A method for monitoring and early warning of wall temperature of a plasma burner based on artificial intelligence

Abstract

This invention discloses an artificial intelligence-based method for monitoring and early warning of plasma burner wall temperature, relating to the field of wall temperature monitoring technology. The method includes the following steps: dividing the plasma burner wall to obtain a monitoring zone sequence, establishing a state operation dictionary, and constructing a joint characterization framework for wall temperature operation status based on the monitoring zone sequence and the state operation dictionary; collecting historical wall temperature data, constructing a wall temperature correlation stability baseline generation model, inputting the joint characterization framework for wall temperature operation status and historical wall temperature data into the wall temperature correlation stability baseline generation model, outputting a wall temperature correlation stability baseline, and generating a wall temperature deviation description; constructing a wall temperature monitoring model, collecting real-time wall temperature data and inputting it into the wall temperature monitoring model, outputting the correlation damage degree and the ranking of contributing factors; generating a risk level based on the correlation damage degree, and generating a plasma burner wall temperature monitoring and early warning signal based on the risk level and the ranking of contributing factors.

Description

Technical Field

[0001] This invention relates to the field of wall temperature monitoring and early warning technology, specifically to an artificial intelligence-based method for monitoring and early warning of wall temperature in plasma burners. Background Technology

[0002] As a key component of core equipment such as power plant boilers and industrial heating devices, plasma burners directly reflect the stability and safety of combustion through their wall temperature. Abnormal wall temperature can lead to serious accidents such as localized ablation, material fatigue, and even combustion explosions. Therefore, accurate monitoring and early warning of wall temperature are of great practical significance for ensuring continuous equipment operation and reducing maintenance costs. With the increasing demands for combustion efficiency and safety standards in industrial production, traditional monitoring methods are no longer sufficient to meet the needs for accurate early warning under complex operating conditions. There is an urgent need to build an intelligent monitoring technology system that can adapt to the operating characteristics of burners.

[0003] Currently, the most widely used method in the field of plasma burner wall temperature monitoring is the single-point wall temperature alarm method based on a fixed threshold. This method divides the burner wall into several independent monitoring points, presets a fixed threshold for each monitoring point, and uses temperature sensors to collect wall temperature data at each point in real time. When the real-time wall temperature data at a certain monitoring point exceeds the fixed threshold, an alarm signal is directly triggered, completing the initial response to the abnormal wall temperature.

[0004] However, this single-point wall temperature alarm method based on a fixed threshold only focuses on whether the wall temperature data of a single monitoring point exceeds the limit. It cannot capture the spatial and temporal correlation of burner wall temperature, making it difficult to identify the correlation between operating status and wall temperature response, let alone locate the core cause of the abnormality. Ultimately, the alarm signals are mostly isolated prompts of single-point over-limit, which cannot assist maintenance personnel in taking targeted measures. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based method for monitoring and early warning of plasma burner wall temperature, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence, comprising the following steps: Step S1: Divide the plasma burner wall to obtain a monitoring zone sequence, establish a state operation dictionary, and construct a joint characterization framework for wall temperature operation state based on the monitoring zone sequence and the state operation dictionary; Step S2: Collect historical wall temperature data of the plasma burner, construct a wall temperature correlation stability baseline generation model based on meta-learning algorithm, input the joint characterization framework of the wall temperature operating state and the historical wall temperature data into the wall temperature correlation stability baseline generation model, output the wall temperature correlation stability baseline, and generate a wall temperature deviation description. Step S3: Construct a wall temperature monitoring model based on a graph convolutional-long short-term memory network. Train the wall temperature monitoring model using the wall temperature correlation stability baseline and wall temperature deviation description to obtain a trained wall temperature monitoring model. Collect real-time wall temperature data and input it into the trained wall temperature monitoring model. Output the correlation damage degree and cause contribution ranking. Step S4: Based on the degree of associated damage, generate a risk level for wall temperature monitoring; based on the risk level and the ranking of contributing factors, generate an early warning signal for plasma burner wall temperature monitoring.

