Heat dissipation system fault detection method based on self-attention mechanism

Abstract

The invention provides a heat dissipation system fault detection method based on a self-attention mechanism, and belongs to the field of fault detection, and the method comprises the following steps: defining a fault type through a heat dissipation system fault type definition module, detecting the fault type of the matched fan by monitoring the change of the fan current effective value, the THD value and the phase difference parameter, and importing the parameter data corresponding to the defined fault type into the historical database; the method comprises the following steps: collecting data associated with fault detection through a fault data set acquisition module, acquiring three-phase current from a control panel, storing waveform data in a register, performing data processing on a current signal, and calculating to obtain a current effective value; performing FFT analysis on the current to obtain a THD value and a phase; performing FFT frequency domain analysis to obtain a current THD and a phase; and comparing the processed data with historical data, calculating a distortion size X, and comparing the distortion size X with a parameter diagnosis index to obtain a preliminary diagnosis result.

Description

Technical Field

[0001] The present invention belongs to the field of fault detection, and in particular relates to a heat dissipation system fault detection method based on a self-attention mechanism. Background Art

[0002] The cooling system plays a vital role in modern electronic and industrial equipment, ensuring that the equipment operates within the normal operating temperature range, thereby improving performance and extending service life. However, as the complexity and power density of equipment increase, fault detection of the cooling system becomes increasingly important. Traditional fault detection techniques often rely on simple temperature sensors or manual inspections, which have multiple defects and are difficult to adapt to the needs of modern equipment. Traditional fault detection methods usually lack real-time performance and rely on regular manual inspections and maintenance, which cannot detect potential faults in the system in a timely manner. This delay may cause the equipment to fail to take timely measures when a fault occurs, resulting in serious equipment damage or downtime, resulting in a decrease in production efficiency. Although simple temperature sensors can provide basic temperature data, it is difficult to fully reflect the actual working status of the cooling system. Abnormal temperature does not necessarily mean that the cooling system fails, but may be caused by other factors. Therefore, single temperature monitoring cannot provide sufficient fault warning information.

[0003] Existing fault detection methods are usually unable to handle complex multi-dimensional data, especially in large-scale systems where the cooling system may be affected by multiple factors, such as ambient temperature, load changes, and material aging. Traditional detection methods often focus on a single indicator and ignore the relationship and influence between data, resulting in insufficient accuracy and reliability of fault diagnosis. In addition, existing technologies often lack intelligence and automation capabilities. Many detection systems still rely on manual experience for fault diagnosis, are easily affected by human subjective factors, and have difficulty handling complex fault modes.

[0004] Existing fault detection systems generally have the problem of insufficient ability to identify abnormal patterns. In cooling systems, faults may manifest themselves in a variety of ways, and a single monitoring method is difficult to fully capture these anomalies, resulting in missed detections or false detections. Traditional methods also often lack data-driven analysis capabilities and are unable to use big data and machine learning technologies to mine potential fault modes and patterns. Many traditional fault detection methods perform poorly in terms of adaptability and flexibility, making them difficult to effectively deploy in different application scenarios, limiting the possibility of their widespread use.

[0005] Therefore, developing a cooling system fault detection method based on the self-attention mechanism has become an urgent need in the industry. The self-attention mechanism can effectively process complex data relationships and capture the mutual influence between multi-dimensional data, thereby improving the accuracy and real-time performance of fault detection. By utilizing deep learning and the self-attention mechanism, this method can learn the normal working mode of the cooling system from historical data, identify potential abnormal behaviors in a timely manner, and provide more comprehensive fault warnings. This new method can not only improve the intelligence level of fault detection, but also achieve more efficient and flexible cooling system management, thereby improving the reliability and performance of the equipment. Summary of the invention

[0006] The present invention proposes a heat dissipation system fault detection method based on a self-attention mechanism, which provides a solution to the limitations of real-time fault handling methods. Faults can be discovered in advance, and corresponding maintenance measures can be taken to avoid production losses caused by faults.

