An intelligent operation and maintenance system for machine pumps fusing deep learning and fine-tuning large models

Abstract

The application discloses a kind of fusion deep learning and fine-tuning big model's machine pump intelligent operation and maintenance system, belong to industrial equipment intelligent operation and maintenance, artificial intelligence and industrial internet technology field.Specifically include: data acquisition processing module, deep learning module, agent module, BS interaction module;Data acquisition processing module is used to filter the vibration acceleration signal of the collected machine pump and carry out noise reduction, obtains vibration acceleration filter signal, utilizes frequency domain integration algorithm to integrate vibration acceleration filter signal into vibration velocity signal, to extract machine pump rotor operating state feature;The data acquisition processing module is also used to pass vibration acceleration filter signal to deep learning module, pass through the fault diagnosis model of well-trained in deep learning module output machine pump rolling bearing operating state category, based on network transmission protocol, rotor operating state feature and rolling bearing operating state category are pushed to BS interaction module, for the agent module of embedding BS interaction module call.The application innovatively fuses the real-time acquisition analysis capability of data acquisition processing module, the accurate perception capability of deep learning and the cognitive reasoning capability of big language model agent, relies on B / S architecture, and creates a set of "state monitoring, fault diagnosis, three-dimensional visualization, operation and maintenance decision" integrated intelligent platform.

Description

Technical Field

[0001] This invention relates to the fields of intelligent operation and maintenance of industrial equipment, artificial intelligence and industrial internet technology, specifically to an intelligent operation and maintenance system for pumps and motors that integrates deep learning and fine-tuning of large models. Background Technology

[0002] Pumps and motors, as core equipment in industrial fluid transport systems, are widely used in key fields such as petrochemicals, power, metallurgy, and water conservancy, and are hailed as the "heart" of industrial systems. Their working principle is based on fluid mechanics, using impeller rotation to generate centrifugal force for fluid transport. However, pumps and motors are prone to various failures during long-term operation, typical failure modes including bearing failure, misalignment, imbalance, rubbing, loosening, cavitation, condensation, and poor lubrication. However, due to the complexity of pumps and motors and the diversity of their operating environments, the technical threshold for fault diagnosis is high. Traditional manual inspection and diagnosis are inefficient and costly, no longer meeting current needs, and there is an urgent need for more efficient monitoring and diagnostic methods.

[0003] Deep learning, as one of the most cutting-edge technologies in the field of machine learning, has demonstrated enormous application potential in the field of rolling bearing fault diagnosis. Firstly, deep learning can automatically extract key features of pump rolling bearings from large amounts of historical operating data without relying on manually designed feature engineering. This not only significantly improves the accuracy and reliability of diagnosis but also reduces the development difficulty of the diagnostic system. Secondly, deep learning models have powerful nonlinear modeling capabilities, enabling them to capture complex fault modes of pump rolling bearings, thereby achieving accurate identification of fault types. Furthermore, deep learning supports end-to-end learning, allowing fault diagnosis results to be directly extracted from pump operating status parameters, greatly simplifying the entire diagnostic process.

[0004] Furthermore, while large-scale modeling technology has demonstrated great potential, it still has significant limitations in industrial pump and compressor maintenance scenarios. Although general-purpose large-scale models are knowledgeable, they lack domain-specific terminology and fault mechanism knowledge, making them prone to "illusions" and providing seemingly reasonable but actually erroneous maintenance suggestions. In addition, general-purpose large-scale models typically require massive computing resources for training and inference, making direct deployment in industrial fields extremely costly and impractical. How to "train" general-purpose large-scale models into knowledgeable "pump and compressor experts" while reducing reliance on computing power is a pressing problem that the industry needs to solve.

[0005] Therefore, the key issue that needs to be addressed is how to combine lightweight deep learning technology to achieve efficient edge perception, and at the same time use fine-tuning large model technology to achieve cognitive reasoning and interpretable decision-making under limited computing power, so as to build an intelligent pump operation and maintenance system that integrates "state monitoring, fault diagnosis, 3D visualization, and operation and maintenance decision-making". Summary of the Invention

[0006] This invention provides an intelligent operation and maintenance system for pumps and turbines that integrates deep learning and fine-tuning of large models. This system innovatively combines the real-time data acquisition and analysis capabilities of a data acquisition and processing module, the precise perception capabilities of deep learning, and the cognitive reasoning capabilities of a large language model (LLM) agent. Based on a B / S (browser / server) architecture, it creates an intelligent platform integrating "status monitoring, fault diagnosis, 3D visualization, and operation and maintenance decision-making." The system aims to address the pain points of traditional pump and turbine operation and maintenance, such as difficulty in fault identification, high reliance on expert experience, and unintuitive human-computer interaction, thereby improving the operational reliability and maintenance efficiency of pump and turbine units.

[0007] The technical solution of this invention is:

[0008] A pump intelligent operation and maintenance system integrating deep learning and fine-tuning of a large model includes: a data acquisition and processing module, a deep learning module, an intelligent agent module, and a B / S interaction module. The data acquisition and processing module filters and reduces noise from the acquired vibration acceleration signals of the pump to obtain a filtered vibration acceleration signal. It then uses a frequency domain integration algorithm to integrate the filtered vibration acceleration signal and convert it into a vibration velocity signal to extract the pump rotor operating state features. The data acquisition and processing module transmits the filtered vibration acceleration signal to the deep learning module, outputs the pump rolling bearing operating state category through a fault diagnosis model trained in the deep learning module, and pushes the rotor operating state features and rolling bearing operating state category to the B / S interaction module based on a network transmission protocol for use by the intelligent agent module embedded in the B / S interaction module.

[0009] Furthermore, the training of the fault diagnosis model in the deep learning module specifically involves: preprocessing and labeling the vibration acceleration signals in the acquired pump fault dataset; performing image conversion on the labeled vibration acceleration filtered signals to construct a training set; and inputting the training set into the Ultra-Light LSR-Net deep learning model for training to obtain the trained fault diagnosis model.

