Fault detection method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to a fault detection method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring an operation data time sequence of a boiler steam turbine system; performing fault detection on the boiler steam turbine system according to the operation data time sequence through a preset fault detection model; wherein the preset fault detection model is obtained by training according to fault training samples and non-fault training samples in advance, the non-fault training samples are obtained by screening historical operation data, and the fault training samples comprise a first fault training sample and a second fault training sample; the first fault training sample is obtained by screening the historical operation data, and the second fault training sample is generated through a trained time sequence generative adversarial network. By adopting the method, the fault detection accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of power technology, and in particular to a fault detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] With the rapid development of the modern power industry, thermal power units, as a core component of the power system, have attracted much attention regarding their operational safety and stability. Boiler turbine systems are complex in structure and operate in harsh environments; therefore, timely and accurate fault detection is crucial for ensuring the safe and stable operation of the units and improving overall efficiency.

[0003] Among various fault detection technologies, data-driven methods are widely valued for their strong adaptability and rapid response. These methods construct identification models by mining and analyzing historical operating data of the unit, and under ideal conditions of sufficient and evenly distributed data, they can usually achieve high fault identification accuracy.

[0004] However, in actual industrial scenarios, due to the complexity and variability of on-site data, the constructed models often exhibit underfitting, leading to a decrease in the accuracy of fault detection and frequent missed and false alarms. Summary of the Invention

[0005] Therefore, it is necessary to provide a fault detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the accuracy of fault detection in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a fault detection method, including:

[0007] Obtain the time series of operating data for the boiler turbine system;

[0008] The boiler turbine system is used to detect faults based on the time series of operating data using a pre-set fault detection model.

[0009] The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from historical operating data, and the second fault training sample is generated by a trained temporal generative adversarial network.

[0010] In one embodiment, fault detection of the boiler turbine system is performed based on a time series of operating data, including:

[0011] Time convolution processing is performed on the time series of operational data to obtain fault characteristics;

[0012] Based on the fault characteristics, detect the target fault type and the target probability value of the target fault type in the boiler turbine system in a preset future time period.

[0013] In one embodiment, temporal convolution processing is performed on the runtime data time series to obtain fault features, including:

[0014] Causal dilation convolution is performed on the time series data to obtain residual data;

[0015] By aggregating residual data and runtime data time series, fault characteristics are obtained.

[0016] In one embodiment, the operational data time series includes multiple operational data time series composed of various operational data; the operational data time series are subjected to temporal convolution processing to obtain fault features, including:

[0017] Convert each runtime data time series into embedded features;

[0018] Enhanced features are obtained by using an attention mechanism to augment the embedded features of each loss.

[0019] Temporal convolution is performed on each enhancement feature to obtain fault features.

[0020] In one embodiment, the temporal generative adversarial network includes an embedding layer, a generator, a discriminator, and an embedding recovery layer; training the model of the temporal generative adversarial network includes:

[0021] Acquire noise data and filter third fault training samples from historical operation data;

[0022] The embedding layer converts noisy data into noisy embeddings and converts the third fault training sample into a fault training sample embedding.

[0023] Initial fault generation samples are generated based on noise embedding using a generator.

[0024] By embedding the recovery layer, the fault generation sample is mapped to the data dimension corresponding to the noise data to obtain the target fault generation sample, and the noise embedding is mapped to the data dimension corresponding to the noise data to obtain the target fault training sample.

[0025] The discriminator is used to distinguish between real and fake target fault generated samples and target fault training samples to obtain the discrimination result.

[0026] The discrimination loss of the discriminator is determined based on the discrimination result, the generation loss of the generator is determined based on the difference in data density distribution between the target fault generated sample and the target fault training sample, and the embedding transformation loss is determined based on the difference in data density distribution between the third fault training sample and the target fault training sample.

[0027] The discriminative loss, generation loss, and embedding transformation loss are fused to obtain the fusion loss. The temporal generative adversarial network is then iteratively optimized based on the fusion loss.

[0028] In one embodiment, obtaining fault training samples includes:

[0029] Select non-fault training samples and first fault training samples from historical operating data;

[0030] The number of fault samples to be generated is determined based on the number of non-fault training samples, the number of first fault training samples, and the preset sample ratio.

[0031] The trained temporal generative adversarial network generates a second fault training sample, and the number of second fault training samples generated is determined based on the number of fault samples to be generated.

[0032] Secondly, this application also provides a fault detection device, comprising:

[0033] The acquisition module is used to acquire the operating data time series of the boiler turbine system, which includes at least the main steam temperature time series.

[0034] The fault detection module is used to detect faults in the boiler turbine system based on the time series of operating data using a preset fault detection model. The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening historical operating data, and the second fault training sample is generated by a trained time series generative adversarial network.

[0035] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0036] Obtain the time series of operating data for the boiler turbine system;

[0037] The boiler turbine system is used to detect faults based on the time series of operating data using a pre-set fault detection model.

[0038] The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from historical operating data, and the second fault training sample is generated by a trained temporal generative adversarial network.

[0039] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0040] Obtain the time series of operating data for the boiler turbine system;

[0041] The boiler turbine system is used to detect faults based on the time series of operating data using a pre-set fault detection model.

[0042] The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from historical operating data, and the second fault training sample is generated by a trained temporal generative adversarial network.

[0043] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0044] Obtain the time series of operating data for the boiler turbine system;

[0045] The boiler turbine system is used to detect faults based on the time series of operating data using a pre-set fault detection model.

[0046] The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from historical operating data, and the second fault training sample is generated by a trained temporal generative adversarial network.