[0007] The technical solution of the present invention is implemented as follows: a method for detecting heat dissipation system faults based on a self-attention mechanism, the method comprising the following steps: The fault type is defined through the cooling system fault type definition module, and the fault type of the matching fan is detected by monitoring the changes in the fan current effective value, THD value, and phase difference parameters, and the parameter data corresponding to the defined fault type is imported into the historical database; The fault data set acquisition module collects data related to fault detection, collects three-phase current from the control board and stores the waveform data in the register, processes the current signal, and calculates the effective value of the current; performs FFT analysis on the current to obtain the THD value and phase; and performs FFT frequency domain analysis to obtain the current THD and phase; Compare the processed data with the historical data, calculate the distortion size X, compare the distortion size X with the parameter diagnosis index, and obtain the preliminary diagnosis result; and use the distortion size, current effective value, phase, current THD, and temperature as feature inputs, and use the feature encoder to obtain feature self-attention embedding, that is, the representation of the input features in a high-dimensional space; Then, the feature self-attention embedding is used as the input of the decoder and as the initial hidden state of the decoder for autoregressive prediction. The real-time operation data of the cooling system is collected, including the effective value of the current, phase, current THD, and temperature as feature inputs. At the same time, the trained self-attention neural network model is loaded, and the features are input into the self-attention neural network model. The output results are calculated through forward propagation to obtain the prediction of the cooling system fault type.

[0008] Compared with the existing technology, this method of heat dissipation system fault detection based on self-attention mechanism has several significant innovations and advantages. First, this method accurately defines the fault type through the heat dissipation system fault type definition module, and combines the effective value of the fan current, total harmonic distortion (THD) value and phase difference parameter changes to perform fault matching. The advantage of this method is that it can analyze specific fault types, while traditional fault detection methods usually rely on single parameter monitoring and fail to fully consider the relationship between multiple parameters, resulting in reduced accuracy of fault identification. Secondly, the concept of historical database is introduced in the method. By importing the parameter data corresponding to the defined fault type into the historical database, it can not only achieve long-term tracking of fault modes, but also provide data support for subsequent fault warning and maintenance decisions. This data-driven approach is relatively rare in traditional systems, and existing technologies often rely on experience or static standards and lack dynamic updating and learning capabilities.

[0009] The design of the fault data set acquisition module enables the system to collect three-phase current from the control board and store the waveform data in the register for processing. The frequency domain analysis of the current signal is achieved by performing fast Fourier transform (FFT) analysis on the effective value, THD value and phase of the current signal. This process is relatively lacking in the existing technology. Traditional methods usually rely only on simple monitoring in the time domain and cannot fully capture the frequency domain characteristics of the current signal, resulting in insufficient fault identification capabilities.

[0010] The method compares the processed data with the historical data and calculates the distortion size X to obtain a preliminary diagnosis result. This process makes fault detection more accurate and comprehensive through comparative analysis and the relationship between distortion characteristics and parameter diagnostic indicators. Traditional methods often lack this dynamic comparative analysis, resulting in the inability to identify potential fault risks in a timely manner.

[0011] The introduction of feature self-attention embedding is a highlight of this method. By using feature encoders to obtain feature self-attention embedding, the method can represent input features in a high-dimensional space and capture the complex relationships between different features. This self-attention mechanism is not common in traditional fault detection methods. Existing technologies usually rely on shallow feature extraction and are difficult to effectively capture complex patterns and information.

[0012] The use of decoders and the introduction of autoregressive predictions enable the system to predict future failure modes based on historical data. This deep learning-based prediction capability is rarely used in existing technologies. Traditional methods usually rely on simple linear models or empirical formulas and lack the ability to proactively analyze future failures.

[0013] By collecting real-time operation data of the cooling system and loading the trained self-attention neural network model, this method can accurately predict the fault type of the cooling system. This combination of dynamic monitoring and intelligent analysis effectively improves the system's response speed and accuracy to faults. However, existing technologies often lack intelligent features in this regard and cannot achieve real-time and accurate fault prediction. Therefore, this cooling system fault detection method based on the self-attention mechanism not only improves the accuracy and efficiency of fault detection, but also provides a new solution for the intelligent maintenance of modern industrial equipment. This method is significantly superior to traditional technologies in terms of comprehensiveness, accuracy and intelligence of fault detection, and represents the development direction of future fault detection technology.