[0010] Furthermore, the image transformation operation specifically involves: for images of length... The mapping relationship between the vibration acceleration filter signal and the converted two-dimensional grayscale image is expressed by the following formula:

[0011] ;

[0012] in, This represents the pixel brightness value at the corresponding location in a two-dimensional grayscale image. ; Indicates the rounding operation; The amplitude of the vibration acceleration filter signal; and Signals The maximum and minimum values.

[0013] Furthermore, the Ultra-Light LSR-Net deep learning model includes:

[0014] The input layer is used to receive a two-dimensional grayscale image X obtained by image transformation operation;

[0015] The preprocessing module is used to perform preliminary feature encoding on the original input, and the expression is:

[0016] ;

[0017] in, For output of the preprocessing module, This represents the Hardswish activation function, where BN(·) is the batch normalization operation. This represents a convolution operation with a kernel size of 3×3 and a stride of 2, where X represents a single-channel two-dimensional grayscale image;

[0018] Lite-CES Block 1, which takes the output of the preprocessing module as input, is used to extract mid-to-shallow local texture features. ;

[0019] Lite-CES Block 2, which takes the output of Lite-CES Block 1 as input, is used to extract high-level semantic features. ;

[0020] A global average pooling layer, wherein the global average pooling layer uses high-level semantic features The input is used to extract the global mean and construct a one-dimensional feature vector;

[0021] A fully connected classifier takes a one-dimensional feature vector as input and outputs the classification result of the running state category.

[0022] Furthermore, the Lite-CES Block 1 specifically comprises:

[0023] The output of the preprocessing module is divided into two parts along the channel dimension: ;

[0024] in, and These represent the outputs of the preprocessing module. The two sets of channel features after splitting;

[0025] right use And a depthwise convolution operation with a stride of 1 is obtained , represented as:

[0026] ;

[0027] Output of depthwise convolution operation use Pointwise convolution to obtain , represented as:

[0028] ;

[0029] Output of pointwise convolution and After stitching, a Channel Shuffle operation is introduced to obtain mid-to-shallow local texture features. , represented as:

[0030] .

[0031] Furthermore, the Lite-CES Block 2 specifically comprises:

[0032] Local texture features of the middle and shallow layers Divided into two parts along the channel dimension:

[0033] ;

[0034] in, and These represent local texture features in the middle and shallow layers, respectively. The two sets of channel features after splitting;

[0035] right Use in sequence And depthwise convolution with stride of 2, Pointwise convolution operation to obtain , represented as:

[0036] ;

[0037] right Lightweight downsampling mapping is used to obtain , represented as:

[0038] ;

[0039] in, Indicates a lightweight downsampling map;

[0040] The output after pointwise convolution and the output of lightweight downsampling mapping After concatenation, a ChannelShuffle operation is introduced to obtain high-level semantic features. , is represented as:

[0041] .

[0042] Furthermore, the intelligent agent module specifically comprises: firstly, completing industrial document parsing and vector database construction; then, using the constructed vector database as a knowledge source, employing RAG and data distillation techniques to construct a high-quality domain fine-tuning dataset; fine-tuning the high-quality domain fine-tuning dataset using LoRA to form a large-scale intelligent diagnostic model for pumps and rigs, which is then encapsulated to obtain the intelligent agent module; finally, embedding the intelligent agent module into the BS interaction module, which generates interpretable pump and rig fault diagnosis conclusions and maintenance suggestions based on the pump rotor operating status characteristics, rolling bearing operating status categories, and vector database retrieval results.

[0043] The beneficial effects of this invention are:

[0044] This invention proposes an intelligent operation and maintenance system architecture for pumps and turbines that integrates deep learning and fine-tuning of large models. The system includes a data acquisition and processing module, a deep learning module, an intelligent agent module, and a B / S interaction module. It combines LabVIEW's efficient data acquisition capabilities with Python's deep learning analysis advantages and achieves real-time cross-platform data interaction through network transmission protocols, thereby improving system development efficiency and flexibility. The data acquisition and processing module collects pump operating status parameters in real time, extracts rotor operating status characteristics through time-frequency domain analysis, and transmits the preprocessed vibration acceleration signal to the deep learning module. The deep learning module provides the rolling bearing operating status category, and the rotor operating status characteristics and rolling bearing operating status category are transmitted to the agent module. The agent module performs logical reasoning to provide pump operating status assessment and maintenance suggestions. The interactive interface adopts a B / S architecture, imports the pump's 3D model, supports rotation, scaling, and other operations, allows for in-depth viewing of the internal structure, and displays the pump operating status in real time. Vibration data is mapped to the corresponding positions in the 3D model for visualization. When vibration is abnormal, the abnormal area is displayed intuitively, along with the rolling bearing health status and rotor operating status characteristics. Through interaction with the agent, the large language model agent integrates the deep learning rolling bearing diagnostic conclusions and rotor operating status characteristics to generate interpretable pump fault diagnosis conclusions and maintenance suggestions. This enables convenient multi-terminal access and system integration, forming a complete closed loop from status monitoring, fault diagnosis, visualization display to maintenance decision execution, significantly improving pump operating reliability, maintenance efficiency, and human-machine interaction experience. Attached Figure Description

[0045] Figure 1This is a structural block diagram of the present invention.

[0046] Figure 2 This is a diagram illustrating the principle of frequency domain integration.

[0047] Figure 3 This is a diagram illustrating the principle of image conversion operations.

[0048] Figure 4 This is the architecture diagram of the Ultra-Light LSR-Net deep learning model.

[0049] Figure 5 This is a block diagram of the intelligent agent module.

[0050] Figure 6 These are the loss function and accuracy curves during the training process under the model of this invention.

[0051] Figure 7 These are the performance results of each running state category in the test set under the model of this invention.

[0052] Figure 8 This is the confusion matrix of the test set under the model of this invention.

[0053] Figure 9 These are the performance results for each running state category of the test set under the ablation comparison model.

[0054] Figure 10 It is the confusion matrix of the test set under the ablation comparison model.

[0055] Figure 11 This is the initial login interface for the BS interaction module.

[0056] Figure 12 It is a visual monitoring interface.

[0057] Figure 13 This is the second visual monitoring interface.

[0058] Figure 14 This is the third visual monitoring interface.