[0007] Preferably, obtaining the monitoring partition sequence and establishing the status operation dictionary includes: Guided by the thermal risk mechanism of the burner wall, it is first divided into several monitoring zones along the axial and circumferential directions. Each monitoring zone has a clear thermal response semantic, and the collection of all these monitoring zones constitutes the monitoring zone sequence. ; After completing the spatial dimension division, the equipment operating status needs to be standardized and semantically processed to define the key event types affecting the wall temperature. Therefore, an operating status dictionary needs to be established. ; The spatial structure and operational status dictionary defined by the monitoring partition sequence P. Based on the defined event semantics, the continuous data generated in actual operation is standardized into two independent time series aligned under a unified time base, where the first independent time series is the wall temperature state vector sequence {T( ),T( ),…,T( )},in Is At any given time, the wall temperature state vector is composed of temperature readings from all M monitoring zones; the second independent time sequence is the operational state event sequence. .

[0008] Preferably, the framework for constructing the joint characterization of wall temperature operating status includes: The core operation of constructing a joint characterization framework for wall temperature operating states is to convert the wall temperature state vector sequence {T( ),T( ),…,T( The event sequence S is precisely time-aligned with the runtime event sequence S, for each event instance in the runtime event sequence S. The joint characterization framework for wall temperature operating status will automatically extract a corresponding analysis time window from the wall temperature state vector sequence. .

[0009] Preferably, the method for constructing a wall temperature correlation stability baseline generation model based on a meta-learning algorithm includes: A meta-learning algorithm is used to construct a baseline generation model for wall temperature-related stability. Wall temperature-related stability baseline generation model Taking the support set data as input, its output is the predicted wall temperature-related stability baseline. The optimization objective of the model parameter θ is to minimize the baseline of stability associated with the predicted wall temperature. Stability baseline associated with wall temperature The difference between them is determined by the composite loss function. To measure; After training through meta-learning, when faced with new equipment, only a small number of event instances from the equipment's normal operation need to be collected as a support set and input into the wall temperature correlation stability baseline generation model. This allows for the rapid generation of a suitable wall temperature-related stability baseline. .

[0010] Preferably, the description of the generated wall temperature deviation includes: Based on the established wall temperature-related stability baseline Generate a wall temperature deviation description for each event instance and calculate the analysis time window. Wall temperature response and wall temperature-related stability baseline At critical junctures Maharanobis distance ; By analyzing the distance to Mahalanobis Correlation coefficient matrix in stability baseline associated with wall temperature The differences are detected by examining the disruption of spatial coordination, and the deviations are categorized into specific descriptions.

[0011] Preferably, the wall temperature monitoring model constructed based on graph convolutional-long short-term memory network includes: In the construction of the wall temperature monitoring model, a core architecture combining graph convolutional networks and long short-term memory networks is adopted, and a three-level attention mechanism is embedded to achieve focus on key features; The wall temperature monitoring model first processes the spatial features of the wall temperature response window through a graph convolutional network layer; the adjacency matrix A of the graph convolutional network layer is constructed based on the spatial adjacency relationship between the monitoring zones. Subsequently, the wall temperature monitoring model introduces a convolutional attention module to adaptively optimize the spatial features extracted by the graph convolutional network layers. This convolutional attention module performs channel attention and spatial attention sequentially. Channel attention first compresses spatial information through global average pooling and max pooling, and then generates channel attention weight vectors through a shared multilayer perceptron. ; Spatial attention is applied to the feature map optimized by channel attention. By concatenating the results of average pooling and max pooling along the channel dimension, and then performing convolution calculations through a convolution kernel, a spatial attention weight map is generated. ; After spatial feature optimization, the sequence features are input into a Long Short-Term Memory (LSTM) network to capture temporal dynamics. The LSM network outputs the hidden state at each time step t. Furthermore, a temporal attention mechanism is introduced, assigning different temporal attention weights to the entire hidden state sequence. ; Ultimately, the comprehensive feature vector of the entire wall temperature response window All hidden states are represented based on temporal attention weights. The weighted sum.

