Fault diagnosis method, medium and equipment for gas-steam cycle water pump

Abstract

The application provides a fault diagnosis method, medium and equipment for a gas-steam cycle water pump, which realizes deep fusion of multi-modal features by constructing a feature encoder and a feature enhancement module based on a convolutional neural network and introducing a cross-attention mechanism. The scheme also adopts a multi-task learning architecture, in which each task output head independently generates a corresponding diagnosis result, and finally integrates to form a comprehensive fault diagnosis report containing fault location, fault degree and fault category. This multi-task architecture not only realizes cross-task knowledge sharing and complementation, but also further improves the accuracy and comprehensiveness of diagnosis. The application also designs a fault diagnosis system and device matched therewith, which can efficiently deploy the trained model on an edge device to realize real-time, multi-dimensional, multi-granularity and extensible fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of circulating water pump fault diagnosis technology, specifically to a fault diagnosis method, medium, and equipment for a gas-steam circulating water pump. Background Technology

[0002] In combined cycle gas turbine units, the circulating water cooling system, as a core subsystem ensuring condenser heat exchange efficiency, directly determines key parameters such as unit vacuum and power generation efficiency. Among these, the circulating water pump, as the core power supply device, has a decisive impact on heat exchange efficiency. However, given that this equipment is constantly exposed to complex operating conditions such as high-speed water flow impact and dynamic load regulation, the drive motor and its associated bearing system are prone to complex failures due to multiple factors including mechanical wear, fatigue damage, foreign object intrusion, and lubrication failure. Therefore, establishing a precise and diversified bearing condition monitoring and fault diagnosis system is of significant engineering importance for ensuring safe equipment operation and improving unit heat exchange efficiency.

[0003] However, existing circulating water pump condition monitoring mainly relies on regular manual inspections and simple threshold alarms, which still have limitations in practical engineering applications, mainly in the following two aspects: 1) Traditional vibration monitoring can only trigger alarms when the bearing is severely damaged, and cannot identify early slight wear, resulting in the slow degradation of filter performance not being detected in time; 2) Most methods use single-task models, which require training models separately for different tasks, resulting in low computational efficiency and limiting the popularization and practicality of the methods. Summary of the Invention

[0004] The present invention provides a fault diagnosis method, system and device for a gas-steam circulating water pump, which can at least solve one of the technical problems in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A fault diagnosis method for a gas-fired steam circulating water pump, comprising the following steps performed via computer equipment:

[0007] S100: Collect vibration and current signals of the circulating water pump induction motor under different fault types, fault degrees and fault locations, and divide them into training set, verification set and test set after preprocessing;

[0008] S200. Construct a single-modal feature encoder based on a residual convolutional block cascade structure to extract features from the vibration and current signals of the circulating water pump induction motor. Use a time-frequency feature enhancement module to enhance feature representation and obtain multimodal fusion features based on a shared bottleneck layer and cross attention.

[0009] S300. Based on the multimodal fusion features in step S200, design corresponding task heads according to the number of tasks and task types to achieve diversified fault diagnosis.

[0010] S400: Train the model based on the training set and validation set from step S100 and the model constructed in steps S200 and S300. Update the model parameters by minimizing the multi-task weighted loss function to obtain an optimized multimodal multi-task learning model, and evaluate its performance on the test set from step S100.

[0011] Furthermore, the method for acquiring vibration signals and current signals in step S100 of the present invention includes:

[0012] Both vibration and current signals were acquired at a sampling rate of 64kHz. The vibration signal was processed by a 30kHz low-pass filter, and the current signal was processed by a 25kHz low-pass filter.

[0013] The filtered vibration signal and current signal are segmented according to a preset sliding window to construct a dataset;

[0014] The window length is 2048 sample points, and the window step size is 1024 sample points.

[0015] Furthermore, the multimodal fusion feature acquisition method of the present invention includes:

[0016] S210: Construct a single-mode feature encoder based on a residual convolutional block cascade structure to extract features from vibration and current signals;

[0017] Design a feature encoder based on cascaded one-dimensional residual convolutional blocks to extract local features from vibration / current signals;

[0018] The number of cascaded layers is 3;

[0019] A single residual convolutional block is represented as:

[0020]

[0021] in, and For the first Input and output features of layer residual convolutional blocks. This represents a one-dimensional convolution operation. This indicates a batch normalization operation, used to normalize the output of convolutions. Represents a non-linear activation function;

[0022] The nonlinear activation function is preferably the ReLU function;

[0023] Final feature output for,

[0024] ;

[0025] S220: Construct a time-frequency feature enhancement module to enhance the feature signals of the extracted vibration and current signals;

[0026] Input features The frequency domain features are mapped to the frequency domain through Fourier transform, and then extracted and normalized through point convolutional layer 1 and normalization layer. Subsequently, a nonlinear representation is introduced through activation function.

