Large compressor fault prediction method, device and equipment

Abstract

The invention discloses a fault prediction method, device and equipment for a large compressor. The method comprises the following steps: acquiring multi-sensor data; preprocessing the acquired multi-sensor data, and performing two-dimensional sample construction; based on a Bi-LSTM network, time sequence feature extraction and data decoupling are carried out; based on a CNN (Convolutional Neural Network) and Transform structure, multi-mode feature extraction and fusion are carried out; establishing a multi-instance learning framework, and performing fault prediction on the acquired multi-sensor data; and carrying out model training and optimization, and carrying out model performance evaluation. According to the method, the decomposed multi-periodic sequence is converted into the two-dimensional tensor through the data decoupling technology, and the recognition and processing capacity of the model on the periodic mode is improved; by combining the dynamic characteristics of the time series data with the strong classification capability of multi-instance learning, the accuracy, universality and real-time performance of fault prediction are improved, the fault mode recognition capability is improved, and the prediction reliability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of compressor fault prediction, and in particular to a method, device and equipment for predicting faults of large compressors. Background Art

[0002] With the rapid development of industrial automation and intelligent manufacturing, large compressors play a vital role in industries such as energy, chemicals, and manufacturing. The efficient operation of compressors is directly related to production efficiency and product quality. However, as equipment operation time increases, frequent compressor failures not only reduce production efficiency but can also cause serious safety incidents and increase maintenance costs.

[0003] Traditional compressor fault prediction methods rely primarily on manual feature extraction based on expert experience. This is not only labor-intensive but also has low detection accuracy, making it difficult to meet the high-precision requirements of complex systems. With the rapid development of deep learning technology, deep learning methods have gradually become an important tool in the field of fault diagnosis due to their ability to automatically extract deep-level feature information. Among the many deep learning methods, convolutional neural networks (CNNs), due to their weight sharing and sparse connectivity, possess powerful feature extraction capabilities and generalization properties, significantly improving fault diagnosis accuracy and thus becoming widely used in engineering fault diagnosis.

[0004] In addition to CNNs, recurrent neural networks (RNNs) and their variants, long short-term memory (LSTM) and gated recurrent units (GRUs), are also widely used in fault prediction. RNNs are primarily designed to process interdependent time series data and are capable of preserving state information in hidden layers. However, due to limitations in their network structure, RNNs are not effective at processing long-term dependent sequences and are prone to vanishing or exploding gradients. As an improvement on RNNs, LSTMs, by introducing input, forget, and output gates, effectively control the flow of information, addressing the shortcomings of RNNs in long-term sequence prediction. Compared to LSTMs, GRUs have a simpler structure, consisting only of update and reset gates, reducing network complexity and computational effort while maintaining performance.

[0005] However, the above prior art still has the following defects: Lack of universality: The existing compressor fault prediction method design is not universal, and single-scale feature mining can easily lead to large errors in the prediction results.

[0006] Low prediction accuracy: Existing methods have deficiencies in the use of time series information and fail to fully exploit the temporal dynamic characteristics of the data, affecting the accuracy of the prediction.

[0007] Poor practicality: Existing fault prediction methods lack sufficient practicality in actual applications and are unable to meet the real-time monitoring and prediction needs in complex industrial environments.

[0008] Limited fault pattern recognition capabilities: Existing methods have poor recognition capabilities for complex fault patterns, making it difficult to accurately distinguish different types of faults, which reduces the reliability of predictions. Summary of the Invention

[0009] The main purpose of the present invention is to provide a large compressor fault prediction method and device to solve the technical problems of poor accuracy, practicality, universality and reliability of large compressor fault prediction methods in the existing technology.

[0010] To achieve the above objectives, the present invention provides a large compressor fault prediction method, which includes the following steps: S10, acquiring multi-sensor data; S20, preprocessing the acquired multi-sensor data and constructing a two-dimensional sample; S30, performing time series feature extraction and data decoupling based on a Bi-LSTM network; S40, performing multi-modal feature extraction and fusion based on a CNN and Transformer structure; S50, establishing a multi-instance learning framework, and performing fault prediction on the acquired multi-sensor data; S60, performing model training and optimization, and performing model performance evaluation.