[0012] Preferably, the degree of correlation disruption obtained from the output includes: By using a fully connected layer and a sigmoid activation function, the integrated feature vector is mapped to a scalar, representing the degree of association disruption. The calculation formula is as follows: ; in, It refers to the degree of disruption to the relationship; It is a comprehensive feature vector; and σ represents the trainable weights and bias parameters; σ is the Sigmoid function.

[0013] Preferably, the output of the cause contribution ranking includes: The evaluation of the "state event chain" refers to multiple consecutive operational state events occurring within the same analysis time window or a short adjacent period. The contribution score (C) of the state event chain C is calculated using the following formula: ; in, It is a state event chain; 'e' is an event in the state event chain. It is the moment when event e occurs; It corresponds to a specific time point. Time attention weights; It is the spatial attention weight vector that affects the monitoring partition of event e; Take the maximum value of the vector; Finally, the wall temperature monitoring model sorts all identified state event chains in descending order of their contribution scores and outputs this sorted list as the cause contribution ranking.

[0014] Preferably, the risk levels of the generated wall temperature monitoring include: Based on the degree of associated damage Risk quantification and classification are performed, with risk level determined by a comprehensive risk score R that integrates the degree of damage, duration, and type of deviation. The calculation formula is as follows: ; in, It refers to the degree of disruption to the relationship; It is the duration of destruction factor; p is the deviation type factor; and It is the normalization function; α, β, γ are the weighting coefficients; Based on the range of the comprehensive risk score R, the risk level is divided into three levels: weak, medium, and high, which correspond to different warning levels and response requirements. When the consistency index H exceeds the preset threshold, the warning for the corresponding risk level is triggered.

[0015] Preferably, the consistency index is calculated as follows: ; in, is the degree of association disruption in the q-th window; Q is the size of the criterion window; It is the threshold for determining an anomaly; I( ) is an indicator function; q represents the change in the degree of disruption of the association between adjacent windows; q is the window index.

[0016] Beneficial effects This invention provides an artificial intelligence-based method for monitoring and early warning of plasma burner wall temperature, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The artificial intelligence-based plasma burner wall temperature monitoring and early warning method extracts the general correlation features of operating status and wall temperature response from a large amount of historical stable data of mature burners of the same type through meta-learning algorithm. Then, with the help of transfer learning, it can quickly generate a wall temperature correlation stability baseline adapted to the new equipment by fine-tuning with only a small amount of stable data of the newly commissioned burner. This effectively solves the problem of the difficulty in baseline construction caused by insufficient data in the cold start stage of the new equipment, greatly shortens the baseline deployment cycle, and ensures the accuracy of the baseline and the operating characteristics of the new equipment. It also reduces the risk of baseline deviation caused by missing data and provides a reliable judgment benchmark for subsequent correlation damage identification.

[0017] (2) This AI-based plasma burner wall temperature monitoring and early warning method, through a wall temperature monitoring model constructed based on graph convolution-long short-term memory network, can deeply integrate the spatial and temporal dimensions of wall temperature monitoring: graph convolution accurately captures spatial correlation patterns by modeling the spatial correlation between wall temperature monitoring intervals; long short-term memory network effectively extracts the temporal dependence between operating status and wall temperature response, realizing in-depth mining of temporal correlation patterns. This wall temperature monitoring model breaks through the limitations of traditional methods of isolated monitoring. It is precisely by relying on this ability to deeply mine spatiotemporal correlation patterns that it can accurately identify correlation destruction phenomena such as spatial coordination anomalies or temporal response instability, and realize early identification of coking trends. This breaks through the traditional DCS's delayed alarm mode based on fixed thresholds, eliminates the coking problem from the root, and avoids nozzle burnout and equipment shutdown caused by coking. At the same time, the model combines spatiotemporal features to achieve accurate positioning of core causes, providing maintenance personnel with a clear direction for fault tracing.