[0027] After returning to the time domain via inverse Fourier transform, the residual is element-wise added to the input features to form the residual, and then passed through a second point convolutional layer to obtain the output features. ,

[0028]

[0029]

[0030] Step S230: Construct a multimodal feature fusion module based on a shared bottleneck layer and cross-attention to fuse the enhanced feature signals of vibration and current signals and obtain fused features;

[0031] First, a global query vector is generated through a shared bottleneck layer. ,

[0032]

[0033] In this context, SA represents the self-attention computation method, and Maxpooling is the max pooling operation used for downsampling. , These are the output features of time-frequency feature enhancement 1 and time-frequency feature enhancement 2, respectively.

[0034] Subsequently, in the vibration signal branch, bonds are generated through convolutional layer 1 and convolutional layer 2. Sum In the current signal branch, bonds are generated through convolutional layer 3 and convolutional layer 4. Sum ;

[0035] Furthermore, the correlation and complementarity of dual-path features are dynamically learned through cross-attention.

[0036]

[0037]

[0038]

[0039] Here, CA is a cross-attention calculation method that uses a shared global query vector Q to jointly model common and dissimilar features among modalities. For global query vector For vibration signal branch pairs The output after weighted aggregation For global query vector For current signal branch pairs The output after weighted aggregation;

[0040] By concatenating the two output features, the fused features are obtained. .

[0041] Furthermore, the method for designing corresponding task headers based on the number of tasks and task types in step S300 of this invention to achieve diversified fault diagnosis includes:

[0042] Design a corresponding number of task heads based on the number and type of diagnostic tasks, and use fully connected layers and Softmax layers to realize the mapping from feature space to probability space, outputting diagnostic analysis results of fault location, fault degree and fault category;

[0043] For newly added diagnostic tasks, design corresponding task headers based on the diagnostic task type and add them to the existing task queue in parallel.

[0044] Furthermore, the task header of the present invention includes:

[0045] Diagnosis of fault location, fault severity, and fault type.

[0046] Furthermore, the method for optimizing the weighted loss of multiple tasks in step S400 of the present invention includes:

[0047] Joint loss calculation:

[0048] ;

[0049] in, For the task The weight coefficients are adjusted to prevent the network from converging to a local optimum for a particular task. For the corresponding task Cross-entropy loss;

[0050] Minimize the joint loss using the gradient descent-based optimization algorithm Adam. Update the model parameters.

[0051] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0052] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0053] As can be seen from the above technical solution, this invention provides an innovative fault diagnosis method, system, and device. This method fully utilizes the complementary modal information between vibration signals and current signals. By constructing a feature encoder and feature enhancement module based on a convolutional neural network and introducing a cross-attention mechanism, deep fusion of multimodal features is achieved. Furthermore, this invention employs a multi-task learning architecture, where each task output head independently generates corresponding diagnostic results, ultimately integrating them into a comprehensive fault diagnosis report containing fault location, fault severity, and fault category. This multi-task architecture not only achieves cross-task knowledge sharing and complementarity but also further improves the accuracy and comprehensiveness of the diagnosis. This invention also designs a corresponding fault diagnosis system and device, capable of efficiently deploying the trained model on edge devices to achieve real-time, multi-dimensional, multi-granular, and scalable fault diagnosis. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the method for diagnosing bearing faults in a circulating water pump.

[0055] Figure 2 The diagram shows the structural block of a circulating water pump bearing fault diagnosis model based on multimodal multi-task learning.

[0056] Figure 3 Here is a block diagram of the feature encoder module structure;

[0057] Figure 4 Here is a block diagram of the time-frequency feature enhancement module;

[0058] Figure 5 Here is a block diagram of the feature fusion module based on cross-attention;

[0059] Figure 6 For each task, there are accuracy and loss curves.