[0011] Optionally, step S10 includes the following steps: S110, installing multiple types of sensors at key locations of a large compressor; S120, setting a data acquisition frequency to 100 times per second; S130, acquiring multi-sensor data.

[0012] Optionally, step S20 includes the following steps: S210, performing noise filtering on the acquired multi-sensor data based on a bandpass filter; S220, performing outlier detection and processing based on a Z-score method; S230, filling missing values ​​in the data based on a linear interpolation method; S240, normalizing each sensor data to a unified scale range based on a Min-Max normalization method; S250, segmenting and stacking the preprocessed multi-sensor data according to a set time period to construct a multi-instance data set containing multiple time periods.

[0013] Optionally, the step S250 includes the following steps: S2510, setting an appropriate time period length based on the periodic features extracted by Fourier transform; S2520, dividing the entire data set into multiple time period samples according to the set period length; S2530, using a sliding window method to generate continuous time period samples; S2540, constructing a two-dimensional tensor structure.

[0014] Optionally, step S30 includes the following steps: S310, constructing a three-layer Bi-LSTM network; S320, inputting multi-period multi-sensor two-dimensional samples into the Bi-LSTM network, and extracting time series features through forward and backward propagation; S330, based on the extracted time series features, decoupling multi-instance data into short-term features and long-term features; S340, reconstructing the decoupled short-term features and long-term features into a high-dimensional tensor.

[0015] Optionally, the step S40 includes the following steps: S410, extracting local patterns and detail information in short-term features based on a multi-layer convolutional neural network; S420, extracting global dependencies and patterns in long-term features based on the self-attention mechanism in the Transformer structure; S430, performing multi-modal feature fusion based on feature mapping technology and multi-head attention fusion mechanism.

[0016] Optionally, the step S50 includes the following steps: S510, establishing a multi-instance learning model and inputting the fused comprehensive features into the multi-instance learning model; S520, performing hybrid loss function design and domain adversarial training; S530, performing fault prediction on the acquired multi-sensor data based on the trained multi-instance learning model.

[0017] Optionally, step S60 includes the following steps: S610, dividing the collected multi-instance data set into a training set, a validation set and a test set; S620, initializing the parallel convolutional neural network and Transformer module; S630, based on a 5-fold cross-validation strategy, using the Adam optimizer to optimize the parameters; S640, based on a hybrid loss function, optimizing the classification performance and feature similarity of the model; S650, evaluating the model performance based on the confusion matrix, ROC curve and AUC value, accuracy and recall evaluation indicators. In addition, to achieve the above-mentioned purpose, an embodiment of the present application also provides a large compressor fault prediction device, which includes: a data acquisition module for acquiring multi-sensor data; a data preprocessing module for preprocessing the acquired multi-sensor data and performing two-dimensional sample construction; a data decoupling module for performing time series feature extraction and data decoupling; a feature extraction module for performing multi-mode feature extraction and fusion; a fault prediction module for establishing a multi-instance learning framework and performing fault prediction on the acquired multi-sensor data; and a model performance evaluation module for performing model training and optimization, and performing model performance evaluation.

[0018] In addition, to achieve the above-mentioned purpose, an embodiment of the present application also provides a computer-readable storage medium, which includes instructions. When the instructions are run on a computer, the computer executes the large compressor fault prediction method described in any embodiment of the present application.

[0019] In addition, to achieve the above-mentioned purpose, an embodiment of the present application also provides a computing device, which includes: at least one processor, a memory and an input and output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the large compressor fault prediction method described in any embodiment of the present application.