[0018] (3) This AI-based plasma burner wall temperature monitoring and early warning method incorporates an attention mechanism into the wall temperature monitoring model. The convolutional attention module automatically filters key input features, prioritizing the weights of features that significantly affect associated damage and suppressing interference from irrelevant features. The temporal attention module focuses on key time segments where anomalies occur, amplifying the feature signals of abnormal association patterns. The introduction of the dual attention mechanism enables the wall temperature monitoring model to focus on core information more efficiently and reduce noise interference. This not only improves the accuracy and robustness of associated damage identification but also reduces the false alarm rate. More importantly, this focusing ability makes the cause contribution ranking results output in step S3 more targeted and interpretable. Based on this ranking, maintenance personnel can quickly identify the core operating parameters that cause abnormal wall temperatures and reverse the abnormal operating mode to quickly correct the combustion conditions. This achieves a shift from experience-based trial and error after traditional alarms to data-driven precise correction, significantly enhancing the practical value of the wall temperature monitoring model in complex industrial conditions.

[0019] (4) This AI-based plasma burner wall temperature monitoring and early warning method can effectively prevent automatic shutdown of the plasma generator due to excessively high wall temperature by predicting the wall temperature rise trend in advance and intervening in a timely manner. This is particularly important for units that frequently participate in deep peak shaving and need to operate at less than 30% of rated load for extended periods. It significantly improves the boiler's stable combustion safety under low load conditions, eliminates the risk of boiler flameout caused by plasma shutdown, and provides reliable technical support for the unit to safely and flexibly respond to the grid's deep peak shaving commands. Attached Figure Description

[0020] Figure 1 This invention proposes an artificial intelligence-based method for monitoring and early warning of plasma burner wall temperature.

[0021] Figure 2 The present invention provides a hierarchical diagram describing wall temperature deviation in a plasma burner wall temperature monitoring and early warning method based on artificial intelligence.

[0022] Figure 3 This invention presents a method for monitoring and warning the wall temperature of a plasma burner based on artificial intelligence, which generates a hierarchical diagram of the plasma burner wall temperature monitoring and warning signals. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Currently, the following problems exist in practical applications: 1. Lag in monitoring methods: Due to the hysteresis of heat conduction and the installation location of sensors, when the DCS displays an alarm for excessive wall temperature, obvious coking often already exists in the burner nozzle area. This lag of "coking first, alarm later" makes it impossible for maintenance personnel to intervene in the early stages of overheating. Once coking forms, it can affect combustion efficiency at best, and cause nozzle burnout or equipment failure at worst.

[0025] 2. Direct Threat to Low-Load Stable Combustion Safety: When the wall temperature rises sharply, the plasma generator often automatically shuts down due to interlocking protection logic to ensure equipment safety. For units that frequently participate in deep peak shaving and rely on the plasma system to maintain low-load combustion, the unexpected shutdown of the plasma system will directly disrupt the boiler's combustion stability, and in severe cases, may even lead to boiler flameout or unplanned shutdown, posing a significant challenge to grid security and unit economy.

[0026] 3. Lack of actionable corrective guidance: Traditional alarms only output an isolated "temperature over-limit" signal, which cannot inform maintenance personnel of the specific reasons for the over-temperature (such as improper primary air ratio, coal powder fineness deviation, plasma power fluctuation, etc.). This causes on-site personnel to make blind adjustments based on experience after the alarm, often missing the best intervention window.

[0027] This embodiment provides a technical solution to the above problems: an artificial intelligence-based method for monitoring and warning the wall temperature of a plasma burner. Specifically, the following artificial intelligence-based method for monitoring and warning the wall temperature of a plasma burner is provided; please refer to [link / reference]. Figures 1-3 The method includes the following steps: Step S1: Divide the plasma burner wall to obtain a monitoring partition sequence, establish a state operation dictionary, and construct a joint characterization framework for wall temperature operation state based on the monitoring partition sequence and the state operation dictionary.

[0028] The core objective of this step is to construct a unified joint characterization framework for wall temperature monitoring of plasma burners. This framework effectively correlates spatial temperature monitoring points with temporal equipment operation events, laying the foundation for subsequent analysis of wall temperature response patterns to operational status. The construction of the joint characterization framework for wall temperature operation begins with the spatial structural division of the burner wall.

[0029] The thermal risk mechanism refers to the spatiotemporally correlated abnormal evolution of temperature in different regions of the burner wall during dynamic operation, triggered by events. This abnormal evolution is the fundamental physical process leading to localized overheating, thermal stress concentration, and ultimately, failure of the wall. This thermal risk mechanism is based on physical laws derived from historical operating data and analysis of typical failure cases.