[0047] The aforementioned fault detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product first acquire the time series of operating data of the boiler turbine system, and then perform fault detection using a pre-set fault detection model. This fault detection model is trained using fault training samples and non-fault training samples. The non-fault training samples and some fault training samples are selected from historical operating data, while another portion of the fault training samples are generated through a trained temporal generative adversarial network (GAN). By introducing fault training samples generated by the temporal GAN, the quantity and diversity of fault samples can be effectively expanded, enabling the fault detection model to learn fault mode characteristics more effectively and comprehensively during training, thereby effectively alleviating the underfitting problem caused by the scarcity of fault samples. Finally, by using a fully trained fault detection model, the normal operating state and fault state of the boiler turbine system can be more accurately distinguished, significantly reducing the false alarm rate and false alarm rate of fault detection. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a diagram illustrating the application environment of a fault detection method in one embodiment of this application;

[0050] Figure 2 This is a flowchart illustrating the fault detection process in one embodiment of this application;

[0051] Figure 3 This is a flowchart illustrating the model training process of a temporal generative adversarial network in one embodiment of this application.

[0052] Figure 4 This is a schematic diagram of the fault detection results of multiple models in one embodiment of this application;

[0053] Figure 5 This is a structural block diagram of a fault detection device in one embodiment of this application;

[0054] Figure 6 This is an internal structural diagram of a computer device in one embodiment of this application. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0056] In one exemplary embodiment, such as Figure 1 As shown, a fault detection method is provided. This embodiment illustrates the application of this method to a terminal, where the terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. It is understood that this method can also be applied to servers, and can also be applied to systems including terminals and servers, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps 102-104. Wherein:

[0057] Step 102: Obtain the time series of operating data of the boiler turbine system.

[0058] The boiler turbine system is the core energy conversion device in a thermal power plant, responsible for the continuous and stable conversion of the chemical energy of fuel into electrical energy. The boiler turbine system mainly consists of two parts: the boiler and the turbine. On the boiler side, fuel burns in the furnace, releasing heat energy. The working fluid, water, flows sequentially through the economizer, water-cooled walls, and superheaters, ultimately forming high-temperature, high-pressure superheated steam, i.e., main steam. On the turbine side, the main steam is introduced through the main steam valve, expands, and performs work, driving the turbine rotor and the connected generator rotor to rotate at high speed, thus completing the conversion of thermal energy, mechanical energy, into electrical energy.

[0059] Given the complex structure of boiler turbine systems, the tight coupling of operating parameters, and the harsh conditions of operation under high temperature and high pressure, their operating status directly affects the safety, economy, and reliability of the unit. Abnormalities in key parameters of the boiler turbine system, such as significant fluctuations in main steam temperature or pressure, may induce a series of chain reactions of equipment damage, including metal material creep, thermal stress fatigue, and turbine water erosion, which may lead to unplanned shutdowns or accidents in severe cases.

[0060] Therefore, fault detection of boiler turbine systems based on operational data has a clear engineering necessity. This technical approach involves real-time acquisition and analysis of multi-source time-series data reflecting the system status (such as main steam temperature, pressure, vibration, and flow rate) to construct a data-driven model, aiming to achieve early identification and warning of abnormal operating conditions and initial faults.

[0061] Operating data can refer to the physical parameters that are measured and recorded in real time by various sensors during the operation of the boiler turbine system. These parameters include at least the main steam temperature (measured at the boiler superheater outlet), and may also include the main steam pressure (measured at the superheater outlet), superheater inlet feedwater flow rate, superheater outlet steam flow rate, flue gas temperature (measured at the superheater flue gas inlet and outlet), flue gas flow rate (measured at the superheater flue gas side), desuperheating water valve opening, and blower speed.

[0062] Operational data time series can refer to a set of operational data arranged in chronological order. For example, the main steam temperature value is recorded once per minute and continuously recorded for 24 hours, forming an ordered set of 1440 data points, which is a main steam temperature time series.

[0063] For example, a sensor network deployed on key equipment such as boilers, steam turbines, and pipelines continuously collects operational data reflecting the system's status. This operational data is recorded and stored in real time by the data acquisition system, forming a multi-dimensional time series of operational data ordered by timestamps.

[0064] In some feasible implementations, the runtime data time series can be preprocessed before being input into the model, and then the preprocessed runtime data time series can be input into the fault detection model for fault detection. The process of acquiring and preprocessing the runtime data time series may include:

[0065] (1) Data acquisition: The acquisition object is the multi-dimensional real-time monitoring data of the main steam temperature system of the boiler turbine system. The acquisition frequency is set to 1Hz~5Hz, which can be dynamically adjusted according to the unit's operating stability.

[0066] (2) Outlier cleaning: A cleaning strategy based on the 3σ principle is adopted. First, the mean μ and standard deviation σ are calculated for the single-parameter time series data. Data that exceed the range of [μ-3σ, μ+3σ] are marked as potential outliers. The marked outliers are filled by linear interpolation (the interval between adjacent valid data is ≤5 time steps).

[0067] (3) Blank value processing: For blank values ​​caused by data transmission interruption, a combination of forward filling and backward filling is used to fill them; if the blank duration is >10s, a data re-acquisition command is triggered, and a fill value is generated based on the data model of similar operating conditions in the same period in history (matched by unit load, coal consumption and feedwater temperature).

[0068] (4) Standardization: The Z-Score standardization method is used to normalize the cleaned numerical data to eliminate the interference of different parameter magnitudes on model training.

[0069] Step 104: Using a preset fault detection model, fault detection is performed on the boiler turbine system based on the time series of operating data. The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from historical operating data, and the second fault training sample is generated by a trained time series generative adversarial network.

[0070] Among them, the fault detection model can refer to a mathematical model or neural network model that has been trained and has fixed parameters before being put into use, and is used to determine whether there is a fault in the boiler turbine system at present or in a preset future period based on the input data.

[0071] Fault training samples can refer to data samples labeled as faults, which are used to enable the fault detection model to learn the features of the fault state during the model training process.

[0072] Non-fault training samples refer to data samples labeled as non-faulty, which are used to allow the fault detection model to learn the features of normal operating conditions during the model training process.