[0020] Compared with the prior art, the present invention has the following beneficial effects: The large compressor fault prediction method provided in the embodiment of the present application converts the decomposed multi-period sequence into a two-dimensional tensor through data decoupling technology, retains rich periodic information, and improves the model's recognition and processing capabilities of periodic patterns; extracts local patterns and detailed information in short-term features through multi-layer CNN, enhancing feature expression capabilities; extracts global dependencies and patterns in long-term features through the Transformer structure, improving the comprehensiveness and depth of features; by combining the dynamic features of time series data with the powerful classification capabilities of multi-instance learning, and introducing advanced feature extraction and fusion methods, it significantly improves the accuracy, universality and real-time performance of fault prediction, and improves the fault pattern recognition capability and the reliability of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flowchart of a large compressor fault prediction method provided in an embodiment of the present application; Figure 2 A structural block diagram of a large compressor fault prediction device provided in an embodiment of the present application; Figure 3 Another flow chart of the large compressor fault prediction device provided in an embodiment of the present application; Figure 4Another flow chart of the large compressor fault prediction device provided in an embodiment of the present application; Figure 5 Another flow chart of the large compressor fault prediction device provided in an embodiment of the present application; Figure 6 This is a structural block diagram of the data decoupling module provided in an embodiment of the present application; Figure 7 A schematic diagram of the structure of the medium provided in the embodiment of the present application; Figure 8 A schematic diagram of the structure of a computing device provided in an embodiment of the present application.

[0022] The realization of the objectives, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0023] It should be understood that the specific embodiments described herein are merely for explaining the present application and are not intended to limit the present application. Rather, these embodiments are provided to make the present disclosure more thorough and complete and to fully convey the scope of the present disclosure to those skilled in the art.

[0024] In order to solve the above technical problems, the present invention provides a method for predicting large compressor failures. Figure 1 、 Figure 3-Figure 5 As shown, the method may include the following steps: S10, acquiring multi-sensor data.

[0025] In an exemplary embodiment, step S10 may include the following steps: S110, installs various types of sensors in key locations of large compressors; S120, setting the data collection frequency to 100 times per second; S130, acquiring multi-sensor data.

[0026] Key components of large compressors include the air inlet, compression chamber, exhaust port, blades, and bearings. These sensors, including temperature, pressure, vibration, and flow sensors, collect multi-dimensional operational data in real time, comprehensively monitoring the compressor's operating status. The data collection frequency is set at 100 times per second to capture rapid, dynamic changes in equipment operation. All sensor data is synchronized using timestamps to ensure temporal consistency across different sensor data, providing a reliable foundation for subsequent data processing and analysis.

[0027] S20, preprocessing the acquired multi-sensor data and constructing a two-dimensional sample.

[0028] In an exemplary embodiment, step S20 may include the following steps: S210, performing noise filtering on the acquired multi-sensor data based on a bandpass filter; S220, outlier detection and processing based on the Z-score method; S230, filling missing values ​​in the data based on linear interpolation to ensure the continuity and integrity of the data; S240, based on the Min-Max normalization method, normalizes the sensor data to a unified scale range ([−1,1]), using the Min-Max normalization method to eliminate the dimensional differences between different sensor data; S250 , segmenting and stacking the pre-processed multi-sensor data according to a set time period to construct a multi-instance data set containing multiple time periods.

[0029] Specifically, noise filtering uses a bandpass filter (such as a Butterworth filter) to remove high-frequency and low-frequency noise in the data, retaining the effective signal of the equipment operation. The specific settings are a high-pass cutoff frequency of 0.1 Hz and a low-pass cutoff frequency of 50 Hz, thereby filtering out irrelevant frequency components.

[0030] Furthermore, when performing outlier detection and processing, outliers with a Z-score exceeding 3 are removed or corrected to ensure the authenticity and reliability of the data.

[0031] In an exemplary embodiment, step S250 may include the following steps: S2510, based on the periodic features extracted by Fourier transform, set an appropriate time period length, with 60 seconds as one period; S2520, dividing the entire data set into multiple time period samples according to the set period length; S2530 uses a sliding window approach to generate continuous time period samples. For example, the data from the 1st to the 60th second is used as the first sample, the data from the 2nd to the 61st second is used as the second sample, and so on, to generate continuous and overlapping time period samples. This approach ensures that each sample contains rich time step information, enhancing the model's ability to learn time series features. S2540, construct a two-dimensional tensor structure.

