A method and system for early warning of key equipment faults of thermal power generating units based on MultiPatchFormer

Abstract

The application discloses a thermal power unit key equipment fault early warning method and system based on MultiPatchFormer, and relates to the technical field of thermal power unit fault early warning. The method acquires thermal power unit equipment data and operation data and performs preprocessing; the preprocessed data is input into a MultiPatchFormer model to obtain parameter prediction values of future time periods; an abnormal score is obtained by calculating the deviation of the prediction values and actual values, and if the score exceeds a failure threshold, an early warning is triggered. The MultiPatchFormer model is sequentially connected by a multi-scale patch embedding layer, a hierarchical double attention encoder and a multi-step linear decoder: the multi-scale patch embedding layer divides data into non-uniform patches at different step lengths and completes double position encoding; after the encoded patches are bidirectionally refined by the hierarchical double attention encoder, multi-step forward prediction is completed by the multi-step linear decoder adopting a pseudo self-recursive strategy. The bidirectional attention mechanism strengthens data correlation mining and improves parameter prediction accuracy, and provides reliable technical support for thermal power unit key equipment fault early warning.

Description

Technical Field

[0001] This invention belongs to the field of fault early warning technology for thermal power units, and relates to a fault early warning method and system for key equipment of thermal power units based on MultiPatchFormer. Background Technology

[0002] As the core energy supply infrastructure of the power industry, the stable operation of key equipment in thermal power units plays a crucial role in ensuring energy supply security and improving the reliability of the power system. Therefore, fault early warning of key equipment has always been a key topic in the field of reliability research in the power industry. With the widespread adoption of distributed control systems (DCS) and plant-level monitoring information systems (SIS) in thermal power units, the units continuously generate and accumulate massive amounts of multi-dimensional time-series data during operation. This data covers many key parameters such as temperature, pressure, flow rate, and vibration, providing a rich source of information for in-depth analysis of equipment operating status.

[0003] However, the traditionally widely used threshold alarm method has significant shortcomings in fault early warning. This method simply compares equipment parameters with preset thresholds, triggering an alarm once the parameters exceed the threshold range. This "one-size-fits-all" approach is too crude and fails to provide early warning for potential, slowly changing faults. Alarms are often only issued when the fault has developed to a relatively severe stage, resulting in delayed alarms and an inability to take timely and effective maintenance measures. Furthermore, the false alarm rate remains high, causing numerous problems for the normal operation and maintenance management of equipment. In recent years, artificial intelligence technologies such as deep learning have made significant progress in data processing and analysis, demonstrating powerful advantages in processing high-dimensional, nonlinear, and time-series data, bringing new development opportunities to thermal power unit fault early warning technology. By using deep learning algorithms, it is possible to automatically learn the normal operating status patterns and fault precursor characteristics of equipment from massive amounts of historical data, enabling earlier and more accurate predictive maintenance of equipment faults. This drives the evolution of fault early warning technology from traditional rule-based and threshold-based methods to an end-to-end learning model.

[0004] Currently, deep learning algorithms such as RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory Network), CNN (Convolutional Neural Network), and AE (Autoencoder) have been applied to some extent in the field of fault early warning for thermal power units. However, these algorithms have also revealed some limitations in practical applications. High-frequency noise is common in industrial data, which can interfere with the true characteristics of the data. At the same time, the data also has multi-scale periodic patterns and strong coupling characteristics, that is, different parameters are interconnected and influence each other. Existing deep learning algorithms are unable to effectively handle these complex characteristics simultaneously, which to some extent limits the accuracy and reliability of fault early warning. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem that existing fault early warning systems do not consider the strong coupling of industrial data, resulting in inaccurate early warnings, and to provide a fault early warning method and system for key equipment of thermal power units based on MultiPatchFormer.

[0006] To achieve the above objectives, the present invention employs the following technical solution: The first aspect of this invention provides a method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer, comprising the following steps: Acquire equipment data and operating data of thermal power units, and preprocess the operating data of thermal power unit equipment to obtain preprocessed data; The preprocessed data is input into the MultiPatchFormer model, which outputs the predicted values ​​of each parameter for future time periods. Based on the deviation between the predicted value and the corresponding actual value of each parameter, the corresponding anomaly score is calculated; if the anomaly score exceeds the failure threshold, a fault warning is triggered. The MultiPatchFormer model comprises a multi-scale patch embedding layer, a hierarchical dual attention encoder, and a multi-step linear decoder connected in sequence. The multi-scale patch embedding layer divides the preprocessed data into several non-uniform patches with different lengths and performs dual position encoding on each non-uniform patch. The non-uniform patches after dual position encoding enter the hierarchical dual attention encoder for bidirectional feature refinement and then enter the multi-step linear decoder for decoding, outputting the predicted values ​​of each parameter for future time periods. The multi-step linear decoder uses a pseudo-autoregressive strategy for multi-step forward prediction.