[0059] Figure 15 This is the fourth visual monitoring interface.

[0060] Figure 16 It is an analysis and diagnosis Figure 1 .

[0061] Figure 17 It is an analysis and diagnosis Figure 2 .

[0062] Figure 18 It is the result of the intelligent agent generation. Figure 1 .

[0063] Figure 19 It is the result of the intelligent agent generation. Figure 2 . Detailed Implementation

[0064] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0065] Example 1: As Figures 1-19 As shown, a pump intelligent operation and maintenance system integrating deep learning and fine-tuning of a large model includes: a data acquisition and processing module, a deep learning module, an intelligent agent module, and a B / S interaction module; details are as follows:

[0066] I. Data Acquisition and Processing Module:

[0067] The data acquisition and processing module, acting as the system's "sensory nerves," is responsible for acquiring and initially processing high-quality data, laying the foundation for subsequent analysis. This invention, built on LabVIEW and Python, mainly includes the following:

[0068] (1) Pump and motor data acquisition system

[0069] Hardware layer: Composed of sensor modules (such as vibration acceleration sensors, temperature sensors, etc. In this embodiment of the invention, a vibration acceleration sensor is used as the sensor module. The vibration acceleration sensor is installed in the vertical direction of the pump rolling bearing position, and the pump rolling bearing is used to support the pump rotor), data acquisition card and host computer. The sensor module is connected to the data acquisition card, and the data acquisition card is connected to the host computer through a USB interface; for example, the host computer is a computer.

[0070] Software layer: The data acquisition and monitoring module is developed based on LabVIEW. As the host computer software of the hardware layer, it is responsible for controlling the hardware layer and realizing signal acquisition, real-time display and storage.

[0071] Data Layer: The data layer is primarily responsible for the entire data flow and status presentation. The software layer sends commands downwards to control the hardware layer to acquire signals. The acquired raw vibration acceleration signals are transmitted to the host computer in real time and dynamically mapped into time-domain waveforms, spectrum, and amplitude visualizations in the LabVIEW front panel. Simultaneously, the software layer supports online configuration and dynamic adjustment of key acquisition parameters such as sampling frequency and number of sampling points. It also persistently stores the acquired raw acceleration signals, combined with precise timestamps, as TDMS or CSV format files, providing a foundation for subsequent data backtracking and in-depth analysis.

[0072] (2) Signal preprocessing unit

[0073] A signal preprocessing unit was built using Python. This unit uses a bandpass filter to filter and reduce noise in the vibration acceleration signal acquired by the pump data acquisition system, resulting in a filtered vibration acceleration signal. Then, a frequency domain integration algorithm is used to integrate the filtered vibration acceleration signal and convert it into a vibration velocity signal. The frequency domain integration algorithm flow is as follows: Figure 2 As shown, the principle of frequency domain integration is that it first transforms the time-domain acceleration filter signal to the frequency domain through a Fast Fourier Transform (FFT), at which point the signal is decomposed into components of different frequencies. Then, simple algebraic operations (i.e., division by frequency factors) are performed on these frequency components in the frequency domain to complete the integration conversion from acceleration to velocity. During this process, the "low-frequency zeroing" technique can effectively eliminate the integration drift error caused by DC components and low-frequency noise. Finally, the inverse Fast Fourier Transform (IFFT) is used to restore the processed frequency domain signal back to the time domain to obtain an accurate vibration velocity signal.

[0074] (3) Feature extraction unit

[0075] The feature extraction unit is used to convert the vibration velocity signal to the frequency domain using FFT, and extract the amplitude ratio of the rotation frequency and its harmonics as the operating state features of the pump rotor;

[0076] (4) Information transmission unit

[0077] The vibration acceleration filter signal is transmitted to the deep learning module, which outputs the rolling bearing operating state category. Based on the network transmission protocol, the rotor operating state characteristics and rolling bearing operating state category are transmitted to the intelligent agent module. The intelligent agent module performs logical reasoning, covering the pump rolling bearing and rotor, to generate interpretable pump fault diagnosis conclusions and maintenance suggestions. The information transmission unit and the BS interaction module adopt a full-duplex real-time communication mode based on WebSocket, which is divided into the following four key stages:

[0078] 1. Service setup and static resource hosting

[0079] Web server startup: Use uvicorn as an ASGI server to start a FastAPI application instance.

[0080] Front-end page loading: Using StaticFiles and FileResponse, the server hosts and sends the HTML, JS, and CSS files required by the B / S interaction module (browser) to the B / S interaction module. After the browser loads the page, it prepares to establish a live connection.

[0081] 2. Establish a WebSocket persistent connection

[0082] Connection: JavaScript on the browser side (BS interaction module) initiates a WebSocket connection request.

[0083] Routing: FastAPI captures the request via the @app.websocket decorator and establishes a persistent TCP connection.

[0084] Cross-domain support: CORSMiddleware plays a role at this stage, ensuring that the browser allows the WebSocket handshake to pass if the front-end and back-end ports are inconsistent.

[0085] 3. Data Acquisition and Signal Processing

[0086] File reading: The Python backend uses the nptdms.TdmsFile library to read vibration data files in TDMS format acquired and stored by LabVIEW.

[0087] Data transformation: Use NumPy to convert the read raw data into array format, and call the signal preprocessing unit for preprocessing and the feature extraction unit for feature extraction;

[0088] 4. Real-time data push

[0089] Asynchronous streaming: Utilizing the asynchronous features of asyncio and FastAPI, the Python backend transmits data in blocks or by time steps within a loop.

[0090] Sending data: The obtained pump rotor operating status characteristics and rolling bearing operating status categories are pushed to the BS interaction module in real time through the .send_json() or .send_bytes() methods of the WebSocket object, so that the intelligent agent module embedded in the BS interaction module can call them.

[0091] Time synchronization: The browser uses the time or datetime module to timestamp each data packet to ensure that the data displayed on the BS is synchronized with the collection time.