[0073] Temporal Generative Adversarial Networks (GANs) are a class of artificial intelligence models specifically designed to generate time-series data (such as vibration signals and temperature sequences). A temporal GAN ​​consists of at least one generator and one discriminator, which compete against each other to enable the generator to produce realistic fake data.

[0074] For example, after obtaining the operational data time series, the obtained operational data time series can be input into a pre-trained fault detection model. The fault detection model calculates and analyzes the input sequence to obtain the fault detection results of the boiler turbine system.

[0075] As an example, the fault detection result can be a judgment result about the system status, used to indicate whether the boiler turbine system is currently in a normal state or has a fault.

[0076] As another example, the fault detection result may also include at least one fault type and the probability of the presence or occurrence of that fault type.

[0077] Before practical deployment, fault detection models need to be trained to understand the inherent patterns and characteristics of normal and various fault states in boiler turbine systems. The trained fault detection model can then be deployed on application terminals to achieve online monitoring and fault detection of real-time operational data. Model training typically relies on a large number of balanced positive and negative samples. However, in real-world industrial scenarios, the number of available fault samples is limited, far fewer than samples under normal operating conditions, leading to a severe imbalance in the training dataset. This imbalance causes the model to overfit to the majority class (normal samples) while under-learning features of the minority class (fault samples), resulting in a significant decrease in the model's ability to identify abnormal states. This problem directly affects the reliability of fault detection in practical applications.

[0078] Therefore, this embodiment utilizes the trained temporal generative adversarial network to expand the number and diversity of fault samples, enabling the fault detection model to learn fault mode features more effectively and comprehensively during the training process, thereby effectively alleviating the model underfitting problem caused by the scarcity of fault samples.

[0079] For example, before model training, historical operating data can be obtained from the historical operating database of the boiler turbine system. Then, the historical operating data can be labeled manually or by a trained labeling model. Based on the labeling results, the time series of operating data under non-fault conditions can be selected as non-fault training samples, and the time series of operating data under fault conditions can be selected as the first fault training samples.

[0080] Because actual failures occur infrequently, the screening process may require covering several years of historical data to collect as many diverse failure mode samples as possible, but the total number remains limited. Therefore, the first set of collected failure training samples can be input into a temporal generative adversarial network (GAN) for feature learning. After the GAN training converges, a time series of operational data with similar temporal dynamics to real failure data is generated by inputting a random noise vector, serving as the second set of failure training samples. During the generation process, the intensity and variation patterns of failure features in the generated samples can be controlled by adjusting the input parameters of the generative network, thereby obtaining a diverse range of failure training samples.

[0081] Then, the prepared non-fault training samples, the first fault training samples, and the second fault training samples are input into the fault detection model to be trained. The training process can adopt a standard machine learning training procedure, dividing the sample data into a training set and a validation set. By comparing its predicted output with the true label of the sample, the parameters are iteratively optimized so that the trained fault detection model can effectively distinguish between normal and fault states of the system.

[0082] In some feasible implementations, quality control can be performed on the second fault training samples generated by the temporal generative adversarial network.

[0083] For example, quality control may include timing consistency verification, statistical distribution verification, fault feature verification, and dynamic sample screening.

[0084] Temporal consistency verification includes: evaluating the similarity of the temporal trends of the generated samples and the real samples by calculating the dynamic time warping distance between the two.

[0085] Statistical distribution verification includes: conducting significance tests on key statistical indicators (such as mean, variance, skewness, kurtosis, autocorrelation coefficient, etc.) of the generated sample and the real sample (e.g., using a t-test with a significance level of α=0.05), requiring that there are no significant differences between the two in all statistical indicators; at the same time, measuring the difference between the distribution of the generated sample and the distribution of the real sample through KL (relative entropy) divergence, with a KL divergence value ≤0.1, to ensure that the generated sample is highly consistent with the real sample in terms of statistical characteristics.

[0086] Fault feature verification includes: constructing a feature verification rule base for the indicative features of different fault types (such as the feature of slow rise in main steam temperature and drop in flue gas temperature in superheater ash accumulation fault, and the feature of sudden rise in main steam temperature and no response in desuperheating water flow rate in desuperheating water valve jamming fault), and using fuzzy inference algorithm to determine whether the generated sample contains the indicative features of the corresponding fault, and the feature matching degree must be ≥90%.

[0087] Dynamic sample selection involves setting a threshold for the pass rate of generated samples. When the pass rate of a batch of generated samples is less than 70%, the training parameters of the temporal generative adversarial network are adjusted (e.g., increasing the number of hidden layer neurons in the sequence generator and decreasing the learning rate by 0.001 to 0.0005). The network is then retrained for 10 to 20 iterations and samples are generated again until the pass rate is ≥70%, ensuring the effectiveness of the expanded samples. Finally, the ratio of the number of expanded samples to the number of real samples is controlled at 3:1 to 5:1 to achieve sample class balance (e.g., the difference in the number of samples in each class is ≤10%).

[0088] In the aforementioned fault detection method, the operating data time series of the boiler turbine system is first acquired, and then a pre-set fault detection model is used for fault detection. This fault detection model is trained using fault training samples and non-fault training samples. The non-fault training samples and some fault training samples are selected from historical operating data, while another portion of the fault training samples are generated through a trained temporal generative adversarial network (GAN). By introducing fault training samples generated by the temporal GAN, the quantity and diversity of fault samples can be effectively expanded, enabling the fault detection model to learn fault mode characteristics more effectively and comprehensively during training, thereby effectively alleviating the underfitting problem caused by the scarcity of fault samples. Finally, by using a fully trained fault detection model, the normal operating state and fault state of the boiler turbine system can be more accurately distinguished, significantly reducing the false negative and false positive rates of fault detection.

[0089] In one exemplary embodiment, such as Figure 2 As shown, fault detection of the boiler turbine system is performed based on the time series of operating data, including steps 202 to 204. Wherein:

[0090] Step 202: Perform temporal convolution processing on the time series of running data to obtain fault characteristics.