[0014] At present, most of the cooling system fault detection technologies mainly use traditional signal processing methods and machine learning methods. Traditional signal processing methods usually include FFT (Fast Fourier Transform) analysis, wavelet transform, time-frequency analysis and other technologies. These methods can detect faults from the aspects of wind pressure, mechanical vibration, current, etc. of the cooling system. Machine learning methods mainly classify and identify the working status of the motor by establishing a neural network model for fault detection. Although these methods perform well in real-time fault detection, they have some shortcomings in fault prediction. Traditional methods can usually only handle specific types of faults, and require a lot of feature engineering and prior knowledge, which is not easy to expand and migrate. In addition, traditional methods often cannot fully consider the complex relationship between data, and it is difficult to deal with nonlinear problems. To solve the above problems, this study considers training a self-attention neural network model for motor fault detection. For the above problems, this method.

[0015] Defining multiple cooling system fault types can speed up error troubleshooting and better meet the needs of actual engineering projects. Collect data related to fault detection and build a cooling system fault data set.

[0016] The self-attention-based neural network model is more adaptable and versatile. It can learn the dependencies between monitoring data and automatically discover patterns and rules in the data, so as to better capture changes and anomalies in the motor's operating state. At the same time, the original data can be directly modeled to avoid information loss, making the model more accurate and more robust.

[0017] As a preferred implementation, the fault types include fan dirty blockage faults and fan general faults. Before defining the fault type, the fan is preliminarily detected to obtain the power of the fan when it is working, and the fan fault state is determined by setting a monitoring current threshold.

[0018] As a preferred implementation, the power formula of the fan when working is as follows: in Q:Air volume(m 3 / h)、 P :Wind pressure, η 1 Machine efficiency, η 2 Mechanical transmission efficiency.

[0019] As a preferred embodiment, the feature encoder is a coding model, and the model setting method includes the following steps: Get the input feature x=[x1,x2…x4], perform data cleaning and normalization on the feature sequence to obtain the feature sequence, and then divide the data set into training, validation and test sets with a division ratio of 7:2:1; Absolute position encoding is performed on the feature sequence to obtain a feature vector sequence; The feature vector sequence is input into the Transformer encoder for encoding. The encoder consists of multiple encoders, each of which consists of two sublayers, one is a multi-head self-attention sublayer, and the other is a feedforward neural network sublayer. Each sublayer has a residual connection and layer normalization operation. The normalization layer uses LayerNorm layer normalization operation; The self-attention layer uses the self-attention mechanism to perform attention calculations with the encoding results as the query matrix, keyword matrix, and value matrix.

[0020] As a preferred implementation, the attention calculation adopts the following formula: in, Q represents the query matrix, K represents the keyword matrix, K T represents the transpose of a matrix, d k represents the dimension of the input feature vector, V represents the value matrix, Attention( Q, K, V ) represents the final output self-attention score, and softmax represents the normalized exponential function.

[0021] As a preferred implementation, the decoder performs autoregressive prediction, and the specific steps are as follows: Encode the encoder output as a Transformer decoder: Build a decoder that adds a multi-head self-attention layer and a feed-forward neural network layer to the encoder; Get the output before the decoder, perform linear transformation and scaling through matrix multiplication and scaling factors, and convert the original vector to the new vector space; The first input to the decoder is a special symbol <sos>, indicating the beginning of the sequence, the decoder starts from <sos>Start generating the output sequence and use the symbols generated in the previous time step as input for the next time step; At each time step, the decoder embeds the symbol generated by the previous time step into a continuous vector space through an embedding layer. The decoder connects the representation output by the encoder and the embedding vector of the previous time step, inputs them into the multi-head self-attention layer of the decoder, and calculates the attention score of the current time step. By weighting the encoder output representation, the decoder captures the information associated with the current time step in the input sequence. The decoder concatenates the weighted representation with the embedding vector to obtain the probability distribution of the output symbol of the current time step, which is used to predict the symbol of the next time step, and obtains the predicted output sequence y1, y2...yT; The predicted output sequence is sent to a softmax layer for fault classification prediction. By optimizing the cross entropy loss function and using the Adam optimizer to optimize the network parameters, the model convergence is accelerated.