[0032] Specifically, each time period sample is composed of data from multiple sensors, forming a two-dimensional tensor. Rows represent different sensors (such as temperature, pressure, vibration, and flow sensors), and columns represent time steps (such as data points per second). This two-dimensional tensor structure not only preserves the local variation characteristics of the time series but also integrates comprehensive information from multiple sensors, providing multi-dimensional and multi-angle fault signatures, laying a solid foundation for subsequent feature extraction and classification.

[0033] S30, based on the Bi-LSTM (bidirectional long short-term memory) network, performs time series feature extraction and data decoupling.

[0034] In an exemplary embodiment, step S30 may include the following steps: S310, build a three-layer Bi-LSTM network; S320 inputs multi-period multi-sensor two-dimensional samples into the Bi-LSTM network and extracts time series features through forward and backward propagation. The network outputs a time series feature vector, which represents the state change of the device at each time step. S330, based on the extracted time series features, decoupling the multi-instance data into short-term features and long-term features; S340, reconstructs the decoupled short-term features and long-term features into a high-dimensional tensor.

[0035] The Bi-LSTM network contains 256 hidden units per layer. The bidirectional structure can simultaneously capture forward and backward time-dependent information, fully extracting the temporal features in the data.

[0036] In this embodiment, short-term features are output through the last layer of the Bi-LSTM network to obtain feature vectors that reflect the instantaneous state changes of the device in a short period of time. Long-term features are extracted from the middle layer output of the Bi-LSTM network through a global average pooling operation to reflect the long-term operating trend of the device.

[0037] In addition, reference Figure 6 The decoupled short-term and long-term features are reconstructed into a high-dimensional tensor for subsequent feature fusion and classification. Short-term and long-term features are concatenated to form a comprehensive feature vector. A fully connected layer converts the comprehensive feature vector into a 256-dimensional feature vector and reconstructs it into a 16x16 two-dimensional tensor. This high-dimensional tensor can simultaneously reflect both local and global characteristics of device operation, providing an effective data representation for subsequent feature fusion and classification.

[0038] S40, based on CNN (multi-layer convolutional neural network) and Transformer structure, performs multi-modal feature extraction and fusion.

[0039] In an exemplary embodiment, step S40 may include the following steps: S410, extracts local patterns and details from short-term features based on a multi-layer convolutional neural network; S420, based on the self-attention mechanism in the Transformer structure, extracts global dependencies and patterns in long-term features; S430 performs multimodal feature fusion based on feature mapping technology and multi-head attention fusion mechanism.

[0040] The specific process of step S410 is as follows: First, a three-layer CNN is constructed, with each layer using convolution kernels of different sizes (e.g., 3x3 and 5x5) to capture local features at different scales. Each convolution layer is followed by a ReLU activation function and a max pooling layer to progressively extract higher-level features. The high-dimensional tensor is then input into the CNN, where convolution operations extract detailed information from short-term features and generate short-term feature maps. Each convolution layer is followed by nonlinear activation and pooling operations to enhance feature representation and reduce feature dimensionality.

[0041] In step S420, a Transformer-based encoder module is constructed, comprising a multi-head self-attention layer and a feedforward neural network layer. Each layer contains eight attention heads to extract long-term dependencies from different perspectives. The high-dimensional tensor is input into the Transformer encoder, where the self-attention mechanism captures long-term dependencies and generates long-term feature representations. Each Transformer encoder layer is followed by a residual connection and layer normalization to enhance feature stability and expressiveness.

[0042] In addition, when performing multimodal feature fusion, the short-term features extracted by CNN and the long-term features extracted by Transformer are mapped to the same 256-dimensional feature space through fully connected layers (or 1x1 convolutional layers) to ensure the consistency and comparability of the features.