[0091] It should be noted that conventional fault detection models lack the ability to learn the temporal characteristics inherent in operational data. The operating parameters of boiler turbine systems exhibit significant dynamic characteristics and temporal dependencies. Traditional methods, such as least squares support vector machines, due to inherent structural limitations, fail to fully exploit the time-series patterns in the data and neglect the intrinsic correlations between state parameters at adjacent time points. Furthermore, they have a weak overall perception of long-range causality and abnormal patterns in complex operating conditions and are easily affected by local fluctuations.

[0092] Temporal convolution processing refers to a neural network operation specifically designed for processing time-series data. It extracts local dependencies and temporal patterns by sliding convolutional kernels along the time dimension. Fault detection models can include temporal convolutional networks to implement temporal convolution processing.

[0093] Fault characteristics can refer to key information extracted from raw operating data that can characterize potential abnormal states. For example, abnormal fluctuation patterns, inflection points in the trend of change, and frequency characteristics of the main steam temperature can indicate that a fault is about to occur.

[0094] For example, the acquired time series of operational data is input into the temporal convolutional network of the fault detection model. This temporal convolutional network can consist of multiple layers, each layer using a predefined convolutional kernel to perform convolution operations on the time axis of the input sequence. Through this layer-by-layer local perception and combination, the network can progressively extract deep, abstract feature representations related to equipment condition degradation or fault initiation from the raw, potentially noisy, operational data. The output of the temporal convolutional network is the fault feature.

[0095] Step 204: Based on the fault characteristics, detect the target fault type and the target probability value of the target fault type in the boiler turbine system in a preset future time period.

[0096] Among them, the preset future time period can refer to a future time interval set according to the required advance warning time.

[0097] The target fault type can refer to the specific fault category that the fault detection model is predefined as to predict.

[0098] The target probability value can refer to the confidence level of the target fault type in a future time period, as output by the fault detection model. It can be a value between 0 and 1. For example, assuming the fault detection result is "superheater leak: 0.85", it means that there is an 85% confidence level that the superheater leak fault will occur in a preset future time period.

[0099] Fault detection models can also include classifiers to classify faults.

[0100] For example, the fault features obtained by temporal convolution are input into the classifier of the fault detection model. The classifier processes the input features and calculates the probability value of each fault type occurring in a preset future time period. Then, the fault detection result is output according to the preset output rules, such as outputting the fault type with the highest probability value, or outputting the fault type with a probability value higher than the preset probability threshold. The fault detection result includes the target fault type that meets the preset output rules and the target probability value corresponding to the target fault type.

[0101] In this embodiment, temporal convolution processing effectively extracts temporal features from operational data, capturing long-term dependencies and temporal patterns, overcoming the shortcomings of traditional methods in learning temporal characteristics. By outputting specific future fault types and their probabilities, it provides operators with clear and quantifiable early warning information, enabling proactive intervention and preventative maintenance, and significantly improving the safety and reliability of the boiler turbine system.

[0102] In one exemplary embodiment, temporal convolution processing is performed on the runtime data time series to obtain fault characteristics, including:

[0103] Causal dilation convolution is performed on the time series of operational data to obtain residual data; the residual data and the time series of operational data are aggregated to obtain fault characteristics.

[0104] It should be noted that conventional time-series feature extraction methods often struggle to simultaneously extract long-term dependencies and local detailed features when processing operational data of boiler turbine systems. Because system operating parameters exhibit dynamic characteristics across multiple time scales, traditional convolutional methods are prone to limitations in their receptive field or loss of detail when processing such data, resulting in extracted features that fail to fully reflect the equipment's operating status.

[0105] Causal dilated convolution refers to convolution operations that handle causal constraints and increase the receptive field through dilation coefficients, ensuring that the output depends only on the current and past inputs and does not use future information. For example, assuming a one-dimensional convolution kernel with a dilation coefficient of 2 is used on the main steam temperature time series, the calculation at each position only uses the data points at that position and before it, but the dilation mechanism can capture historical dependencies that are far away.

[0106] Residual data can refer to the difference information or higher-order features extracted after specific processing, representing the patterns implicit in the original data. For example, time-series features that are not easily observed directly, such as the rate of temperature change and the periodicity of fluctuations, can be extracted from main steam temperature data through causal dilatation convolution.

[0107] Temporal convolutional networks can be used to create causal dilated convolutional networks.

[0108] For example, the runtime data time series is input into a causal dilated convolutional network, which performs convolution operations along the time dimension using a preset dilation factor. As the convolutional kernel slides along the time axis, it only considers data from the current time point and previous time points, ensuring that each output feature point does not contain future information. The dilation mechanism inserts zero values ​​between kernel elements, causing the receptive field to grow exponentially with each convolution operation, thereby capturing long-term temporal dependencies with fewer layers. After the convolution operation, a non-linear activation function is used to process the residual data, which retains high-order temporal patterns and anomalous signs from the original time series, especially those pre-fault features spanning long time spans. The obtained residual data is then added point-by-point to the original runtime data time series, ensuring that the two sequences are aligned along the time dimension to obtain fault features. These fault features retain the complete information of the original runtime data while incorporating the deep features extracted through causal dilated convolution.

[0109] In this embodiment, the synergistic effect of causal dilated convolution and residual aggregation effectively overcomes the shortcomings of traditional methods in simultaneously capturing features across multiple time scales during feature extraction. Causal dilated convolution ensures the effective capture of long-term dependencies, while the residual aggregation mechanism guarantees the complete preservation of local detailed features. This combination enables the extracted fault features to contain both macroscopic trend information and retain microscopic fluctuation features, providing a more comprehensive and discriminative feature representation for accurately identifying early faults and complex fault modes, thereby significantly improving the reliability of the fault detection system.