[0022] After adopting the above technical solution, the beneficial effects of the present invention are: Accuracy: The self-attention mechanism can learn the dependencies between the various monitoring data in the sequence, so as to better capture the subtle changes and abnormal conditions in the operation of the cooling system. Compared with the traditional characteristic equation, the method based on the self-attention neural network can more accurately capture the complex patterns in the data, thereby achieving higher accuracy.

[0023] Interpretability: The self-attention mechanism of this method can directly show the relationship between each monitoring data in the sequence, so as to better explain the prediction results of the model. This helps engineers understand the decision-making process of the model and can help engineers determine the maintenance measures for specific faults.

[0024] Adaptability: This method can automatically learn the dependencies between the elements in the sequence and adaptively adjust the model parameters according to different data features. This makes the method more adaptive and robust and can adapt to different data and environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0026] Figure 1 This is a diagram showing the relationship between the air volume and air pressure of the fan in the present invention; Figure 2 This is a flow chart of the fault diagnosis algorithm of the present invention; Figure 3 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Example like Figure 1 As shown, a method for detecting heat dissipation system faults based on a self-attention mechanism comprises the following steps: The fault type is defined through the cooling system fault type definition module, and the fault type of the matching fan is detected by monitoring the changes in the fan current effective value, THD value, and phase difference parameters, and the parameter data corresponding to the defined fault type is imported into the historical database; The fault data set acquisition module collects data related to fault detection, collects three-phase current from the control board and stores the waveform data in the register, processes the current signal, and calculates the effective value of the current; performs FFT analysis on the current to obtain the THD value and phase; and performs FFT frequency domain analysis to obtain the current THD and phase; Compare the processed data with the historical data, calculate the distortion size X, compare the distortion size X with the parameter diagnosis index, and obtain the preliminary diagnosis result; and use the distortion size, current effective value, phase, current THD, and temperature as feature inputs, and use the feature encoder to obtain feature self-attention embedding, that is, the representation of the input features in a high-dimensional space; Then, the feature self-attention embedding is used as the input of the decoder and as the initial hidden state of the decoder for autoregressive prediction. The real-time operation data of the cooling system is collected, including the effective value of the current, phase, current THD, and temperature as feature inputs. At the same time, the trained self-attention neural network model is loaded, and the features are input into the self-attention neural network model. The output results are calculated through forward propagation to obtain the prediction of the cooling system fault type.

[0029] This application document adopts a cooling system fault detection method based on a self-attention neural network model, which mainly includes a cooling system fault type definition module, a fault data set acquisition module, a self-attention neural network training model, and a fault type prediction. The present invention uses a self-attention neural network model module to learn the cooling system data features and their corresponding fault types, firstly performs feature encoder processing through input features, then calculates the attention score through the self-attention mechanism, and finally embeds the feature self-attention as the input of the decoder for autoregressive prediction, and performs fault classification prediction through the softmax layer. The use of a self-attention neural network model can fully consider the dependency between data features and improve the accuracy of cooling system fault detection.

[0030] The fan in this application document is a key device for the heat dissipation system of rail trains, such as high voltage, traction, and air conditioning. It is an important guarantee for safe operation. Common problems of the heat dissipation system during operation include: fan dirtiness and blockage, and fan failure. Through the corresponding detection technology, the working parameters that can represent the fault trend can be measured. By conducting in-depth analysis of these parameters, the fault type of the AC traction motor of the electric locomotive can be determined.