[0043] This embodiment uses a multi-head attention fusion mechanism to perform a high-level fusion of mapped short-term and long-term features. The mapped short-term features are used as queries, and the long-term features are input into the multi-head attention module as keys and values. The attention weights are calculated using the head-attention mechanism. Different attention heads can extract relationships between features from different perspectives, enhancing the comprehensive representation of features. The outputs of the multi-head attention are concatenated and linearly transformed to generate a comprehensive fault prediction feature representation.

[0044] S50, establishing a multi-instance learning framework and performing fault prediction on the acquired multi-sensor data.

[0045] In an exemplary embodiment, step S50 may include the following steps: S510: Establish a multi-instance learning model and input the integrated features into the multi-instance learning model. Through joint training at the instance level and the packet level, accurate classification and prediction of fault modes are achieved. This step mainly includes the following process: (1) Instance-level classifier: For the features of each time step, an independent fully connected classifier is constructed to output the failure probability.

[0046] (2) Packet-level aggregation: The classification results of all instances are pooled together to generate packet-level fault prediction results. Packet-level aggregation can integrate prediction information from multiple time steps and improve the accuracy of the overall prediction.

[0047] (3) Joint training mechanism: Through joint training of instance-level classification and packet-level aggregation, the classification performance of the model at different levels is optimized, and the comprehensive judgment ability of fault modes is improved; S520, designing hybrid loss functions and conducting domain adversarial training; S530 , based on the trained multi-instance learning model, fault prediction is performed on the acquired multi-sensor data.

[0048] This embodiment uses a hybrid loss function and domain adversarial training to improve the classification performance and generalization ability of the model.

[0049] Hybrid loss function design: The cross-entropy loss function is used to measure the classification error:

[0050] in, is the actual category label, is the model prediction probability, is the sample size, and other parameters are conventional parameters in this field.

[0051] Contrastive Loss: It is used to enhance the similarity between feature vectors, promote the aggregation of feature vectors of the same category and the separation of feature vectors of different categories.

[0052]

[0053] in, is the distance between eigenvectors, is the boundary, and other parameters are conventional parameters in this field.

[0054] Domain Adversarial Loss: By introducing a domain discriminator, the model's generalization ability in different operating environments is optimized.

[0055]

[0056] in, is the loss function of the domain discriminator, is the loss function.

[0057] The main operation process of Domain Adversarial Training is as follows: Domain discriminator design: Build an independent neural network to identify the domain to which the data belongs (such as different equipment, different working conditions).

[0058] Gradient Reversal Layer (GRL): A gradient reversal layer is introduced between the main model and the domain discriminator so that during backpropagation, the feature extraction part of the main model maximizes the classification error of the domain discriminator, thereby learning domain-invariant feature representations.

[0059] Adversarial training mechanism: The gradient of the main model is adjusted through the gradient reversal layer, which enables the model to learn features that are not affected by the domain and improves the model's generalization ability in different operating environments.

[0060] In step S530, the real-time multi-sensor data is segmented and preprocessed according to a set period and then fed into the time series feature extraction and data decoupling module. Time series features are extracted using a Bi-LSTM, combined with a CNN and Transformer to extract multimodal features. A comprehensive feature representation is generated using a multi-head attention fusion mechanism. This comprehensive feature representation is fed into a multi-instance learning model, which outputs the fault probability and category, generating a prediction result. The prediction result is displayed in real time on the monitoring interface, triggering an early warning mechanism and supporting preventive maintenance decisions, thereby reducing equipment downtime and lowering maintenance costs.

[0061] S60, perform model training and optimization, and evaluate model performance.

[0062] In an exemplary embodiment, step S60 may include the following steps: S610, dividing the collected multi-instance data set into a training set, a validation set, and a test set to ensure a balanced distribution of various types of fault samples; S620, initialize the parallel convolutional neural network and Transformer modules; S630, based on a 5-fold cross-validation strategy, uses the Adam optimizer for parameter optimization; S640, based on a hybrid loss function, optimizes the model’s classification performance and feature similarity; S650, performs model performance evaluation based on confusion matrix, ROC curve and AUC value, precision and recall evaluation indicators.