[0110] In an exemplary embodiment, the runtime data time series includes multiple runtime data time series composed of various runtime data; the runtime data time series are subjected to temporal convolution processing to obtain fault features, including:

[0111] The time series of each running data is converted into embedded features; the embedded features of each loss are enhanced through an attention mechanism to obtain enhanced features; and the enhanced features are processed by temporal convolution to obtain fault features.

[0112] It should be noted that in the multivariate operating data of boiler turbine systems, different parameters exhibit significant differences in their sensitivity and contribution to faults. Assigning equal importance to all operating data, or relying on prior knowledge for fixed weight allocation, makes it difficult to adapt to the dynamic changes in the characteristics and correlations of the boiler turbine system under different operating conditions. This results in key fault symptoms being overwhelmed by a large number of normal parameters, affecting the accuracy of fault detection.

[0113] For example, firstly, an embedding layer or fully connected layer is used to linearly transform the data at each time step of each running data time series, converting them from original one-dimensional scalar values ​​into real-number vectors of a specified dimension. This transformation process preserves the temporal order of the original sequence and unifies running data with different physical dimensions and numerical ranges into the same feature representation space. Then, all embedded features are used as input, and an attention scoring function is used to calculate the importance score of each embedded feature relative to the global context. The obtained attention scores are normalized to obtain the attention weights corresponding to each embedded feature, ensuring that the sum of all attention weights is 1. Then, a scalar multiplication operation is performed on each embedded feature and its corresponding attention weight to amplify the contribution of important features and suppress the influence of secondary features. The weighted embedded features at all time steps maintain the original temporal order, forming an enhanced feature sequence, i.e., enhanced features. This enhanced feature, while maintaining the original temporal structure, highlights the parameter features that are more critical for fault detection. Finally, the enhanced features are input into a temporal convolutional network for feature extraction to obtain fault features.

[0114] In some feasible implementations, the enhanced features are obtained by enhancing the embedded features of each loss through an attention mechanism. This can be achieved by: first, performing linear transformations on the embedded features of multiple input runtime data time series using three different learnable weight matrices WQ, WK, and WV to generate a query matrix Q, a key matrix K, and a value matrix V; then, performing matrix multiplication on the transposes of the query matrix Q and the key matrix K to obtain a score matrix with the shape [sequence length, sequence length]. Each element S[i, j] in the score matrix represents the original correlation or similarity between the i-th element (as the query) and the j-th element (as the key) in the runtime data time series; to prevent the subsequent gradient from being too small due to the high feature dimension of the key matrix K, the score matrix can be scaled based on the feature dimension of the key matrix K; then, the scaled score matrix is ​​normalized using a normalized exponential function to obtain the attention weights; finally, the calculated attention weights are multiplied by the value matrix V to obtain the enhanced features. This result highlights features that are more important in the current context and suppresses the influence of secondary features, thereby enhancing the model's sensitivity to key failure symptoms.

[0115] In this embodiment, a unified representation of multi-source heterogeneous data is achieved through embedding transformation, while the attention mechanism dynamically adjusts the importance weights of various data types according to actual operating conditions. This dynamic feature enhancement mechanism ensures a prominent emphasis on key fault symptoms, enabling subsequent temporal convolution processing to more effectively focus on the most relevant fault information, thereby significantly improving the accuracy and reliability of fault detection under complex multivariate operating conditions.

[0116] In one exemplary embodiment, such as Figure 3 As shown, the temporal generative adversarial network includes an embedding layer, a generator, a discriminator, and an embedding recovery layer; the model training of the temporal generative adversarial network includes steps 302 to 314. Wherein:

[0117] Step 302: Obtain noise data and select a third fault training sample from historical operation data.

[0118] It should be noted that, due to the limited number of fault samples and the complex feature distribution of boiler turbine systems, simple data interpolation or perturbation methods can easily cause the generated samples to deviate from the distribution characteristics of the real fault data, failing to meet the requirements of model training for sample quality and diversity.

[0119] Noise data can refer to a data vector randomly sampled from a specific probability distribution (such as a Gaussian distribution or a uniform distribution). For example, a 100-dimensional random vector sampled from a Gaussian distribution with a mean of 0 and a variance of 1.

[0120] The third fault training sample can refer to actual fault data records selected from the historical operation database and labeled with specific fault types.

[0121] For example, a set of noise data is randomly sampled from a preset probability distribution, and at the same time, actual operation data records that conform to a specific fault type are queried and extracted from the system's historical operation database as a third fault training sample.

[0122] Step 304: Through the embedding layer, the noisy data is converted into a noisy embedding, and the third fault training sample is converted into a fault training sample embedding.

[0123] The embedding layer can refer to a neural network layer that maps input data to a low-dimensional feature space.

[0124] Embedding can refer to the low-dimensional, dense feature vector representation obtained after transformation through an embedding layer.

[0125] For example, the acquired noise data and the third fault training sample are respectively input into the embedding layer. The input data is transformed into a low-dimensional dense vector representation of the same dimension through linear transformation or nonlinear mapping by the weight matrix in the embedding layer, namely noise embedding and fault training sample embedding.

[0126] Step 306: Generate initial fault generation samples based on noise embedding using a generator.

[0127] In this context, a generator can refer to a functional module composed of neural networks that is capable of generating new data samples from latent space vectors.

[0128] Initial fault generation samples can refer to fault data samples that have not yet undergone dimensionality recovery, generated by the generator based on noise embedding output.

[0129] For example, after obtaining the noise embedding, the noise embedding is input into the generator, and after forward propagation calculation through the multi-layer neural network inside the generator, the generated data with the same embedding dimension as the fault training sample is output, which is the initial fault generation sample.

[0130] Step 308: By embedding the recovery layer, the fault generation sample is mapped to the data dimension corresponding to the noise data to obtain the target fault generation sample, and the noise embedding is mapped to the data dimension corresponding to the noise data to obtain the target fault training sample.