[0092] II. Deep Learning Module:

[0093] This deep learning module first uses a pump fault dataset as its foundation. The dataset contains vibration acceleration signals of the equipment under normal and different fault conditions, providing sample support for subsequent diagnosis. Next, the pump fault dataset undergoes data preprocessing (using a bandpass filter to filter and reduce noise in the vibration acceleration signals, resulting in filtered vibration acceleration signals) and is labeled. The labeled filtered vibration acceleration signals are then subjected to image transformation. Based on this, the processed samples are divided into training and testing sets according to a preset ratio. The training set is used to learn the feature distribution patterns under different operating conditions, while the testing set is used to evaluate the model's ability to identify unknown samples and its generalization performance. Then, the data is input into the Ultra-Light LSR-Net deep learning model for training, resulting in a trained fault diagnosis model capable of identifying pump operating state categories. Finally, the trained fault diagnosis model is used to identify test samples (which can be filtered vibration acceleration signals provided by the data acquisition and processing module), outputting the corresponding operating state category, thus providing a basis for pump equipment condition monitoring, fault early warning, and maintenance decisions.

[0094] To reduce computational load, improve efficiency, and maintain diagnostic accuracy, the core technologies of the deep learning module involve image transformation operations and the Ultra-Light LSR-Net deep learning model.

[0095] refer to Figure 3 The image conversion operation is as follows:

[0096] The principle of converting a one-dimensional signal to a two-dimensional grayscale image is as follows: For a length of... The vibration acceleration filter signal can be directly reconstructed into an N×N two-dimensional grayscale image. The sampling points of the one-dimensional time-domain signal are mapped to the corresponding pixel values ​​in the two-dimensional grayscale image. The brightness value of each pixel in the two-dimensional grayscale image ranges from 0 to 255, i.e., 8 gray levels. The mapping relationship between the original one-dimensional signal and the converted two-dimensional grayscale image can be expressed by the following formula:

[0097] ;

[0098] in, The amplitude of the vibration acceleration filter signal; and Signals The maximum and minimum values; This represents the pixel brightness value at the corresponding location in a two-dimensional grayscale image. ; This indicates the rounding operation.

[0099] Compared to time-frequency analysis methods such as wavelet transform, short-time Fourier transform, and S-transform, this conversion method has lower time complexity and is more convenient for practical online monitoring and diagnosis. Therefore, this paper adopts this method to convert the time-domain vibration acceleration filtered signals of bearings under different states into corresponding two-dimensional grayscale images, constructs a relevant fault sample library, and then designs an Ultra-LightLSR-Net deep learning model to achieve intelligent identification of bearing operating state categories.

[0100] To achieve efficient recognition of grayscale fault images, this invention designs the Ultra-Light LSR-Net deep learning model (ULLR-Net). This model takes a single-channel two-dimensional grayscale image as input, extracts shallow features through the PreprocessBlock preprocessing module, and progressively completes feature enhancement and spatial downsampling using a two-stage Lite-CES Block module. Finally, it combines Global Average Pooling (GAP) with a fully connected classifier (FullyConnected) to output the running state category. The overall structure is as follows: Figure 4 As shown, the network backbone process is as follows: Input (GrayscaleImage) → Preprocess Block → Lite-CES Block 1 → Lite-CES Block 2 → GAP → Fully Connected → Output (5 Fault Classes). This structure balances classification performance with the number of model parameters, computational complexity, and deployment efficiency, making it suitable for fault diagnosis scenarios that require lightweight and real-time performance. Detailed explanations are as follows:

[0101] 1. Overall Network Structure

[0102] This invention presents the Ultra-Light LSR-Net deep learning model, designed for a six-class classification task. The input is a single-channel grayscale image, and the output is the corresponding six class labels. The entire network consists of the following parts:

[0103] Input Layer: Receives two-dimensional grayscale images;

[0104] Preprocessing Block: Completes shallow feature extraction and preliminary downsampling;

[0105] Lite-CES Block 1: Extracts local texture features in the mid-to-shallow layers and enhances channel interaction;

[0106] Lite-CES Block 2: While maintaining lightweight design, it achieves further downsampling and advanced semantic feature extraction;

[0107] Global Average Pooling (GAP): Compresses spatial features into a channel-level global description;

[0108] Fully Connected Classifier: Outputs classification results for six running state categories.

[0109] Structurally, the network follows the design concept of "shallow preprocessing - lightweight feature extraction - global aggregation - classification output". It retains the ability of convolutional networks to extract local texture and structural information, and reduces model overhead through mechanisms such as depthwise separable convolution, channel splitting and channel rearrangement.

[0110] 2. Input Layer and Preprocessing Module

[0111] (1) Input layer

[0112] The model input is a single-channel two-dimensional grayscale image, which can be represented as:

[0113] ;

[0114] in, Represents a single-channel two-dimensional grayscale image. The dimension is , and These represent the height and width of the input image, respectively. Compared to a three-channel color image, grayscale input reduces redundant color information, making it more suitable for structured visual representation tasks such as fault images or time-frequency maps.

[0115] (2) Preprocessing Block

[0116] The preprocessing module is used to perform preliminary feature encoding on the raw input, and its basic components include: convolution( (Step size of 2), Batch Normalization (BatchNorm), and Hardswish activation function. Their functions can be summarized in the following three points:

[0117] First, extract shallow local features. Through... Convolution encodes edges, textures, and local structures in an image, providing a basic representation for subsequent lightweight modules.

[0118] Second, reduce the spatial resolution of the feature maps. Using convolutions with stride=2 can achieve spatial compression directly at the front end of the network, expanding the receptive field while reducing subsequent computation.

[0119] Third, it enhances training stability and non-linear expressive power. BatchNorm helps stabilize feature distribution, while Hardswish balances non-linear expressive power with low computational cost, making it suitable for lightweight network design.

[0120] Assume the output of the preprocessing module is Then we have:

[0121] ;

[0122] in, This represents the Hardswish activation function. BN(·) is the batch normalization operation, which independently calculates the mean and variance of the mini-batch for each channel and performs normalization and affine transformation. This represents a convolution operation with a kernel size of 3×3 and a stride of 2, where X represents a single-channel two-dimensional grayscale image.