[0063] In this exemplary embodiment, the CNN module is designed as three convolutional layers, each using 3x3 and 5x5 convolution kernels, respectively, followed by ReLU activation functions and max pooling layers to extract local patterns and details from short-term features. The Transformer module consists of a six-layer Transformer encoder, each with eight attention heads, a feedforward neural network layer, residual connections, and layer normalization, designed to extract global dependencies and patterns from long-term features.

[0064] During model training, we employed a 5-fold cross-validation strategy and parameter optimization using the Adam optimizer. The initial learning rate was set to 0.001. We dynamically adjusted the learning rate based on the validation set performance, using a learning rate scheduler (such as ReduceLROnPlateau) to automatically reduce the learning rate when the validation set loss stopped decreasing, ensuring efficient and stable model training.

[0065] To optimize the model's classification performance and feature similarity, a hybrid loss function is used, consisting of a cross-entropy loss and a contrastive loss. The cross-entropy loss measures classification error, ensuring the model's accuracy on classification tasks; the contrastive loss enhances the similarity between feature vectors, promoting the aggregation of feature vectors of the same category and the separation of feature vectors of different categories. Furthermore, Domain Adversarial Training (Domain Adversarial Training) is introduced. By constructing a domain discriminator and employing a gradient reversal layer (GRL), the model's generalization capabilities across different operating environments are enhanced, ensuring the consistency and accuracy of prediction results.

[0066] To evaluate model performance, we use a variety of metrics, including confusion matrices, receiver operating characteristic (ROC) curves and area under curve (AUC), accuracy, and recall, to comprehensively measure the model's performance in fault prediction. Experiments were conducted on a validation set, recording the model's classification accuracy and AUC for different fault types. Based on these experimental results, model hyperparameters (such as convolution kernel size, number of Transformer layers, and learning rate) were adjusted to continuously optimize model performance. This series of verification and tuning processes ensures the model's efficiency and robustness in practical applications, enabling it to consistently provide accurate fault prediction results.

[0067] In practical applications, a multidimensional sensor dataset collected from industrial sites was used to compare the proposed large-scale compressor fault prediction technology based on a time series multi-instance learning framework with existing methods (such as SVM, traditional LSTM, CNN, Bi-LSTM, and Transformer). The experimental results, shown in Table 1, demonstrate that the proposed method outperforms existing technologies in terms of fault prediction accuracy, AUC, and real-time performance, particularly when processing complex time series data and multiple fault modes.

[0068] Table 1

[0069] The experimental results above demonstrate that the method provided by the present embodiment achieves RMSE and AUC values ​​of 4.8 and 0.96, respectively, on the test set, significantly outperforming other models. This demonstrates that the large-scale compressor fault prediction technology based on the time series multi-instance learning framework offers greater efficiency and accuracy in periodic feature extraction and multi-modal feature fusion. The method of the present invention performs particularly well when processing high-noise data and multiple fault modes, demonstrating strong robustness and generalization capabilities.

[0070] Based on the above embodiments, Figure 2 Another embodiment of the present application further provides a large compressor fault prediction device, the large compressor fault prediction device 200 comprising: A data acquisition module 210 is used to acquire multi-sensor data; The data preprocessing module 220 is used to preprocess the acquired multi-sensor data and construct a two-dimensional sample; Data decoupling module 230, used for time series feature extraction and data decoupling; Feature extraction module 240, for performing multi-modal feature extraction and fusion; The fault prediction module 250 is used to establish a multi-instance learning framework and perform fault prediction on the acquired multi-sensor data; The model performance evaluation module 260 is used to perform model training and optimization, and perform model performance evaluation.

[0071] Specifically, if Figure 6 Said data decoupling module 230 includes a periodic multi-dimensional decoupling module, a dual change modeling module and a change aggregation module; The periodic multidimensional decoupling module includes a periodicity extractor, a period-based reconstructor, and a time-varying decoupler; The dual-variation modeling module includes a long-term variation extractor and a short-term variation extractor; The change aggregation module includes a splicing unit and a linear processing unit.