[0131] Among them, the embedding recovery layer can refer to a neural network layer that realizes the inverse mapping from the embedding space to the original data space.

[0132] The target fault generation sample can refer to a fault sample generated by the generator that, after being transformed by the embedding recovery layer, has the same dimension as the original noise data.

[0133] For example, the initial fault generation sample and the noise embedding are respectively input into the embedding recovery layer. Through the inverse transformation operation in the embedding recovery layer, the input data is mapped from the embedding space back to the same dimensional space as the original noise data to obtain the target fault generation sample and the target fault training sample.

[0134] Step 310: The discriminator is used to distinguish between real and fake target fault generated samples and target fault training samples to obtain the discrimination result.

[0135] Here, the discriminator can refer to a neural network classifier used to distinguish whether the input sample is real data or generated data.

[0136] The discrimination result can refer to the probability value of the discriminator's judgment on the authenticity of the input sample.

[0137] For example, the target fault generated sample and the target fault training sample are input into the discriminator in sequence. The discriminator calculates the probability value of each sample being real data through forward propagation, and forms the discrimination result.

[0138] Step 312: Determine the discrimination loss of the discriminator based on the discrimination result, determine the generation loss of the generator based on the difference in data density distribution between the target fault generated sample and the target fault training sample, and determine the embedding transformation loss based on the difference in data density distribution between the third fault training sample and the target fault training sample.

[0139] Here, discriminant loss can refer to the loss function value that measures the discriminator's ability to distinguish between real and generated samples.

[0140] Generation loss can refer to the loss function value that measures the quality of samples generated by the generator.

[0141] Embedding transformation loss can refer to the loss function value that measures the degree of information retention during the transformation between the embedding layer and the embedding recovery layer.

[0142] For example, the cross-entropy loss of the discriminator is calculated as the discrimination loss based on the discrimination result; the generation loss is calculated by comparing the data distribution difference between the target fault generation sample and the target fault training sample; and the embedding transformation loss is calculated by comparing the data distribution difference between the third fault training sample and the target fault training sample.

[0143] Step 314: Fuse the discriminative loss, generation loss, and embedding transformation loss to obtain the fusion loss, and iteratively optimize the temporal generative adversarial network based on the fusion loss.

[0144] For example, the discrimination loss, generation loss, and embedding transformation loss are weighted and summed according to preset weight coefficients to obtain the fusion loss. The gradient is calculated based on the fusion loss, and the network parameters of the discriminator, generator, embedding layer, and embedding recovery layer are updated sequentially through the backpropagation algorithm.

[0145] In this embodiment, the embedding layer ensures the effective representation of data in the feature space, the embedding recovery layer ensures the dimensionality consistency of the generated samples, and the fusion optimization of multiple loss functions can effectively balance multiple objectives such as sample quality, distribution matching and feature fidelity. This enables the generation of high-quality fault samples that not only conform to the distribution of real fault data but also have sufficient diversity, providing sufficient data support for the subsequent training of the fault detection model.

[0146] In one exemplary embodiment, obtaining fault training samples includes:

[0147] Non-fault training samples and first fault training samples are selected from historical operating data; the number of fault samples to be generated is determined based on the number of non-fault training samples, the number of first fault training samples, and the preset sample ratio; second fault training samples are generated through the trained temporal generative adversarial network, and the number of second fault training samples generated is determined based on the number of fault samples to be generated.

[0148] It should be noted that during the training of fault detection models for boiler turbine systems, a common problem is the scarcity of fault samples and a severe imbalance between the fault samples and normal samples. This data imbalance leads to a significant decrease in the ability of the trained model to identify fault states, making it prone to missed detections and false positives, which seriously affects the practical application of fault detection.

[0149] The preset sample ratio can refer to the ratio of normal samples to faulty samples in the pre-set training dataset.

[0150] For example, historical operating data can be obtained first from the historical operating database of the boiler turbine system. Then, the historical operating data is labeled manually or using a trained labeling model. Based on the labeling results, time series of operating data under non-fault conditions are selected as non-fault training samples, and time series of operating data under fault conditions are selected as the first fault training samples. The number of selected non-fault training samples and the number of first fault training samples are then counted. The theoretically required total number of fault samples is calculated based on a preset sample ratio. The difference between this total number of fault samples and the existing number of first fault training samples is determined as the number of fault samples to be generated. Then, using a trained temporal generative adversarial network model, a corresponding number of second fault training samples are generated according to the number of fault samples to be generated.

[0151] In this embodiment, a preset sample ratio guides the generation of samples, ensuring a balance between positive and negative samples in the training dataset. A temporal generative adversarial network is used to generate high-quality synthetic fault samples, significantly expanding the diversity of fault samples. Therefore, the accuracy and generalization ability of the fault detection model for recognizing various fault states can be effectively improved.

[0152] In one exemplary embodiment, a fault detection scheme based on the fusion of Temporal Generative Adversarial Network (TGAN) and Self-Attention Temporal Convolutional Network (SATCN) is proposed. By fusing TGAN and SATCN, a high-precision fault detection model for small-sample, imbalanced data conditions of main steam temperature anomalies is constructed, achieving the following significant technical effects:

[0153] The detection accuracy is significantly improved. In the simulation experiment of coal-fired unit, compared with the traditional temporal convolutional network, the fault detection accuracy of this embodiment is improved by 7.9%, the F1 score (the weighted average of precision and recall) is improved by 4.8%, the false detection rate is reduced from 4.3% to 0%, and the false negative rate is reduced from 2.4% to 0.2%, demonstrating a stronger fault identification capability.

[0154] Enhanced sample robustness. The fault samples generated by the temporal generative adversarial network are highly consistent with the real data in terms of temporal features and statistical distribution, effectively alleviating the problems of sample scarcity and imbalance. This enables the model to have sufficient learning ability for all types of faults during training, avoiding bias towards the majority class.