[0007] Furthermore, the dual positional encoding is described as follows:

[0008] in, This represents a non-uniform patch after dual positional encoding; Indicates absolute position embedding; Indicates time interval embedding; express Encoder.

[0009] Furthermore, the hierarchical dual attention encoder includes a temporal encoder and a channel encoder; the temporal encoder uses a multi-head self-attention mechanism to calculate temporal attention weights and generate a temporal causal relationship graph; the temporal causal relationship graph is used in the channel encoder to generate feature maps through global average pooling and channel-level self-attention mechanisms.

[0010] Furthermore, the formula for calculating the temporal attention weights in the multi-head self-attention mechanism is as follows:

[0011] in, Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the linear projection dimension; Indicates the number of attention points; Indicates the activation function; This represents the temporal attention weights.

[0012] Furthermore, the formula for calculating the attention weights of the channel-level self-attention mechanism is as follows:

[0013] in, Indicates global average pooling. Indicates channel-level attention weights; Indicates the activation function; Represents the hyperbolic tangent function; , , All of these represent weighting coefficients.

[0014] Furthermore, the predicted values ​​of each parameter for the future time period are described as follows:

[0015] in, Represents a linear activation function; This represents the feature map after bidirectional feature refinement in the hierarchical dual attention encoder. Indicates the preceding The predicted value at any given time.

[0016] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the MultiPatchFormer-based method for early warning of faults in key equipment of thermal power units.

[0017] A third aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the MultiPatchFormer-based method for early warning of faults in key equipment of thermal power units.

[0018] A fourth aspect of the present invention provides a computer program product, the computer program product including computer instructions, the computer instructions instructing a computer to execute the MultiPatchFormer-based method for early warning of faults in key equipment of thermal power units.

[0019] The fifth aspect of this invention provides a fault early warning system for key equipment in thermal power units based on MultiPatchFormer, comprising: The data acquisition module acquires equipment data and operating data of the thermal power unit, and preprocesses the operating data of the thermal power unit to obtain preprocessed data. The prediction module inputs the preprocessed data into the MultiPatchFormer model and outputs the predicted values ​​of each parameter for future time periods. The fault warning module calculates the corresponding anomaly score based on the deviation between the predicted value and the corresponding actual value of each parameter; if the anomaly score exceeds the failure threshold, a fault warning is triggered. The MultiPatchFormer model comprises a multi-scale patch embedding layer, a hierarchical dual attention encoder, and a multi-step linear decoder connected in sequence. The multi-scale patch embedding layer divides the preprocessed data into several non-uniform patches with different lengths and performs dual position encoding on each non-uniform patch. The non-uniform patches after dual position encoding enter the hierarchical dual attention encoder for bidirectional feature refinement and then enter the multi-step linear decoder for decoding, outputting the predicted values ​​of each parameter for future time periods. The multi-step linear decoder uses a pseudo-autoregressive strategy for multi-step forward prediction.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a fault early warning method for key equipment in thermal power units based on MultiPatchFormer. Through a multi-scale patch embedding layer, it uses different time-series data to perform non-uniform segmentation, simultaneously capturing short-term fluctuations and long-term trends, adapting to the inherent multi-time-scale variation characteristics of thermal power unit data. A hierarchical dual-attention encoder combines temporal attention and channel attention mechanisms to model causal relationships and parameter coupling relationships in the time dimension, respectively. Temporal attention can identify fault propagation paths such as "increased vibration → increased temperature," while channel attention focuses on the mutual influence between key parameters (such as vibration, temperature, and pressure), achieving deep feature extraction from strongly coupled industrial data. The multi-step linear decoder employs a pseudo-autoregressive strategy, generating multi-step prediction values ​​in stages. Each step combines historical predictions with encoded features, effectively suppressing the problem of error amplification at each stage in traditional autoregressive methods, and improving the stability and accuracy of long-term predictions. A dynamic failure threshold is constructed using the Exponentially Weighted Moving Average (EWMA) method. The warning boundary is adaptively adjusted based on historical anomaly scores to reduce false alarms caused by data mutations or noise, while improving the detection sensitivity for slowly changing faults. After an anomaly is triggered, the system can locate the fault based on parameter correlations, identify the equipment component to which the abnormal parameters belong (such as turbine bearings, steam pipes, etc.), and record the anomaly time, score, and trend. This provides data support for subsequent fault diagnosis and maintenance, realizing integrated fault management from early warning to location. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of the fault early warning method for key equipment of thermal power units based on MultiPatchFormer of the present invention. Figure 2 This is an example of the early warning effect of a certain fault case in a thermal power unit according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0026] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1 This invention provides a method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer, which can significantly improve the accuracy of fault warnings for thermal power units, achieve early warning as early as possible, and ensure the safe and stable operation of thermal power units. Specifically, the detailed steps of the method of this invention are as follows: S1, Data Preprocessing. First, monitoring data of the thermal power generating units is collected, including equipment data and operational data of the units; then, the collected equipment data and operational data of the thermal power generating units are preprocessed.