[0031] Fan blockage problem: When air cooling is the main heat dissipation method, the filter blockage of the cooling fan will directly affect the heat dissipation efficiency, thereby causing failures in various systems of the EMU and affecting normal operation. However, since most of the fans, traction motors and other equipment are installed at the bottom of the train, the filter is easily affected by dust, catkins and other debris, causing blockage. The current filter cleaning is mainly based on manual periodic inspection, which is inefficient. Therefore, how to effectively judge the filter blockage and formulate a reasonable cleaning cycle is of great significance to ensure the normal operation of the EMU and improve the maintenance efficiency.

[0032] Fan failure problem: The fan is ultimately a motor device, and the biggest feature of the motor is its high symmetry. All forms of failure affect its symmetry, which leads to corresponding changes in the magnetic flux between the stator and the rotor, and then causes changes in the stator current, voltage and magnetic field, resulting in abnormal mechanical vibration. When a motor fails, it often produces characteristic components of specific frequencies in the signal spectrum, thereby affecting the THD (total harmonic distortion) value of the current. In addition, the voltage and current values ​​will also change due to the influence of the fault. Therefore, the fan failure can be detected by monitoring the changes in parameters such as the effective value of the fan current, THD value, phase difference, etc. For a fan, the power formula when it is working is: in Q :Air volume(m 3 / h)、 P :Wind pressure, η 1 Machine efficiency, η 2 Mechanical transmission efficiency, where η 1 , η 2 It is only related to the fan itself and has nothing to do with changes in the external environment. Therefore, the power change of the fan is only positively correlated with the air volume and pressure.

[0033] like Figure 1 As shown in the figure, it is a relationship diagram between the air volume and air pressure of a certain type of centrifugal cooling fan. It can be seen that when the fan filter is blocked, the air volume during fan operation decreases and the air pressure increases, but the product of the air volume and air pressure decreases, that is, the power of the fan decreases. Therefore, the positive correlation between the fan work and the fan current can be used to judge the dirty blockage of the fan, that is, as the degree of dirty blockage increases, the fan work decreases and the fan current decreases. As long as a suitable threshold is set, when the fan current is lower than the threshold, the dirty blockage of the fan can be determined. In addition, since dirty blockage leads to a decrease in air volume and a decrease in heat dissipation efficiency, the dirty blockage can also be judged by monitoring the rate of temperature rise and temperature drop of the heat dissipation object to assist the current, thereby improving the accuracy of diagnosis.

[0034] Whether it is a fan blockage problem or a fault problem, the most important thing is to monitor the change of fan current. In order to prevent mutual misjudgment between the two, it is necessary to use the differences between the two. The blockage process is a process that requires time to accumulate and change slowly, but the failure of the fan is an accidental and sudden problem, so the two types of problems can be effectively distinguished based on this feature. For example, the effective value of the monitored current is recorded. When the current value is lower than the set threshold, the current values ​​of the previous few times are compared and analyzed. If there is a sudden change in the current, it means that the fan has a fault. However, if it is found that the current has shown a slow downward trend before reaching the threshold, it means that the current change is caused by blockage.

[0035] The fault data set acquisition module first needs to collect data related to fault detection. The fault diagnosis algorithm step flow chart is as follows: Figure 2 shown.

[0036] The three-phase current (IA, IB, IC) is collected from the control board and the waveform data is stored in the register, and the temperature data is directly obtained from the network.

[0037] Perform data processing on the current signal. Calculate and obtain the effective value of the current (calculate multiple times to get the average value); perform FFT analysis on the current to obtain the THD value and phase; perform FFT frequency domain analysis to obtain the current THD and phase, which can reflect the operating status of the fan. Once a sudden change in data occurs, it can be determined that the fan is faulty. Here, for the fan fault problem, only the fault judgment function is realized and maintenance suggestions are given, but the fault location cannot be achieved.