[0060] Figure 7 A fault location confusion matrix diagram;

[0061] Figure 8 A confusion matrix diagram for the degree of failure;

[0062] Figure 9 A fault category confusion matrix diagram;

[0063] Figure 10 A block diagram of a circulating water pump condition monitoring system provided in this application;

[0064] Figure 11 This application provides a structural diagram of a circulating water pump condition monitoring device. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0066] like Figure 1 As shown in the embodiment of the present invention, a method for diagnosing bearing faults in a circulating water pump involves the following steps performed using a computer device:

[0067] S100: Collect vibration and current signals of the circulating water pump induction motor under different fault locations, fault degrees and fault categories, and preprocess them to divide them into training set, validation set and test set;

[0068] S200. Construct a single-modal feature encoder based on a residual convolutional block cascade structure to extract features from the vibration and current signals of the circulating water pump induction motor. Use a time-frequency feature enhancement module to enhance feature representation and obtain multimodal fusion features based on a shared bottleneck layer and cross attention.

[0069] S300. Based on the multimodal fusion features in step S200, design corresponding task heads according to the number of tasks and task types to achieve diversified fault diagnosis.

[0070] S400: Train the model based on the training set and validation set from step S100 and the model constructed in steps S200 and S300. Update the model parameters by minimizing the multi-task weighted loss function to obtain an optimized multimodal multi-task learning model, and evaluate its performance on the test set from step S100.

[0071] The following provides a detailed explanation of each step:

[0072] S100: Collect vibration and current signals of the circulating water pump induction motor under different fault types, fault degrees and fault locations, and preprocess them to divide them into training set, validation set and test set;

[0073] During the data annotation process, professional engineers accurately labeled the collected data to indicate bearing fault conditions. Subsequently, necessary preprocessing operations were performed on the raw signals to form a multimodal, multi-task dataset. For subsequent model training and validation, the constructed dataset was divided into training, validation, and test sets at a ratio of 70%, 10%, and 20%, respectively.

[0074] In the data acquisition phase of step S100, the vibration acceleration sensor is installed on the motor housing, and the current signal is directly acquired from the motor phase current. To ensure high data quality and consistency, both the vibration and current signals are acquired at a sampling rate of 64kHz, and the acquisition process is strictly synchronized. During each acquisition, the experimental conditions remain highly consistent, i.e., the motor speed, load torque, and radial force remain consistent, and the ambient temperature is approximately controlled at 40-45℃.

[0075] The vibration signal is processed by a 30kHz low-pass filter, and the current signal is processed by a 25kHz low-pass filter.

[0076] The filtered vibration and current signals are segmented according to a preset sliding window to construct a dataset. The window length is 2048 sample points, and the window step size is 1024 sample points.

[0077] S200. Construct a single-modal feature encoder based on a residual convolutional block cascade structure to extract features from the vibration and current signals of the circulating water pump induction motor. Use a time-frequency feature enhancement module to enhance feature representation and obtain multimodal fusion features based on a shared bottleneck layer and cross-attention.

[0078] like Figure 2 As shown, the circulating water pump bearing fault diagnosis model includes: feature encoder 1, feature encoder 2, time-frequency feature enhancement module 1, time-frequency feature enhancement module 2, multi-modal feature fusion module, task head 1, task head 2, and task head 3;

[0079] The circulating water pump bearing fault diagnosis model receives vibration signals collected by a vibration acceleration sensor, performs local feature extraction via feature encoder 1, and transmits the extracted local features to time-frequency feature enhancement module 1 for mapping. While mining feature information in the frequency domain, it maintains the integrity of time domain information, providing input data for subsequent multimodal feature fusion.

[0080] The circulating water pump bearing fault diagnosis model receives the motor phase current signal, performs local feature extraction via feature encoder 2, and transmits the extracted local features to time-frequency feature enhancement module 2 for mapping. While mining feature information in the frequency domain, it maintains the integrity of time domain information, providing input data for subsequent multimodal feature fusion.

[0081] Among them, feature encoder 1 and feature encoder 2 have the same structure.

[0082] Step S210: Construct a single-mode feature encoder based on a residual convolutional block cascade structure to extract features from vibration and current signals;

[0083] like Figure 3As shown, a feature encoder based on cascaded one-dimensional residual convolutional blocks is designed to extract local features from vibration / current signals. In this embodiment, the number of cascaded layers is 3. One layer consists of a residual convolutional block. It can be represented as:

[0084]

[0085] in, and For the first Input and output features of layer residual convolutional blocks; This represents a one-dimensional convolution operation. In this embodiment, the kernel size is 5 and the stride is 2. This indicates a batch normalization operation, used to normalize the convolution output to speed up network convergence and improve generalization ability. Represents a non-linear activation function;

[0086] The nonlinear activation function is preferably the ReLU function;

[0087] After three layers of residual convolutional blocks are cascaded, the final feature output is... for,

[0088] .