[0155] The global feature extraction capability is enhanced. The introduction of the self-attention mechanism enables the model to adaptively focus on key time steps and feature dimensions, significantly improving the ability to distinguish complex failure modes (such as superheater ash accumulation and fan surge), and enhancing the interpretability and stability of the model.

[0156] In an exemplary embodiment, a simulator system was used to simulate the normal operation of an 800MW coal-fired unit under steady-state conditions and six common faults under abnormal main steam temperature conditions as shown in Table 1. 1600 sets of simulation operation data were collected, and the training samples were expanded from 1000 sets to 3000 sets using a temporal generative adversarial network. The convolution kernel size was set to 8 and the learning rate to 0.001. After testing, the model with expanded temporal generative adversarial network data had a lower false negative rate.

[0157] Table 1. Common System Fault Types and Main Fault Phenomena under Abnormal Main Steam Temperature Conditions

[0158]

[0159] Comparative experiments were conducted on the same test set with common fault detection models such as RNN (Recurrent Neural Network), LSSVM (Least Squares Support Vector Machine), and LSTM (Long Short-Term Memory). The fault detection results for different models are as follows: Figure 4 As shown, the TGAN-SATCN model used in this embodiment has higher fault detection accuracy. Compared with the LSTM, LSSVM and KNN models, the accuracy is improved by 12.3%, 14.8% and 17.8% respectively, and the false negative rate and false positive rate are significantly reduced.

[0160] In some feasible implementations, the training process for SATCN includes:

[0161] (1) Training dataset partitioning: Stratified sampling can be used to divide the preprocessed fault training samples into training set, validation set and test set in a ratio of 7:2:1, and ensure that the fault category distribution in each set is consistent with the overall data.

[0162] (2) Cross-entropy loss is used as the loss function to measure the difference between the SATCN model prediction results and the actual fault labels, thus solving the problem of class judgment error and fault class imbalance in the main steam temperature anomaly detection and fault detection tasks. The formula is as follows:

[0163]

[0164] In the formula, N is the number of samples in the training batch; C is the total number of main steam temperature fault categories (including the normal category, a total of 5 categories, i.e., C=5); y i,c Let be the true label of the i-th sample corresponding to the c-th type of fault, using one-hot encoding; be the probability value of the i-th sample being predicted by the model as the c-th type of fault, taking a value in the range [0,1], and satisfying . This loss function effectively penalizes the model's bias in predicting fault categories, especially for easily confused fault types in the main steam temperature system (such as superheater ash accumulation and flue gas damper failure, both of which can lead to an increase in main steam temperature), thereby enhancing the model's ability to distinguish between different fault characteristics.

[0165] (3) Optimizer selection and parameter settings: The AdamW optimizer (an adaptive moment estimation optimizer with weight decay) is adopted. The initial learning rate is set to 0.001 and the weight decay coefficient is set to 0.01 to prevent the model from overfitting. The cosine annealing learning rate scheduling strategy is adopted, and the learning rate is decayed to 0.8 times the current value every 50 iterations. The minimum learning rate is not lower than 1e-5.

[0166] (4) Model hyperparameter configuration: The causal convolution kernel size of the TCN part is set to 3 to 5, the dilation factor is set in a zigzag sequence of [1,2,4,8] (a total of 4 layers of dilated convolution), the number of residual blocks is set to 6 to 8, and the dropout rate is set to 0.2 to 0.3; the number of attention heads of the self-attention mechanism part is set to 4 to 8, and the model feature dimension is set to 128 to 256; the batch size during model training is set to 32 to 64, and the number of training rounds is set to 300 to 500 rounds.

[0167] This embodiment proposes a fault sample augmentation method based on temporal generative adversarial networks (GANs), effectively alleviating the model overfitting problem caused by scarce fault samples and class imbalance. The temporal GAN ​​constructs a low-dimensional adversarial space through embedding and recovery functions, enabling it to learn the temporal distribution characteristics of real fault data and generate high-quality augmented samples that conform to temporal patterns. This method balances the number of samples in each class, significantly improving the model's ability to identify minority class faults and avoiding overfitting of the majority class by traditional models, thereby enhancing generalization performance and robustness in practical applications.

[0168] This embodiment proposes a temporal convolutional network module incorporating a self-attention mechanism. The temporal convolutional network module itself utilizes causal dilated convolution and residual structures to extract multi-scale temporal features, while the self-attention mechanism adaptively focuses on important features at different time steps by calculating query-key correlation. The SATCN model can improve the model's ability to extract global features from main steam temperature time-series data.

[0169] The TGAN-SATCN model proposed in this embodiment integrates sample augmentation and feature enhancement strategies, which can improve the accuracy of anomaly detection and fault diagnosis. TGAN ensures the quality and balance of input data, while SATCN, through its powerful temporal modeling and attention weighting mechanism, significantly improves the ability to identify early faults and multiple fault modes.

[0170] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0171] Based on the same inventive concept, this application also provides a fault detection device for implementing the fault detection method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more fault detection device embodiments provided below can be found in the limitations of the fault detection method described above, and will not be repeated here.

[0172] In one exemplary embodiment, such as Figure 5 As shown, a fault detection device is provided, including: an acquisition module 502 and a fault detection module 504, wherein:

[0173] The acquisition module 502 is used to acquire the operating data time series of the boiler turbine system, which includes at least the main steam temperature time series.

[0174] The fault detection module 504 is used to detect faults in the boiler turbine system based on the time series of operating data using a preset fault detection model. The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from historical operating data, and the second fault training sample is generated by a trained time series generative adversarial network.

[0175] In one exemplary embodiment, the fault detection module 504 is further configured to:

[0176] Time convolution processing is performed on the time series of operational data to obtain fault characteristics;

[0177] Based on the fault characteristics, detect the target fault type and the target probability value of the target fault type in the boiler turbine system in a preset future time period.