[0123] 3. Lite-CES Block 1

[0124] Lite-CES Block 1 is mainly used for shallow to medium-level feature extraction. Its core idea is to improve feature representation capabilities with low computational cost through channel splitting, depthwise convolution, pointwise convolution and channel rearrangement.

[0125] (1) Channel Split

[0126] First, the input features are divided into two parts along the channel dimension:

[0127] ;

[0128] in, and They represent The two sets of channel features after splitting. The purpose of channel splitting is to reduce the computational cost of convolution, allowing some channels to enter the convolution branch for transformation, while the other channels retain the original information or participate in lightweight fusion.

[0129] (2) Depthwise Conv

[0130] In the main branch, depthwise convolution is used to perform spatial convolution operations independently on each channel. Compared with standard convolution, depthwise convolution does not directly mix different channels, thus significantly reducing the number of parameters and computational cost. If the stride of this layer is 1, it is mainly used for local pattern extraction; if the stride is 2, downsampling can be achieved simultaneously.

[0131] For Lite-CES Block 1, the main branch primarily performs local texture extraction, which can be represented as:

[0132] ;

[0133] (3) Pointwise Conv

[0134] Since depthwise convolution only performs spatial convolution within a channel, the information exchange between channels is weak. Therefore, it is subsequently introduced with... Pointwise convolution for channel fusion:

[0135] ;

[0136] Pointwise convolution can recombine the features of each channel, improving the network's discriminative ability.

[0137] (4) Channel Shuffle

[0138] After channel splitting and grouping, information isolation issues may exist between different channel groups. To enhance cross-channel information exchange, a Channel Shuffle operation is introduced after concatenating the main branch and the retained branch:

[0139] ;

[0140] This operation rearranges the channels, enabling subsequent convolutions to perceive information from different channel groups, thereby improving feature utilization efficiency.

[0141] As can be seen from the above technical solutions, the main functions of Lite-CES Block 1 include: preserving shallow texture and local structural information; performing channel interaction under low overhead conditions; and providing more stable and richer intermediate representations for subsequent high-level feature learning.

[0142] 4. Lite-CES Block 2 (Stride=2)

[0143] The role of Lite-CES Block 2 is to perform spatial downsampling while further extracting deep semantic features. Since this module is set to stride=2, it not only undertakes the task of feature enhancement, but also reduces the feature map size, expands the receptive field, and further compresses the computational cost.

[0144] The key difference between Lite-CES Block 2 and Lite-CES Block 1 is that, in order to ensure that the feature map sizes of the two branches are consistent before splicing, the side branches should not use pure Identity, but should use Light Skip Connection, such as average pooling or lightweight convolutional downsampling.

[0145] (1) Main branch: downsampling feature extraction

[0146] The main branch still uses depthwise convolution + pointwise convolution for feature transformation. Let the input be... Then the main branch output can be expressed as:

[0147] ;

[0148] Among them, the stride=2 of the depthwise convolution achieves spatial downsampling; , and They represent The two sets of channel features after splitting.

[0149] (2) Light Skip Connection

[0150] To maintain the same spatial resolution as the main branch output of Lite-CES Block 2, the bypass branch uses lightweight downsampling mapping, such as: Average Pooling; Depthwise Conv with stride=2; or Pooling + 1×1 Conv. Its output is denoted as:

[0151] ;

[0152] in, This represents a lightweight downsampling map.

[0153] (3) Splicing and channel rearrangement

[0154] The features of the two branches are then concatenated, and cross-group information exchange is enhanced through Channel Shuffle:

[0155] .

[0156] As can be seen from the above technical solution, the Lite-CES Block 2 used in this invention has the following advantages: it reduces spatial resolution, improving subsequent computational efficiency; it expands the network's receptive field, enabling the model to acquire a wider range of structural information; it retains more effective features through dual-branch fusion, reducing information loss caused by downsampling; and it utilizes ChannelShuffle to enhance the multi-channel feature mixing capability. This module allows the network to gradually transition from local texture representation to more discriminative high-level semantic features while maintaining lightweight design.

[0157] 5. Global Average Pooling and Classifier Design

[0158] (1) Global Average Pooling (GAP)

[0159] After passing through two levels of Lite-CES Blocks, the network obtains high-level semantic feature maps. To reduce the number of parameters and avoid the strong dependence of traditional fully connected layers on spatial dimensions, this invention employs global average pooling to compress the space of each channel:

[0160] ;

[0161] in, Indicates the first The global mean of each channel. and The feature map space size, High-level feature map No. passage, spatial location The pixel value at that location.

[0162] (2) Fully Connected Classifier

[0163] The one-dimensional feature vector output by GAP is fed into the fully connected layer to classify six types of running states (normal, inner race fault, outer race fault 1, outer race fault 2, outer race fault 3, rolling element fault). The output can be represented as:

[0164] ;

[0165] in, The predicted scores are for the six operational status categories, and g is the global mean of all channels. The resulting feature vectors are optimized using a combination of Softmax and cross-entropy loss during the training phase.

[0166] Based on the above, the overall topology of the model of this invention is simple and easy to build. It uses the Lite-CES Block as its core, employing operations such as depthwise convolution, pointwise convolution, channel splitting, and channel rearrangement to effectively reduce model complexity while maintaining feature representation capabilities. Simultaneously, the model combines a preprocessing module, global average pooling, and a fully connected classifier. Based on this, the resulting model can efficiently identify six categories of operating states, providing an effective structural foundation for subsequent lightweight deployment and real-time fault diagnosis. Specifically: Firstly, shallow convolution helps preserve edge and texture information, and the subsequent Lite-CESBlock can gradually expand the receptive field and learn higher-level discriminative features. Secondly, traditional standard convolution performs spatial convolution and channel mixing simultaneously, resulting in high computational cost. This invention, however, employs a "Depthwise Conv + Pointwise Conv" decomposition in the Lite-CES Block, separating spatial modeling from channel fusion, thus significantly reducing the number of parameters and FLOPs. Simultaneously, at the input of each Lite-CES Block, the feature map is divided into two groups along the channel dimension: a main branch and a reserved branch. Convolution transformation is performed only on the main branch, while the reserved branch is directly passed to subsequent fusion. This strategy reduces the number of channels participating in convolution operations, further reducing computational overhead, while preserving the original information of the reserved branch helps mitigate feature loss. Thirdly, in lightweight grouped structures, channels easily form information silos. This invention uses Channel Shuffle, which adds almost no extra overhead but effectively promotes cross-channel feature interaction, improving the expressive power of the lightweight model. Fourthly, by using the Hardswish function, both expressive power and computational efficiency can be balanced, making it more suitable for embedded or edge deployments compared to complex activation functions. Fifthly, this invention uses GAP to replace large-scale fully connected layers. By using GAP, two-dimensional features are directly compressed into channel descriptions, avoiding parameter explosion caused by high-dimensional flattening, which is more conducive to controlling model complexity.