[0027] Specifically, monitoring data from thermal power generating units is inherently prone to quality issues due to sensor noise, communication interruptions, and transient operational fluctuations. These issues include missing values, outlier interference, and redundant features. These problems directly degrade the performance of MultiPatchFormer because incomplete sequences disrupt patch embedding, and outliers distort the calculation of attention weights in the encoder. This embodiment uses LightGBM to clean the monitoring data from thermal power generating units, followed by normalization.

[0028] S2, train the MultiPatchFormer model and calculate the threshold.

[0029] S201, Training the MultiPatchFormer model. This invention designs the MultiPatchFormer to capture the multi-scale temporal variations and cross-parameter dependencies commonly found in monitoring data of thermal power generating units. MultiPatchFormer comprises three basic modules: a multi-scale patch embedding layer, a hierarchical dual-attention encoder, and a multi-step linear decoder. These components work together to build an end-to-end framework capable of enabling early detection of potential faults.

[0030] First, the multi-scale patch embedding layer in the MultiPatchFormer model transforms the original multivariate time series data—that is, the equipment and operational data of thermal power units—into a structured feature representation while preserving the temporal granularity. For the original multivariate time series... (in It is the sequence length. (This refers to the number of parameters). This embodiment uses overlapping patch segmentation with different step sizes, specifically including the following steps: (1) Adaptive patch partitioning: The input sequence uses a step size The data is divided into non-uniform patches to capture short-term fluctuations (e.g., 8 steps) to long-term trends (e.g., 32 steps). Each patch... go through Linear projection of convolution onto dimension .

[0031] (2) Perform dual position encoding on the above non-uniform patch, specifically, using absolute position embedding. Embedded with relative time interval The fusion process preserves the chronological order; the specific formula is as follows: (1) in, This represents a non-uniform patch after dual-position encoding. This hybrid encoding strategy can reduce information loss caused by downsampling.

[0032] Secondly, the hierarchical dual attention encoder in the MultiPatchFormer model consists of stacked temporal encoder and channel encoder modules, enabling bidirectional feature refinement, specifically: (1) Temporal Encoder: Employs a multi-head self-attention mechanism (MSA) to calculate temporal attention weights and generate a temporal causal relationship graph. For example, calculating the... The time step and the first Temporal attention weights at each time step are used to identify the causal relationship of "abnormal vibration → temperature rise". (Includes...) Size, including time step and The temporal attention weights are calculated as follows: (2) in, Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the linear projection dimension; Indicates the number of attention points; Indicates the activation function; This represents the temporal attention weights.

[0033] The temporal encoder is followed by a feedforward network (FFN) with a GELU activation function to enhance the nonlinearity of local features.

[0034] (2) Channel Encoder: Captures the correlation between parameters through global average pooling and channel-level self-attention mechanisms. For feature maps (in (This refers to the number of image tiles). The attention weights for each key parameter in the feature map are calculated as follows: (3) in, This indicates global average pooling, while This indicates an emphasis on key parameters such as turbine vibration and boiler temperature.

[0035] This dual attention mechanism, combined with layer normalization and residual connections, stabilizes the training process. Each encoder block updates features as follows: (4) Finally, the multi-step linear decoder in the MultiPatchFormer model employs a pseudo-autoregressive strategy to generate reliable long-term prediction results: (1) Phased prediction: future sequences in It is generated in the first step, where the first step is... Step input the previously predicted value and coding features To predict .