[0089] Step S220: Construct a time-frequency feature enhancement module to enhance the feature signals of the extracted vibration and current signals;

[0090] like Figure 4 As shown, the time-frequency feature enhancement module based on Fourier transform enhances the feature extraction capability of the previous level, extracts feature information from the frequency domain, and at the same time preserves the integrity of the time domain information.

[0091] Input features The input features are mapped to the frequency domain via Fourier transform, then extracted and normalized by a point convolutional layer 1 and a normalization layer. A nonlinear representation is then introduced via an activation function. After returning to the time domain via inverse Fourier transform, the residual is element-wise added to the input features to form a residual, which is then passed through a point convolutional layer 2 to obtain the output features. ,

[0092]

[0093]

[0094] Step S230: Construct a multimodal feature fusion module based on a shared bottleneck layer and cross-attention to fuse the enhanced feature signals of vibration and current signals and obtain fused features;

[0095] like Figure 5As shown, a multimodal feature fusion module based on a shared bottleneck layer and cross-attention is designed to achieve deep fusion of current signal features and vibration signal features, fully explore the complementarity of the two modes, and provide accuracy for multi-task diagnosis.

[0096] Specifically, a global query vector is first generated through a shared bottleneck layer. The formula is as follows:

[0097]

[0098] Here, SA (Self-attention) is the self-attention calculation method, and Maxpooling is the max pooling operation for downsampling. , These are the output features of time-frequency feature enhancement 1 and time-frequency feature enhancement 2, respectively.

[0099] Subsequently, in the vibration signal branch, bonds are generated through convolutional layer 1 and convolutional layer 2. Sum In the current signal branch, bonds are generated through convolutional layer 3 and convolutional layer 4. Sum Furthermore, it dynamically learns the correlation and complementarity of dual-path features through cross-attention.

[0100]

[0101]

[0102]

[0103] in, For global query vector For vibration signal branch pairs The output after weighted aggregation For global query vector For current signal branch pairs The output after weighted aggregation. CA (Cross-attention) is a cross-attention calculation method that uses a shared global query vector. This enables joint modeling of common and dissimilar features across modalities. The fused features are obtained by concatenating the two output features. .

[0104] S300. Based on the multimodal fusion features in step S200, design corresponding task heads according to the number of tasks and task types to achieve diversified fault diagnosis.

[0105] The task header includes diagnoses for different fault locations, fault severity, and fault categories, corresponding to... Figure 2 Task 1-3.

[0106] The framework designs a corresponding number of task heads based on the number and type of diagnostic tasks. This embodiment uses fully connected layers and Softmax layers to map the feature space to the probability space, outputting diagnostic analysis results including fault location, fault severity, and fault category. The framework supports incremental task expansion, such as expanding tasks to include bearing remaining life prediction and complex fault diagnosis.

[0107] Specifically, for newly added diagnostic tasks, corresponding task heads can be designed according to the type of diagnostic task and added to the existing task queue in parallel. To achieve efficient model expansion, the following two optimization strategies can be selected: one is to freeze the model parameters designed and trained in step S200 and the model parameters of the existing task heads, and only optimize the model parameters of the newly added task heads; the other optional strategy is to fine-tune all model parameters to adapt to the new task and further improve the overall performance.

[0108] S400: Train the model based on the training set and validation set from step S100 and the model constructed in steps S200 and S300. Update the model parameters by minimizing the multi-task weighted loss function to obtain an optimized multimodal multi-task learning model, and evaluate its performance on the test set from step S100.

[0109] The appropriate objective function is selected based on the number and type of diagnostic tasks. For example, cross-entropy is used for classification tasks, and mean squared error is used for regression prediction tasks. In this embodiment, cross-entropy is selected as the objective function for fault location, fault severity, and fault category. The joint loss is calculated as follows:

[0110] ;

[0111] in, For the task The weight coefficients are adjusted to prevent the network from converging to a local optimum for a particular task. For the corresponding task The cross-entropy loss is calculated. The joint loss is minimized using the gradient descent-based optimization algorithm Adam. The model parameters are then updated. In this embodiment, an early stopping strategy and L2 regularization are chosen to avoid overfitting.