[0167] 6. Experimental Results

[0168] This experiment used publicly available bearing data from Case Western Reserve University as the pump and motor fault dataset. Bearing fault data at 12kHz was selected. The test bearings in the Case Western Reserve University experimental setup were used to support the motor shaft. Single-point faults were introduced into the test bearings using electrical discharge machining (EDM), including inner ring faults, rolling element faults, outer ring fault 1, outer ring fault 2, and outer ring fault 3. The outer ring faults are fixed faults; therefore, their position relative to the bearing's load-bearing area directly affects the vibration response of the motor / bearing system. To quantify this effect, experiments were conducted on the fan-end and drive-end bearings. The outer ring faults were positioned at the 3 o'clock position (directly opposite the load-bearing area), the 6 o'clock position (perpendicular to the load-bearing area), and the 12 o'clock position, respectively.

[0169] When setting the operating status categories in this paper, six states were selected: inner race fault, outer race fault 1, outer race fault 2, outer race fault 3, rolling element fault, and healthy state. Under the same operating conditions, based on the Ultra-Light LSR-Net deep learning model, each fault type was classified and statistically analyzed. The data details are shown in Table 1 below:

[0170] Table 1

[0171]

[0172] The dataset consists of grayscale images of six fault classes, with 100 samples in each class. 80% of these are used as the training set, and 20% as the test set. The model is trained using the training set, and then the test set is imported into the trained model. The loss function and accuracy curves during the training process are shown below. Figure 6 The convergence at 1 (100%) indicates that the trained model is well-trained and has achieved the desired effect. The analysis of various recognition rates on the test set is as follows: Figure 7 As shown, precision, recall, and F1 score are all 1, indicating good recognition performance. The test set confusion matrix is ​​as follows. Figure 8 As shown, the recognition effect is good, reaching 100% accuracy.

[0173] To verify the key role of the Lite-CES-Block model, this invention designed and conducted ablation experiments. In the original complete network, the Lite-CES-Block2 module performs spatial downsampling of the feature maps, which helps to expand the receptive field and extract multi-scale features. In the ablation experiments, we completely removed this module and uniformly replaced the positions where Lite-CES-Block2 was used with Lite-CES-Block1, which does not include downsampling operations, to construct an ablation comparison model.

[0174] The experimental results are as follows, and the recognition rate analysis of various types on the test set is as follows. Figure 9 As shown, the accuracy of identifying two types of faults is less than 90%, and the confusion matrix of the test set is as follows. Figure 10 As shown, after removing the entire Lite-CES-Block2 module, the model's best recognition rate on the test set dropped to 95%. Compared to the baseline model containing the complete module, the recognition rate showed a significant decrease. This result fully demonstrates the importance of the Lite-CES-Block2 module in the feature extraction process: the spatial downsampling capability provided by this module can effectively aggregate local information and expand the network's receptive field, thereby significantly improving the model's ability to represent and classify fault features. Removing this module leads to a lack of multi-scale feature fusion in the network, thus impairing the overall recognition performance.

[0175] In real-world scenarios such as industrial fault diagnosis, noise is inevitable in the data acquisition process (e.g., sensor accuracy limitations, environmental electromagnetic interference). To demonstrate the model's stability under "non-ideal data," Gaussian noise was added to the dataset in this experiment. However, real-world data is subject to noise interference, necessitating further verification of the deep learning model's noise resistance. By adding noise of varying intensities, the model's performance was tested. Compared to the initial model's 100% accuracy, the noise experiment more comprehensively reflects the method's engineering value. The table below shows the accuracy changes with different noise intensities. Typically, 0.02 represents mild noise. The deep learning model's recognition accuracy only began to decline after the noise intensity exceeded 1, demonstrating the strong noise resistance of the Ultra-Light LSR-Net deep learning model proposed in this invention.

[0176] Table 2

[0177]

[0178] III. Intelligent Agent Module:

[0179] The intelligent agent module specifically comprises the following steps: First, it completes industrial document parsing and vector database construction; then, using the constructed vector database as a knowledge source, it employs RAG and data distillation techniques to build a high-quality domain fine-tuning dataset. This high-quality domain fine-tuning dataset is then fine-tuned and trained using LoRA to form a large-scale intelligent pump diagnostic model, providing the core inference engine for the intelligent agent module. After encapsulation, the intelligent agent module is obtained. Finally, the intelligent agent module is embedded into the BS interaction module. Based on the pump rotor operating status characteristics, rolling bearing operating status categories, and vector database retrieval results, the intelligent agent module generates interpretable pump fault diagnosis conclusions and maintenance suggestions.

[0180] The entire technical approach unfolds in the sequence of "vector knowledge base construction—domain model fine-tuning and training—agent embedding," using the processing results of each stage as input for the next, ultimately forming a closed loop for intelligent operation and maintenance of pumps and machinery, comprising equipment status perception, domain knowledge retrieval, intelligent reasoning, maintenance decision-making, and human-machine interaction. The specific process can be summarized as follows:

[0181] (1) Construction of RAGFlow knowledge engine and vector database

[0182] To provide the agent module with a reliable source of knowledge in the pump and motor domain, a local knowledge engine was first built on the RAGFlow platform, and a pump and motor domain vector database was then established on this basis.

[0183] The data sources for the vector database include pump and motor maintenance manuals, equipment manuals, historical fault cases, expert experience, safe operating procedures, inspection records, and relevant professional literature. After importing these diverse and heterogeneous industrial documents into the RAGFlow platform, its deep document understanding capabilities are used to parse the document content.