[0072] Based on the above embodiments, the present application also provides a computer-readable storage medium. Figure 7 The computer-readable storage medium shown is a CD 50, which stores a computer program (i.e., a program product). When executed by a processor, the computer program implements the steps described in the above method implementation, including acquiring multi-sensor data; preprocessing the acquired multi-sensor data and constructing two-dimensional samples; extracting time series features and decoupling data using a Bi-LSTM network; extracting and fusing multi-modal features using CNN and Transformer architectures; establishing a multi-instance learning framework and performing fault prediction on the acquired multi-sensor data; training and optimizing the model, and evaluating model performance. The specific implementation of each step is not repeated here.

[0073] It should be noted that examples of the computer-readable storage medium may also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical or magnetic storage media, which are not listed here one by one.

[0074] In addition, based on the above embodiment, the embodiment of the present application also provides a computing device, Figure 8 A block diagram of an exemplary computing device 60 suitable for implementing embodiments of the present application is shown. The computing device 60 may be a computer system or a server. Figure 8 The computing device 60 shown is only an example and should not limit the functionality and scope of use of the embodiments of the present application.

[0075] like Figure 8 As shown, the components of computing device 60 may include, but are not limited to, one or more processors or processing units 601 , a system memory 602 , and a bus 603 connecting various system components (including system memory 602 and processing unit 601 ).

[0076] The computing device 60 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computing device 60, including volatile and non-volatile media, removable and non-removable media.

[0077] The system memory 602 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 6021 and / or cache memory 6022. The computing device 60 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the ROM 6023 may be used to read and write non-removable, non-volatile magnetic media ( Figure 8 is not shown in the , usually referred to as "hard drive"). Although not in Figure 8 As shown in FIG, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk"), and an optical drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) can be provided. In these cases, each drive can be connected to a bus 603 that connects various system components via one or more data media interfaces. System memory 602 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of various embodiments of the present application.

[0078] A program / utility 6025 having a set (at least one) of program modules 6024 may be stored, for example, in system memory 602. Such program modules 6024 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 6024 generally implement the functions and / or methods of the embodiments described herein.

[0079] The computing device 60 can also communicate with one or more external devices 604 (e.g., a keyboard, a pointing device, a display, etc.). Such communication can be performed through an input / output (I / O) interface 605. Furthermore, the computing device 60 can also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) through a network adapter 606. Figure 8 As shown, the network adapter 606 communicates with other modules of the computing device 60 (such as the processing unit 601) via the bus 603 that connects different system components. Figure 8 Not shown, other hardware and / or software modules may be used in conjunction with computing device 60 .

[0080] The processing unit 601 executes various functional applications and data processing by running programs stored in the system memory 602. These include acquiring multi-sensor data; preprocessing the acquired multi-sensor data and constructing two-dimensional samples; extracting time series features and decoupling data based on a Bi-LSTM network; extracting and fusing multi-modal features based on CNN and Transformer architectures; establishing a multi-instance learning framework and performing fault prediction on the acquired multi-sensor data; training and optimizing the model, and evaluating model performance. The specific implementation of each step will not be repeated here. It should be noted that while the detailed description above mentions several units / modules or sub-units / sub-modules of the large compressor fault prediction device, this division is merely exemplary and not mandatory. In practice, according to embodiments of the present application, the features and functions of two or more units / modules described above may be embodied in a single unit / module. Conversely, the features and functions of a single unit / module described above may be further divided and embodied by multiple units / modules.

[0081] In the description of this application, it should be noted that the terms "first", "second" and "third" are used for descriptive purposes only and should not be understood as indicating or implying relative importance.

[0082] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. There may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some communication interface, the indirect coupling or communication connection of the device or unit may be electrical, mechanical or other forms.

[0084] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0085] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0086] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0087] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed in the present application, or perform equivalent replacements for some of the technical features thereof. These modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the scope of protection of the claims.