[0112] The evaluation results of the optimization process of the multimodal multi-task learning model using the training and validation sets are as follows: Figure 6 As shown, the trends in accuracy and loss are illustrated; the confusion matrix of the classification performance of the fault location, fault severity, and fault category diagnosis tasks on the test set is shown in the figure. Figure 7-9 As shown.

[0113] During model training, each task output head shares the aforementioned fusion features with the backbone network, and parameters are updated simultaneously through a multi-task joint optimization strategy to achieve cross-task knowledge sharing and complementarity. In the inference phase, each task output head outputs its corresponding diagnostic results, forming a comprehensive fault diagnosis report that includes fault location, fault severity, and fault category, thereby achieving multi-dimensional and multi-granular bearing condition determination.

[0114] like Figure 10 As shown, this embodiment also provides a fault diagnosis system for a circulating water pump, including:

[0115] Data acquisition module: includes vibration sensor, current sensor and corresponding amplification and filtering circuit, and uses analog-to-digital converter to convert analog signals into digital signals;

[0116] Data processing module: used to implement the preprocessing step in step S100 of this embodiment;

[0117] Fault diagnosis module: Used to extract and enhance features from the various modal data processed by the data processing module, and to fuse the extracted multimodal features, using multiple task heads to perform diverse fault diagnosis.

[0118] Comprehensive Management Module: Used for event management and fault early warning of fault diagnosis results.

[0119] On the other hand, this embodiment also provides a fault diagnosis device for a circulating water pump, such as... Figure 11 As shown, it includes a vibration sensor, a current sensor, a data acquisition input port, a filter amplifier circuit 1, a filter amplifier circuit 2, an analog-to-digital converter 1, an analog-to-digital converter 2, a data storage and processing unit, a communication interface (Ethernet port, USB expansion port), a power input port, an edge computing unit, an HMI, a buzzer, a fault indicator light, and a comprehensive management unit.

[0120] In this embodiment, vibration and current sensors are mounted on the water pump to collect real-time operating status data. After the analog signal enters the system through the acquisition input interface, it first passes through a filtering and amplification circuit to remove noise interference and enhance the signal amplitude, and then is converted into a digital signal by an analog-to-digital converter. The digital signal is transmitted to the central processing unit via the system bus and stored in local memory, while a fault diagnosis algorithm in the edge computing unit is invoked for real-time analysis. The integrated management unit manages the system based on the fault diagnosis results. If an abnormal state is detected, a buzzer and fault indicator light are immediately triggered, and alarm information is uploaded to the host computer via an Ethernet interface or USB expansion port.

[0121] Preferably, the vibration sensor in this embodiment can be a 336C04, deployed on the motor housing; the current sensor can be a LEM Typ CKSR 15-NP, which collects the motor input current.

[0122] Preferably, the edge computing unit employs a high-performance embedded GPU or AI acceleration chip, such as NVIDIA Jetson Nano, to deploy and run the multimodal multitasking model of the present invention.

[0123] Preferably, the integrated management unit, in conjunction with the processor, completes task scheduling to ensure the efficient operation of functions such as monitoring, analysis, alarm, and communication.

[0124] In summary, the present invention fully utilizes the complementary modal information between vibration and current signals. By constructing a feature encoder and feature enhancement module based on a convolutional neural network and introducing a cross-attention mechanism, deep fusion of multimodal features is achieved. Furthermore, the present invention employs a multi-task learning architecture, where each task output independently generates corresponding diagnostic results, which are ultimately integrated to form a comprehensive fault diagnosis report containing fault location, fault category, and fault severity. This multi-task architecture not only achieves cross-task knowledge sharing and complementarity but also further improves the accuracy and comprehensiveness of the diagnosis. The present invention also designs a corresponding fault diagnosis system and device, which can efficiently deploy the trained model on edge devices to achieve real-time, multi-dimensional, multi-granular, and scalable fault diagnosis.

[0125] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0126] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0127] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the fault diagnosis methods for gas-steam circulating water pumps in the above embodiments.

