image, wherein the annotation results include balance and imbalance; Preprocessing the annotated historical impeller still image to obtain a historical standard impeller still image; the preprocessing includes: acquiring the historical impeller still image corresponding to the standard balanced impeller to obtain a standard image, annotating the centroid of the standard image to obtain a standard centroid position; registering and aligning the historical impeller still image according to the standard image, and adjusting the historical impeller still image to the same size and proportion as the standard image; Extracting features from the historical standard impeller still image to obtain comprehensive features of the impeller; the comprehensive features of the impeller include impeller shape features, structural abnormality features and center of mass features; The impeller shape features include impeller curvature, blade edge curvature and blade tip angle; The structural abnormality characteristics include impeller defect characteristics, surface contamination characteristics and surface corrosion characteristics; The center of mass characteristics include the blade center of mass position, the center of mass position and the center of mass error distance; the blade center of mass position includes the center of mass positions of M blades of the impeller; the center of mass error distance represents the distance between the predicted center of mass position and the standard center of mass position, and the predicted center of mass is the overall center of mass of the impeller.

7. A method for detecting faults of a jet pump for a ship according to claim 6, characterized in that: The method for obtaining the predicted centroid position comprises the following steps: Using an image segmentation algorithm to segment each blade in the historical standard impeller still image separately, to obtain M blade segmentation regions; For each of the blade segmentation regions, a regional centroid calculation method is used to obtain the centroid position of the blade; The average of the M blade mass center positions is calculated to obtain the predicted mass center position.

8. A method for detecting faults of a jet pump for a ship according to claim 1, characterized in that: The components are tested to obtain the comprehensive test results including: Obtain the historical usage time of the inspected spray pumps in the spray pump operation and maintenance test database, classify the inspected spray pumps into time periods according to the historical usage time, and obtain historical usage time labels; According to the usage time label, the corresponding historical spray pump audio in the spray pump operation and maintenance test database is obtained; the historical usage time label of the inspected spray pump corresponding to the historical spray pump audio is the same as the usage time label; Obtaining operation and maintenance test data corresponding to the historical spray pump audio, wherein the operation and maintenance test data includes a fault label and a specific fault component; Calculate component failure frequencies according to the operation and maintenance test data; sort the component failure frequencies from high to low to obtain component failure ranking; According to the order in which the component failures are ranked, the component parts are tested in sequence until a failure is identified, thereby obtaining the comprehensive test result.

9. A jet pump fault detection system for a ship, characterized in that: include: The sound data acquisition module is used to collect the working sound data of the test pump to be tested during the test process to obtain the audio to be tested; A usage time acquisition and classification module is used to acquire the usage time of the spray pump to be inspected, classify the spray pump to be inspected into time periods according to the usage time, and obtain a usage time label; A sound anomaly detection module is used to input the audio to be inspected into a corresponding spray pump sound anomaly detection model according to the usage time label to identify abnormal audio samples; the spray pump sound anomaly detection model is obtained based on historical spray pump audio training and is used to detect whether the spray pump to be inspected has a fault; A clustering analysis module is used to cluster all the abnormal audio samples to obtain a clustering result, wherein the clustering result is composed of abnormal audio of the jet shaft, abnormal audio of the impeller and abnormal audio of the impeller bracket; according to the clustering result, all the abnormal audio samples are divided into abnormal samples of the jet shaft, abnormal samples of the impeller and abnormal samples of the impeller bracket; The impeller static balance detection module is used to disassemble the impeller abnormal sample to obtain the impeller; place the impeller on a balance bracket to perform a free rotation test, and collect images of the impeller in a static state by rotating the impeller multiple times to obtain an impeller static image; input the impeller static image into a trained impeller balance classification model to obtain a static balance detection result; the impeller balance classification model is used to identify the balance state of the impeller according to the impeller static image; The unknown abnormality detection module is used to obtain an unknown abnormality sample according to the static balance detection result, comprehensively disassemble the unknown abnormality sample to obtain multiple components; and detect the components to obtain a comprehensive detection result.

10. A marine jet pump fault detection system according to claim 9, characterized in that: The system further comprises: A jet pump shaft bending detection module is used to disassemble the abnormal jet pump shaft sample to obtain the jet pump shaft; perform bending detection on the jet pump shaft to obtain a bending detection result; The impeller bracket deformation detection module is used to disassemble the abnormal sample of the impeller bracket to obtain the impeller bracket; perform deformation detection on the impeller bracket to obtain the deformation detection result.

Images