[0088] Furthermore, although the operations of the method of the present application are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

Claims

1. A large compressor fault prediction method, characterized in that: The large compressor failure prediction method comprises the following steps: S10, acquiring multi-sensor data; S20, preprocessing the acquired multi-sensor data and constructing a two-dimensional sample; S30, based on Bi-LSTM network, performs time series feature extraction and data decoupling; S40, based on CNN and Transformer structures, performs multi-modal feature extraction and fusion; S50, establishing a multi-instance learning framework and performing fault prediction on the acquired multi-sensor data; S60, perform model training and optimization, and evaluate model performance.

2. The large compressor fault prediction method according to claim 1, characterized in that: Step S10 includes the following steps: S110, installs various types of sensors in key locations of large compressors; S120, setting the data collection frequency to 100 times per second; S130, acquiring multi-sensor data.

3. The large compressor failure prediction method according to claim 1, characterized in that: The step S20 includes the following steps: S210, performing noise filtering on the acquired multi-sensor data based on a bandpass filter; S220, outlier detection and processing based on the Z-score method; S230, filling missing values ​​in the data based on linear interpolation; S240, based on the Min-Max normalization method, normalizes the sensor data to a unified scale range; S250 , segmenting and stacking the pre-processed multi-sensor data according to a set time period to construct a multi-instance data set containing multiple time periods.

4. The large compressor failure prediction method according to claim 3, characterized in that: The step S250 includes the following steps: S2510, setting an appropriate time period length based on the periodic features extracted by Fourier transform; S2520, dividing the entire data set into multiple time period samples according to the set period length; S2530, using a sliding window method to generate continuous time period samples; S2540, construct a two-dimensional tensor structure.

5. The large compressor fault prediction method according to claim 1, characterized in that: The step S30 includes the following steps: S310, build a three-layer Bi-LSTM network; S320, inputs multi-period multi-sensor two-dimensional samples into the Bi-LSTM network and extracts time series features through forward and backward propagation; S330, based on the extracted time series features, decoupling the multi-instance data into short-term features and long-term features; S340, reconstructs the decoupled short-term features and long-term features into a high-dimensional tensor.

6. The large compressor fault prediction method according to claim 1, characterized in that: The step S40 includes the following steps: S410, extracts local patterns and details from short-term features based on a multi-layer convolutional neural network; S420, based on the self-attention mechanism in the Transformer structure, extracts global dependencies and patterns in long-term features; S430 performs multimodal feature fusion based on feature mapping technology and multi-head attention fusion mechanism.

7. The large compressor fault prediction method according to claim 1, characterized in that: The step S50 includes the following steps: S510, establishing a multi-instance learning model, and inputting the fused comprehensive features into the multi-instance learning model; S520, designing hybrid loss functions and conducting domain adversarial training; S530 , based on the trained multi-instance learning model, fault prediction is performed on the acquired multi-sensor data.

8. The large compressor fault prediction method according to claim 1, characterized in that: The step S60 includes the following steps: S610, dividing the collected multi-instance data set into a training set, a validation set, and a test set; S620, initialize the parallel convolutional neural network and Transformer modules; S630, based on a 5-fold cross-validation strategy, uses the Adam optimizer for parameter optimization; S640, based on a hybrid loss function, optimizes the model’s classification performance and feature similarity; S650, performs model performance evaluation based on confusion matrix, ROC curve and AUC value, precision and recall evaluation indicators.

9. A large compressor fault prediction device, characterized in that: include: Data acquisition module, used to acquire multi-sensor data; The data preprocessing module is used to preprocess the acquired multi-sensor data and construct two-dimensional samples; Data decoupling module, used for time series feature extraction and data decoupling; Feature extraction module, used for multi-modal feature extraction and fusion; Fault prediction module, used to establish a multi-instance learning framework and perform fault prediction on the acquired multi-sensor data; The model performance evaluation module is used for model training and optimization, as well as model performance evaluation.

10. A computing device, characterized in that The computing device comprises: at least one processor, memory, and input-output unit; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the large compressor fault prediction method according to any one of claims 1 to 8.

Images