[0184] For scanned documents, optical character recognition (OCR) technology is used to extract text information. For complex industrial documents containing titles, tables, images, and chapter structures, layout analysis methods are used to identify title levels, table structures, image descriptions, and chapter relationships. Through these processes, structured extraction of document content such as pump and motor fault comparison tables, equipment parameter specifications, maintenance flowcharts, and safety operating procedures is achieved, reducing the context loss and semantic fragmentation caused by traditional fixed-length text slicing.

[0185] After document parsing is completed, the parsing results are cleaned and intelligently segmented. While preserving the contextual relationships such as fault phenomena, fault causes, diagnostic basis, and maintenance measures, duplicate content and irrelevant information are removed to form several semantically complete knowledge fragments.

[0186] Subsequently, an embedding model (exemplarily Qwen3-Embedding-8B) is invoked to convert each knowledge fragment into a high-dimensional vector, and the high-dimensional vector, along with the corresponding original text, document source, and chapter position, are stored in a vector database. Simultaneously, a hybrid retrieval mechanism combining keyword retrieval and vector similarity retrieval is employed to provide traceable knowledge support for subsequent domain fine-tuning data construction and agent reasoning.

[0187] (2) Fine-tuning of the large-scale intelligent diagnostic model for pumps and motors

[0188] First, using the constructed vector database as the knowledge source, a high-performance general-purpose large model (exemplarily GPT-4) is invoked as the teacher model. Relevant knowledge fragments are obtained through Retrieval Augmentation (RAG), and knowledge distillation is performed. The system automatically generates chain of thought data containing the complete reasoning process of "fault phenomenon - fault cause - diagnostic basis - maintenance measures," transforming expert experience and domain knowledge into a high-quality fine-tuned dataset for the pump and motor domain.

[0189] During the fine-tuning process, a high-quality fine-tuning dataset in the field of pumps and machinery is input into the base model (for example, it can be the Qwen2.5-7B large language model). By adjusting the low-rank adaptation parameters in the model, the model learns the structure of pump and machinery equipment, professional terminology, fault mechanisms, diagnostic rules, maintenance specifications and safe operating procedures.

[0190] Compared to full-parameter fine-tuning, LoRA fine-tuning only requires updating a small number of additional parameters. It can reduce GPU memory usage and training costs while retaining the original language understanding and text generation capabilities of the base model, and enhance the model's professional reasoning capabilities in pump and turbine fault diagnosis and maintenance decision-making scenarios.

[0191] After fine-tuning, a comprehensive intelligent diagnostic model for pumps and motors is obtained. This model can comprehensively analyze the operating status of pumps and motors based on the characteristics of the pump rotor's operating status, the categories of rolling bearing operating status, and the results of vector database retrieval, and generate interpretable pump and motor fault diagnosis conclusions and maintenance suggestions.

[0192] (3) Embedding of intelligent agent module

[0193] After obtaining the large-scale intelligent diagnostic model for pumps and motors, a pump and motor fault diagnosis intelligent agent module was encapsulated. The B / S interaction module was embedded through the API interface, realizing the interoperability and integrated interface between the dialogue system and the pump and motor 3D visualization, equipment monitoring and operation logs.

[0194] Set up an intelligent interactive window on the web platform. When a user initiates a pump fault diagnosis request, the current pump rotor operating status characteristics and the rolling bearing operating status category output by the deep learning module are sent to the intelligent agent dialog box. The intelligent agent module will then sequentially complete knowledge retrieval, status analysis, and comprehensive reasoning to generate interpretable pump fault diagnosis conclusions and maintenance suggestions.

[0195] IV. BS Interaction Module

[0196] As the "face" of the system, this module provides a convenient and intuitive web-based interactive experience, enabling access from multiple devices. It mainly includes the following:

[0197] 3D visualization model: Load a high-fidelity 3D digital twin model of the pump in the browser, allowing users to rotate, zoom, and section the pump to view the internal structure of the pump, such as the impeller, rolling bearing, and seals.

[0198] Equipment monitoring: Establish a spatial mapping relationship between sensor modules and 3D digital twin models, and dynamically map the operating status categories output by the deep learning module to the corresponding positions in the model in the form of color changes, etc.

[0199] Operation Log: The extracted rotor operation status features and the operation status categories output by the deep learning module are displayed in real time in the sidebar of the interface.

[0200] Intelligent Agent Dialog Box: An embedded intelligent agent dialog box allows users to input rotor operating status characteristics and recognition results from the deep learning module. The intelligent agent module then generates interpretable pump and machine fault diagnosis conclusions and maintenance suggestions.

[0201] Diagnostic Analysis: Provides professional charting tools such as spectrum analysis graphs and waveform graphs for senior engineers to perform secondary confirmation and in-depth analysis. Supports historical data query and comparison to help analyze fault evolution trends and optimize preventive maintenance strategies.

[0202] This system, through the close collaboration of the four modules mentioned above, forms a complete closed loop of "perception (data acquisition) → cognition (deep learning + intelligent agent) → expression (BS visualization) → action (O&M suggestions) → feedback (human analysis / data feedback)". Based on this system, predictive maintenance can be carried out efficiently, improving equipment reliability, accurately identifying early faults, avoiding unplanned downtime, improving O&M efficiency, achieving automated diagnosis and intelligent suggestions, shortening troubleshooting time (MTTR), optimizing the O&M personnel experience, and making complex data simple and easy to understand based on zero-threshold natural language interaction and immersive 3D visualization.

[0203] The operation of the BS interaction module can be performed as follows:

[0204] (1) Login

[0205] The initialization login interface of the BS interaction module is as follows: Figure 11 As shown, you can only log in to the system after entering the correct username and password. The login interface retains the intelligent agent dialog box, which makes it convenient for users without permission to use the pump and motor fault diagnosis intelligent agent.