[0178] In one exemplary embodiment, the fault detection module 504 is further configured to:

[0179] Causal dilation convolution is performed on the time series data to obtain residual data;

[0180] By aggregating residual data and runtime data time series, fault characteristics are obtained.

[0181] In one exemplary embodiment, the runtime data time series includes multiple runtime data time series composed of various runtime data; the fault detection module 504 is further configured to:

[0182] Convert each runtime data time series into embedded features;

[0183] Enhanced features are obtained by using an attention mechanism to augment the embedded features of each loss.

[0184] Temporal convolution is performed on each enhancement feature to obtain fault features.

[0185] In an exemplary embodiment, the temporal generative adversarial network includes an embedding layer, a generator, a discriminator, and an embedding recovery layer; the fault detection device further includes a training model, the training module being used for:

[0186] Acquire noise data and filter third fault training samples from historical operation data;

[0187] The embedding layer converts noisy data into noisy embeddings and converts the third fault training sample into a fault training sample embedding.

[0188] Initial fault generation samples are generated based on noise embedding using a generator.

[0189] By embedding the recovery layer, the fault generation sample is mapped to the data dimension corresponding to the noise data to obtain the target fault generation sample, and the noise embedding is mapped to the data dimension corresponding to the noise data to obtain the target fault training sample.

[0190] The discriminator is used to distinguish between real and fake target fault generated samples and target fault training samples to obtain the discrimination result.

[0191] The discrimination loss of the discriminator is determined based on the discrimination result, the generation loss of the generator is determined based on the difference in data density distribution between the target fault generated sample and the target fault training sample, and the embedding transformation loss is determined based on the difference in data density distribution between the third fault training sample and the target fault training sample.

[0192] The discriminative loss, generation loss, and embedding transformation loss are fused to obtain the fusion loss. The temporal generative adversarial network is then iteratively optimized based on the fusion loss.

[0193] In one exemplary embodiment, the training module is further configured to:

[0194] Select non-fault training samples and first fault training samples from historical operating data;

[0195] The number of fault samples to be generated is determined based on the number of non-fault training samples, the number of first fault training samples, and the preset sample ratio.

[0196] The trained temporal generative adversarial network generates a second fault training sample, and the number of second fault training samples generated is determined based on the number of fault samples to be generated.

[0197] Each module in the aforementioned fault detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0198] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a fault detection method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0199] Those skilled in the art will understand that Figure 6The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0200] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0201] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0202] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0203] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0204] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0205] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0206] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A fault detection method, characterized in that, The method includes: Obtain the time series of operating data for the boiler turbine system; The boiler turbine system is used to detect faults based on the time series of the operating data using a preset fault detection model. The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by screening from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by screening from the historical operating data, and the second fault training sample is generated by a trained temporal generative adversarial network.

2. The method according to claim 1, characterized in that, The fault detection of the boiler turbine system based on the time series of the operating data includes: The time series of the operational data is subjected to temporal convolution processing to obtain fault characteristics; Based on the fault characteristics, the target fault type and the target probability value of the target fault type of the boiler turbine system in a preset future time period are detected.

3. The method according to claim 2, characterized in that, The step of performing temporal convolution processing on the time series of the operational data to obtain fault features includes: The time series of the running data is subjected to causal dilated convolution processing to obtain residual data; By aggregating the residual data and the time series of the operational data, fault characteristics are obtained.

4. The method according to claim 2, characterized in that, The operational data time series includes multiple operational data time series composed of various operational data; the step of performing temporal convolution processing on the operational data time series to obtain fault features includes: Convert the time series of the aforementioned operational data into embedded features; Enhanced features are obtained by using an attention mechanism to augment the embedded features of each loss. Temporal convolution is performed on each of the enhanced features to obtain fault features.

5. The method according to any one of claims 1 to 4, characterized in that, The temporal generative adversarial network includes an embedding layer, a generator, a discriminator, and an embedding recovery layer; the model training of the temporal generative adversarial network includes: Acquire noise data and filter third fault training samples from historical operation data; The embedding layer converts the noise data into noise embeddings and converts the third fault training sample into a fault training sample embedding. The generator generates initial fault generation samples based on the noise embedding. Through the embedding recovery layer, the fault generation sample is mapped to the data dimension corresponding to the noise data to obtain the target fault generation sample, and the noise embedding is mapped to the data dimension corresponding to the noise data to obtain the target fault training sample; The discriminator is used to distinguish between real and fake target fault generated samples and target fault training samples to obtain a discrimination result. The discrimination loss of the discriminator is determined based on the discrimination result, the generation loss of the generator is determined based on the difference in data density distribution between the generated target fault sample and the training target fault sample, and the embedding transformation loss is determined based on the difference in data density distribution between the training third fault sample and the training target fault sample. The discriminative loss, the generation loss, and the embedding transformation loss are fused to obtain a fusion loss, and the temporal generative adversarial network is iteratively optimized based on the fusion loss.

6. The method according to any one of claims 1 to 4, characterized in that, Obtaining the fault training samples includes: Select non-fault training samples and first fault training samples from the historical operating data; The number of fault samples to be generated is determined based on the number of non-fault training samples, the number of the first fault training samples, and the preset sample ratio. The trained temporal generative adversarial network generates a second fault training sample, the number of which is determined based on the number of fault samples to be generated.

7. A fault detection device, characterized in that, The device includes: The acquisition module is used to acquire the operating data time series of the boiler turbine system, wherein the operating data time series includes at least the main steam temperature time series; The fault detection module is used to detect faults in the boiler turbine system based on the time series of the operating data using a preset fault detection model. The preset fault detection model is trained in advance based on fault training samples and non-fault training samples. The non-fault training samples are obtained by filtering from historical operating data. The fault training samples include a first fault training sample and a second fault training sample. The first fault training sample is obtained by filtering from the historical operating data, and the second fault training sample is generated by a trained temporal generative adversarial network.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images