[0036] (2) Noise Suppression: By decomposing the prediction task into sub-problems, it reduces error accumulation compared to single-step decoding. The final output is in the form of: (5) In this embodiment, the input sequence of MultiPatchFormer is determined by a step size. The model is divided into non-uniform blocks, enabling it to capture changes from short-term fluctuations (8 steps) to long-term trends (32 steps). The MultiPatchFormer model is trained for 150 epochs with a learning rate of 0.0005. During training, the deviation between the predicted and actual values ​​of all device input parameters is calculated. Then, an exponentially weighted moving average (EWMA) method is used to determine the upper and lower thresholds for each parameter.

[0037] EWMA is an effective tool for smoothing abrupt changes in data. Using EWMA to construct a statistical score set is as follows: (6) in, Indicates a time index. It is the historical score in the The weight given to the set of times when making the estimate. Belongs to the interval In this embodiment, It was set to 0.3; Represents the anomaly score at time t; initial value The average of historical scores; This indicates the deviation between the predicted value and the actual value.

[0038] The upper and lower limits of the EWMA chart (which are also the failure thresholds) are calculated as follows: (7) in, and These represent the mean and standard deviation of the scores in the training dataset, respectively. It is a constant value, and its size is 1.25 times the maximum anomaly score of the training set.

[0039] S3, test data is input into the model trained in step S2 for testing.

[0040] The proposed model is tested using test data to verify its effectiveness. Specifically, the preprocessed test data from step S1 is input into the MultiPatchFormer model trained in step S2 to obtain the predicted sequence of the test data. The prediction accuracy and generalization ability of the model are verified by calculating the deviation between the actual values ​​of each monitored parameter in the test data and the model's predicted values.

[0041] S4, Fault Location For real-time acquired and preprocessed monitoring data of thermal power units, the predicted values ​​of each parameter are calculated using a trained MultiPatchFormer model, and then the deviation between the predicted and actual values ​​is calculated. Then, substitute the values ​​into formula (6) to calculate the real-time anomaly score. .

[0042] Real-time anomaly score Compare with the upper and lower warning thresholds calculated by formula (7): If > or < If the corresponding parameter is found to be abnormal, a fault warning signal will be triggered; if exist[ , If the condition is within the specified range, the equipment is considered to be operating normally.

[0043] Fault localization employs an anomaly parameter tracing method: Statistical analysis of the parameters triggering early warning signals is performed, and combined with the equipment structure and parameter correlations of the thermal power unit, the equipment component to which the anomaly parameter belongs is determined. For example, if both "turbine bearing vibration" and "turbine bearing temperature" parameters trigger an early warning simultaneously, the fault is determined to occur in the turbine bearing component; if only the "main steam pressure" parameter triggers an early warning, the main steam pipeline and pressure regulating device are the primary focus of the investigation. Simultaneously, the time of the anomaly, the anomaly score, and the parameter change trend are recorded to provide data support for subsequent fault diagnosis and maintenance.

[0044] See Figure 2 , Figure 2 This paper presents a comparison of the early warning effects of turbine vibration parameters in a thermal power unit. In the figure, the solid line represents the actual monitored values, the dashed line represents the MultiPatchFormer predicted values, and the shaded area represents the EWMA dynamic threshold range. Through field case testing, the model successfully issued an early warning for non-drive-end bearing failures 7 hours in advance, verifying the model's effectiveness in early fault detection and location.

[0045] One embodiment of the present invention provides a fault early warning system for key equipment of thermal power units based on MultiPatchFormer, comprising: The data acquisition module acquires equipment data and operating data of the thermal power unit, and preprocesses the operating data of the thermal power unit to obtain preprocessed data. The prediction module inputs the preprocessed data into the MultiPatchFormer model and outputs the predicted values ​​of each parameter for future time periods. The fault warning module calculates the corresponding anomaly score based on the deviation between the predicted value and the corresponding actual value of each parameter; if the anomaly score exceeds the failure threshold, a fault warning is triggered. The MultiPatchFormer model comprises a multi-scale patch embedding layer, a hierarchical dual attention encoder, and a multi-step linear decoder connected in sequence. The multi-scale patch embedding layer divides the preprocessed data into several non-uniform patches with different lengths and performs dual position encoding on each non-uniform patch. The non-uniform patches after dual position encoding enter the hierarchical dual attention encoder for bidirectional feature refinement and then enter the multi-step linear decoder for decoding, outputting the predicted values ​​of each parameter for future time periods. The multi-step linear decoder uses a pseudo-autoregressive strategy for multi-step forward prediction.