[0206] (2) Visual monitoring interface

[0207] After logging in, as follows Figure 12 As shown, the central area presents a 3D digital twin model of the equipment, which can be rotated and zoomed in to observe its internal structure. The left side displays a list of equipment. The right side shows a real-time data monitoring interface, displaying vibration data characteristic values ​​and trends. The right side also includes the connection status with the data acquisition and processing module, as well as the operation log. The operation log displays the system's operation records and the equipment's operating status, including rotor operating status characteristics and deep learning diagnostic conclusions, such as... Figures 13-15 .

[0208] (3) Analysis and diagnosis

[0209] The analysis and diagnostic module includes vibration trends, waveforms, and spectrum data, facilitating secondary verification by professional engineers and improving the accuracy of equipment diagnostics. Figure 16 , Figure 17 .

[0210] (4) Intelligent agent

[0211] like Figure 18 , 19 By clicking the pump fault diagnosis intelligence agent icon, the intelligence agent dialog box can be triggered. The equipment operating status can be copied and pasted from the operation log with one click. The pump fault intelligence agent can complete reasoning analysis and give pump fault diagnosis conclusions and operation and maintenance suggestions. Combined with human confirmation, it was found that the conclusions and suggestions given by the intelligence agent are consistent with human judgment and fully meet the operation and maintenance requirements.

[0212] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A pump intelligent operation and maintenance system integrating deep learning and fine-tuning of a large model, characterized in that, include: Data acquisition and processing module, deep learning module, intelligent agent module, and B / S interaction module; The data acquisition and processing module is used to filter and reduce noise in the acquired vibration acceleration signal of the pump to obtain a filtered vibration acceleration signal. The filtered vibration acceleration signal is then integrated into a vibration velocity signal using a frequency domain integration algorithm to extract the rotor operating state features of the pump. The data acquisition and processing module is also used to transmit the filtered vibration acceleration signal to the deep learning module. The fault diagnosis model trained in the deep learning module outputs the operating state category of the pump rolling bearing. Based on the network transmission protocol, the rotor operating state features and the rolling bearing operating state category are pushed to the BS interaction module for use by the intelligent agent module embedded in the BS interaction module.

2. The intelligent operation and maintenance system for pumps and turbines integrating deep learning and fine-tuning large models according to claim 1, characterized in that, The training of the fault diagnosis model in the deep learning module is specifically as follows: the vibration acceleration signals in the acquired pump fault dataset are preprocessed and labeled; the labeled vibration acceleration filtered signals are transformed into images to construct a training set; the training set is input into the Ultra-Light LSR-Net deep learning model for training to obtain the trained fault diagnosis model.

3. The intelligent operation and maintenance system for pumps and turbines integrating deep learning and fine-tuning large models according to claim 2, characterized in that, The image conversion operation specifically involves: for images of length... The mapping relationship between the vibration acceleration filter signal and the converted two-dimensional grayscale image is expressed by the following formula: ； in, This represents the pixel brightness value at the corresponding location in a two-dimensional grayscale image. ; Indicates the rounding operation; The amplitude of the vibration acceleration filter signal; and Signals The maximum and minimum values.

4. The intelligent operation and maintenance system for pumps and turbines integrating deep learning and fine-tuning large models according to claim 2, characterized in that, The Ultra-Light LSR-Net deep learning model includes: The input layer is used to receive a two-dimensional grayscale image X obtained by image transformation operations; The preprocessing module is used to perform preliminary feature encoding on the original input, and the expression is: ； in, For output of the preprocessing module, This represents the Hardswish activation function, where BN(·) is the batch normalization operation. This represents a convolution operation with a kernel size of 3×3 and a stride of 2, where X represents a single-channel two-dimensional grayscale image; Lite-CES Block 1, which takes the output of the preprocessing module as input, is used to extract mid-to-shallow local texture features. ; Lite-CES Block 2, which takes the output of Lite-CES Block 1 as input, is used to extract high-level semantic features. ; A global average pooling layer, wherein the global average pooling layer uses high-level semantic features The input is used to extract the global mean and construct a one-dimensional feature vector; A fully connected classifier takes a one-dimensional feature vector as input and outputs the classification result of the running state category.

5. The intelligent operation and maintenance system for pumps and turbines that integrates deep learning and fine-tuning of a large model according to claim 4, characterized in that, The Lite-CES Block 1 specifically refers to: The output of the preprocessing module is divided into two parts along the channel dimension: ； in, and These represent the outputs of the preprocessing module. The two sets of channel features after splitting; right use And a depthwise convolution operation with a stride of 1 is obtained , represented as: ； Output of depthwise convolution operation use Pointwise convolution to obtain , represented as: ； Output of pointwise convolution and After stitching, a Channel Shuffle operation is introduced to obtain mid-to-shallow local texture features. , represented as: 。 6. The intelligent operation and maintenance system for pumps and turbines that integrates deep learning and fine-tuning of a large model according to claim 4, characterized in that, The Lite-CES Block 2 specifically refers to: Local texture features of the middle and shallow layers Divided into two parts along the channel dimension: ； in, and These represent local texture features in the middle and shallow layers, respectively. The two sets of channel features after splitting; right Use in sequence And depthwise convolution with stride of 2, Pointwise convolution operation to obtain , represented as: ； right Lightweight downsampling mapping is used to obtain , represented as: ； in, Indicates a lightweight downsampling map; The output after pointwise convolution and the output of lightweight downsampling mapping After concatenation, a Channel Shuffle operation is introduced to obtain high-level semantic features. , represented as: 。 7. The intelligent operation and maintenance system for pumps and turbines integrating deep learning and fine-tuning large models according to claim 1, characterized in that, The intelligent agent module specifically comprises: first, completing industrial document parsing and vector database construction; then, using the constructed vector database as a knowledge source, employing RAG and data distillation techniques to construct a high-quality domain fine-tuning dataset; fine-tuning the high-quality domain fine-tuning dataset using LoRA to form a large-scale intelligent diagnostic model for pumps and rigs, which is then encapsulated to obtain the intelligent agent module; finally, embedding the intelligent agent module into the BS interaction module, which generates interpretable pump and rig fault diagnosis conclusions and maintenance suggestions based on the pump rotor operating status characteristics, rolling bearing operating status categories, and vector database retrieval results.

Images