[0038] Compare the processed data with the historical data (normal operation), calculate the distortion size X (based on the percentage of historical data), and compare the distortion size X with the parameter diagnosis indicators shown in Table 1 to obtain preliminary diagnosis results. Different distortion parameters correspond to different state types;

[0039] Self-attention neural network model training Using distortion size (percentage), current RMS, phase, current THD, and temperature as feature inputs, a feature encoder is used to obtain feature self-attention embedding, that is, the representation of input features in a high-dimensional space. The feature encoder is an encoding model that includes the following steps: Obtain the input feature x=[x1,x2…x4], first perform data cleaning and normalization on the feature sequence to obtain the feature sequence, and then divide the data set into training, validation and test sets with a division ratio of 7:2:1.

[0040] Absolute position encoding is performed on the feature sequence to obtain a feature vector sequence; The feature vector sequence is input into the Transformer encoder for encoding. The encoder consists of multiple encoders, each of which consists of two sublayers, one is a multi-head self-attention sublayer and the other is a feedforward neural network sublayer. Each sublayer has a residual connection and a layer normalization operation. The normalization layer uses the LayerNorm layer normalization operation.

[0041] The self-attention layer uses the self-attention mechanism to calculate the attention using the encoding results as the query matrix, keyword matrix, and value matrix. The calculation formula is: in, Q represents the query matrix, K represents the keyword matrix, K T represents the transpose of a matrix, d k represents the dimension of the input feature vector, V represents the value matrix, Attention( Q, K, V ) represents the final output self-attention score, softmax represents the normalized exponential function, and the formula is as follows:

[0042] The feature self-attention embedding is then used as the input of the decoder and as the initial hidden state of the decoder for autoregressive prediction. The specific steps are as follows: Encode the encoder output as the Transformer decoder: First, build the decoder. The decoder adds a multi-head self-attention layer and a feedforward neural network layer on the basis of the encoder. Get the previous output of the decoder and perform linear transformation and scaling through a matrix multiplication and a scaling factor to convert the original vector to a new vector space. The first input of the decoder is a special symbol <sos>, indicating the start of the sequence. The decoder starts with <sos>Start generating the output sequence and take the symbols generated in the previous time step as input for the next time step.

[0043] At each time step, the decoder first embeds the symbol generated in the previous time step into a continuous vector space through an embedding layer. Then, the decoder connects the representation of the encoder output and the embedding vector of the previous time step, inputs them into the decoder's multi-head self-attention layer, calculates the attention score of the current time step, and by weighting the encoder output representation, the decoder can capture the information related to the current time step in the input sequence. The decoder concatenates the attention-weighted representation with the embedding vector to obtain the probability distribution of the output symbol of the current time step, which is used to predict the symbol of the next time step. Finally, the predicted output sequence y1, y2 … yT is obtained. The predicted output sequence is sent to a softmax layer for fault classification prediction. By optimizing the cross entropy loss function, the network parameters are optimized with the Adam optimizer to accelerate the convergence of the model. Predicting the type of motor fault Collect the real-time operation data of the cooling system, including the current effective value, phase, current THD, and temperature as feature inputs, and load the trained self-attention neural network model at the same time. Input the features into the self-attention neural network model, calculate the output results through forward propagation, and obtain the prediction of the cooling system fault type. This patent discloses a method for detecting heat dissipation system faults. Most current heat dissipation system fault detection methods are usually unable to handle environmental noise and interference well, and these factors will affect the accuracy of the model. At the same time, this method provides solutions to some limitations of real-time fault handling methods. For example, for equipment with a long downtime, their operating status cannot be monitored in real time. At the same time, if the fault can be discovered in advance, corresponding maintenance measures can be taken to avoid production losses caused by the fault. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.< / sos> < / sos> < / sos> < / sos>