[0046] In one embodiment of the present invention, an electronic device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a fault early warning method for key equipment in thermal power units based on MultiPatchFormer.

[0047] One embodiment of the present invention provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that more specific examples (a non-exhaustive list) of the computer-readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.

[0048] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0049] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0050] One or more instructions stored in a computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer in the above embodiments.

[0051] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer, characterized in that, Includes the following steps: Acquire equipment data and operating data of thermal power units, and preprocess the operating data of thermal power unit equipment to obtain preprocessed data; The preprocessed data is input into the MultiPatchFormer model, which outputs the predicted values ​​of each parameter for future time periods. Based on the deviation between the predicted value and the corresponding actual value of each parameter, the corresponding anomaly score is calculated; If the abnormal score exceeds the failure threshold, a fault warning is triggered; The MultiPatchFormer model comprises a multi-scale patch embedding layer, a hierarchical dual attention encoder, and a multi-step linear decoder connected in sequence. The multi-scale patch embedding layer divides the preprocessed data into several non-uniform patches with different lengths and performs dual position encoding on each non-uniform patch. The non-uniform patches after dual position encoding enter the hierarchical dual attention encoder for bidirectional feature refinement and then enter the multi-step linear decoder for decoding, outputting the predicted values ​​of each parameter for future time periods. The multi-step linear decoder uses a pseudo-autoregressive strategy for multi-step forward prediction.

2. The method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer according to claim 1, characterized in that, The dual positional encoding is described as follows: in, This indicates a non-uniform patch after dual positional encoding; Indicates absolute position embedding; Indicates time interval embedding; express Encoder.

3. The method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer according to claim 1, characterized in that, The hierarchical dual attention encoder includes a timing encoder and a channel encoder; The timing encoder uses a multi-head self-attention mechanism to calculate the timing attention weights and generate a timing causal relationship graph. The temporal causal relationship graph is generated in the channel encoder through global average pooling and channel-level self-attention mechanism to generate feature maps.

4. The method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer according to claim 3, characterized in that, The formula for calculating the temporal attention weights in the multi-head self-attention mechanism is as follows: in, Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the linear projection dimension; Indicates the number of attention points; Indicates the activation function; This represents the temporal attention weights.

5. The method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer according to claim 3, characterized in that, The formula for calculating the attention weights of the channel-level self-attention mechanism is as follows: in, Indicates global average pooling. Indicates channel-level attention weights; Indicates the activation function; Represents the hyperbolic tangent function; , , All of these represent weighting coefficients.

6. The method for early warning of faults in key equipment of thermal power units based on MultiPatchFormer according to claim 1, characterized in that, The predicted values ​​of each parameter for the future time period are described as follows: in, Represents a linear activation function; This represents the feature map after bidirectional feature refinement in the hierarchical dual attention encoder. Indicates the preceding The predicted value at any given time.

7. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the MultiPatchFormer-based method for early warning of faults in key equipment of thermal power units as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the MultiPatchFormer-based method for early warning of faults in key equipment of thermal power units as described in any one of claims 1-6.

9. A computer program product, the computer program product comprising computer instructions, characterized in that, The computer instructions instruct the computer to execute the MultiPatchFormer-based fault early warning method for key equipment of thermal power units as described in any one of claims 1-6.

10. A fault early warning system for key equipment in thermal power units based on MultiPatchFormer, characterized in that, include: The data acquisition module acquires equipment data and operating data of the thermal power unit, and preprocesses the operating data of the thermal power unit to obtain preprocessed data. The prediction module inputs the preprocessed data into the MultiPatchFormer model and outputs the predicted values ​​of each parameter for future time periods. The fault warning module calculates the corresponding anomaly score based on the deviation between the predicted value and the corresponding actual value of each parameter. If the abnormal score exceeds the failure threshold, a fault warning is triggered; The MultiPatchFormer model comprises a multi-scale patch embedding layer, a hierarchical dual attention encoder, and a multi-step linear decoder connected in sequence. The multi-scale patch embedding layer divides the preprocessed data into several non-uniform patches with different lengths and performs dual position encoding on each non-uniform patch. The non-uniform patches after dual position encoding enter the hierarchical dual attention encoder for bidirectional feature refinement and then enter the multi-step linear decoder for decoding, outputting the predicted values ​​of each parameter for future time periods. The multi-step linear decoder uses a pseudo-autoregressive strategy for multi-step forward prediction.

Images