[0128] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0129] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0130] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0131] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method for a gas-steam circulating water pump, characterized in that, Perform the following steps using a computer device: S100: Collect vibration and current signals of the circulating water pump induction motor under different fault locations, fault degrees and fault categories, and divide them into training set, validation set and test set after preprocessing; S200. Construct a single-modal feature encoder based on a residual convolutional block cascade structure to extract features from the vibration and current signals of the circulating water pump induction motor. Use a time-frequency feature enhancement module to enhance feature representation and obtain multimodal fusion features based on a shared bottleneck layer and cross attention. S300. Based on the multimodal fusion features in step S200, design corresponding task heads according to the number of tasks and task types to achieve diversified fault diagnosis. S400: Train the model based on the training set and validation set from step S100 and the model constructed in steps S200 and S300. Update the model parameters by minimizing the multi-task weighted loss function to obtain an optimized multimodal multi-task learning model, and evaluate its performance on the test set from step S100.

2. The fault diagnosis method for a gas-steam circulating water pump according to claim 1, characterized in that, The methods for acquiring vibration and current signals in step S100 include: Both vibration and current signals were acquired at a sampling rate of 64kHz. The vibration signal was processed by a 30kHz low-pass filter, and the current signal was processed by a 25kHz low-pass filter. The filtered vibration signal and current signal are segmented according to a preset sliding window to construct a dataset; The window length is 2048 sample points, and the window step size is 1024 sample points.

3. The fault diagnosis method for a gas-steam circulating water pump according to claim 1, characterized in that, Multimodal fusion feature acquisition methods include: S210: Construct a single-mode feature encoder based on a residual convolutional block cascade structure to extract features from vibration and current signals; Design a feature encoder based on cascaded one-dimensional residual convolutional blocks to extract local features from vibration / current signals; The number of cascaded layers is 3; One layer of residual convolution block Represented as: in, and For the first Input and output features of layer residual convolutional blocks. This represents a one-dimensional convolution operation. This indicates a batch normalization operation, used to normalize the output of convolutions. Represents a non-linear activation function; The nonlinear activation function is the ReLU function; Final feature output for: ； S220: Construct a time-frequency feature enhancement module to enhance the features extracted from vibration and current signals; Input features The frequency domain features are mapped to the frequency domain through Fourier transform, and then extracted and normalized through point convolutional layer 1 and normalization layer. Subsequently, a nonlinear representation is introduced through activation function. After returning to the time domain via inverse Fourier transform, the residual is element-wise added to the input features to form the residual, and then passed through a second point convolutional layer to obtain the output features. , S230: Construct a multimodal feature fusion module based on a shared bottleneck layer and cross-attention to fuse the enhanced features of vibration signals and current signals to obtain fused features; First, a global query vector is generated through a shared bottleneck layer. , In this context, SA represents the self-attention computation method, and Maxpooling is the max pooling operation used for downsampling. , These are the output features of time-frequency feature enhancement 1 and time-frequency feature enhancement 2, respectively. Subsequently, in the vibration signal branch, bonds are generated through convolutional layer 1 and convolutional layer 2. Sum In the current signal branch, bonds are generated through convolutional layer 3 and convolutional layer 4. Sum ; Furthermore, the correlation and complementarity of dual-path features are dynamically learned through cross-attention. Here, CA is the cross-attention calculation method, which uses a shared global query vector. This enables joint modeling of common and dissimilar features among modalities. For global query vector For vibration signal branch pairs The output after weighted aggregation For global query vector For current signal branch pairs The output after weighted aggregation; By concatenating the two output features, the fused features are obtained. .

4. The fault diagnosis method for a gas-steam circulating water pump according to claim 1, characterized in that, Step S300, which combines the number of tasks and task types to design corresponding task headers to implement diverse fault diagnosis methods, includes: Design a corresponding number of task heads based on the number and type of diagnostic tasks, and use fully connected layers and Softmax layers to realize the mapping from feature space to probability space, outputting diagnostic analysis results of fault location, fault degree and fault category; For newly added diagnostic tasks, design corresponding task headers based on the diagnostic task type and add them to the existing task queue in parallel.

5. The fault diagnosis method for a gas-steam circulating water pump according to claim 4, characterized in that, The task header includes: Diagnosis of fault location, fault severity, and fault type.

6. The fault diagnosis method for a gas-steam circulating water pump according to claim 1, characterized in that, The method for updating the model parameters by minimizing the multi-task weighted loss function in step S400 to obtain the optimized multimodal multi-task learning model includes: Joint loss calculation: ； in, For the task The weight coefficients are adjusted to prevent the network from converging to a local optimum for a particular task. For the corresponding task Cross-entropy loss; Minimize the joint loss using the gradient descent-based optimization algorithm Adam. Update the model parameters.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the method as described in any one of claims 1 to 6.

8. A computer device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 6.

Images