Claims

1. A heat dissipation system fault detection method based on self-attention mechanism, characterized in that: The method comprises the following steps: The fault type is defined through the cooling system fault type definition module, and the fault type of the matching fan is detected by monitoring the changes in the fan current effective value, THD value, and phase difference parameters, and the parameter data corresponding to the defined fault type is imported into the historical database; The fault data set acquisition module collects data related to fault detection, collects three-phase current from the control board and stores the waveform data in the register, processes the current signal, and calculates and obtains the effective value of the current; Perform FFT analysis on the current to obtain the THD value and phase; FFT frequency domain analysis obtains current THD and phase; Compare the processed data with the historical data, calculate the distortion size X, compare the distortion size X with the parameter diagnosis index, and obtain the preliminary diagnosis result; and use the distortion size, current effective value, phase, current THD, and temperature as feature inputs, and use the feature encoder to obtain feature self-attention embedding, that is, the representation of the input features in a high-dimensional space; Then, the feature self-attention embedding is used as the input of the decoder and as the initial hidden state of the decoder for autoregressive prediction. The real-time operation data of the cooling system is collected, including the effective value of the current, phase, current THD, and temperature as feature inputs. At the same time, the trained self-attention neural network model is loaded, and the features are input into the self-attention neural network model. The output results are calculated through forward propagation to obtain the prediction of the cooling system fault type.

2. A method for detecting heat dissipation system faults based on a self-attention mechanism as claimed in claim 1, characterized in that: The fault types include fan dirty blockage faults and fan general faults. Before defining the fault type, the fan is preliminarily detected to obtain the power of the fan when it is working, and the fan fault state is determined by setting a monitoring current threshold.

3. A method for detecting heat dissipation system faults based on a self-attention mechanism as claimed in claim 2, characterized in that: The power formula of the fan when working is as follows: in Q :Air volume(m 3 / h)、 P :Wind pressure, η 1 Machine efficiency, η 2 Mechanical transmission efficiency.

4. The method for detecting heat dissipation system faults based on a self-attention mechanism according to claim 1, characterized in that: The feature encoder is a coding model, and the model setting method includes the following steps: Get the input feature x=[x1,x2…x4], perform data cleaning and normalization on the feature sequence to obtain the feature sequence, and then divide the data set into training, validation and test sets with a division ratio of 7:2:1; Absolute position encoding is performed on the feature sequence to obtain a feature vector sequence; The feature vector sequence is input into the Transformer encoder for encoding. The encoder consists of multiple encoders, each of which consists of two sublayers, one is a multi-head self-attention sublayer, and the other is a feedforward neural network sublayer. Each sublayer has a residual connection and layer normalization operation. The normalization layer uses LayerNorm layer normalization operation; The self-attention layer uses the self-attention mechanism to perform attention calculations with the encoding results as the query matrix, keyword matrix, and value matrix.

5. A method for detecting heat dissipation system faults based on a self-attention mechanism as claimed in claim 4, characterized in that: The attention calculation adopts the following formula: in, Q represents the query matrix, K represents the keyword matrix, K T represents the transpose of a matrix, d k represents the dimension of the input feature vector, V represents the value matrix, Attention( Q, K, V ) represents the final output self-attention score, and softmax represents the normalized exponential function.

6. A method for detecting heat dissipation system faults based on a self-attention mechanism as claimed in claim 1, characterized in that: The decoder performs autoregressive prediction, and the specific steps are as follows: Encode the encoder output as a Transformer decoder: Build a decoder that adds a multi-head self-attention layer and a feed-forward neural network layer to the encoder; Get the output before the decoder, perform linear transformation and scaling through matrix multiplication and scaling factors, and convert the original vector to the new vector space; The first input to the decoder is a special symbol <sos>, indicating the beginning of the sequence, the decoder starts from <sos>Start generating the output sequence and use the symbols generated in the previous time step as input for the next time step;< / sos> < / sos> At each time step, the decoder embeds the symbol generated by the previous time step into a continuous vector space through an embedding layer. The decoder connects the representation output by the encoder and the embedding vector of the previous time step, inputs them into the multi-head self-attention layer of the decoder, and calculates the attention score of the current time step. By weighting the encoder output representation, the decoder captures the information associated with the current time step in the input sequence. The decoder concatenates the weighted representation with the embedding vector to obtain the probability distribution of the output symbol of the current time step, which is used to predict the symbol of the next time step, and obtains the predicted output sequence y1, y2...yT; The predicted output sequence is sent to a softmax layer for fault classification prediction. By optimizing the cross entropy loss function and using the Adam optimizer to optimize the network parameters, the model convergence is accelerated.

Images