[0030] Guided by the thermal risk mechanism of the burner wall, it is first divided into several monitoring zones along the axial and circumferential directions. Each monitoring zone has a clear thermal response semantic, which refers to the coordinated temperature change pattern exhibited by that zone in time and space when various events occur in the burner (such as ignition, load change, soot blowing, etc.). Specifically, spatially, the temperature of each monitoring zone does not change in isolation, but exhibits a specific coordinated pattern based on its physical location and thermal response semantic (e.g., the "flame root zone" heats up first, and the "downstream zone" responds later). Temporally, this spatial coordinated pattern is triggered by specific operating events (such as ignition, load change) and follows a predictable temporal evolution path (e.g., temperature waves propagate sequentially from some zones to others). The collection of all these monitoring zones constitutes the monitoring zone sequence. Its mathematical expression is as follows: ; Where M is the total number of monitoring zones, which is determined by both the burner structural dimensions and the monitoring accuracy requirements; Represents the i-th monitoring partition in the monitoring partition sequence; each monitoring partition (i=1,2,…,M) is determined by its axial position Circumferential position Together with the preset thermal response semantics (such as "anti-mapping flow heating effect" or "reflecting combustion feedback effect"), i is the monitoring partition index.

[0031] After completing the spatial dimension division, the equipment operating status needs to be standardized and semantically processed to define the key event types that can affect the wall temperature (i.e., event types whose triggering has been proven to directly cause changes in the spatiotemporal response pattern of the wall temperature and have been incorporated into the standardized monitoring system). This is achieved by establishing an operating status dictionary, which contains all monitored standard operating event types. The mathematical expression is: ; Wherein, K is the total number of standard event types, derived from the summarization of all controllable operations and observable state changes of the equipment; Represents the k-th event type in the dictionary; each event type in the dictionary Each of the (k=1,2,…,K) pairs has a clearly defined name and a preset duration of influence Δ. (Based on process knowledge or historical data analysis) and the spatial range of the main impact, where the spatial range of the main impact refers to the set of monitoring zones that are most likely to be significantly affected by a certain type of standard event, which can be determined in advance based on process mechanism and historical data analysis; k is the event index.

[0032] The spatial structure and operational status dictionary defined by the monitoring partition sequence P. Based on the defined event semantics, the continuous data generated in actual operation is standardized into two independent time series aligned under a unified time base. To establish accurate causal relationships, a unified time base is required, which is achieved by establishing an artificial timing reference frame. The first independent time series is a wall temperature state vector sequence, whose dimension, order, and physical meaning are uniquely determined by P. The data originates from temperature sensors installed on each monitoring zone. At sampling time The mathematical expression is: ; in, Is At any given time, the wall temperature state vector is composed of the temperature readings of all M monitoring zones; Is At any time, monitor the zone The measured temperature value; n is the index of the sampling time interval. Therefore, from time... arrive The wall temperature state vector sequence can be represented as {T( ),T( ),…,T( )}.

[0033] Meanwhile, the second independent time sequence is the running state event sequence. It records a list of discrete events sorted by occurrence time, and a sequence of running status events. The definition is as follows: ; Where S represents the sequence of operational status events; L is the total number of event instances that occurred during the monitoring period; This represents the l-th event instance in the sequence of running status events, where l is the index of the event instance; It is the event type of the l-th event instance, derived from the runtime status dictionary. A certain event type ; It is the absolute moment when the event instance occurred, which is extracted directly from the time-series log.

[0034] The final constructed joint characterization framework for wall temperature operating state has a core operation of converting the wall temperature state vector sequence {T( ),T( ),…,T( The event sequence S is precisely time-aligned with the runtime event sequence S. For each event instance in the runtime event sequence S... The joint characterization framework for wall temperature operating status will automatically extract a corresponding analysis time window from the wall temperature state vector sequence. This analysis time window It consists of an observation period before the event occurs and a follow-up period after the event occurs, and its mathematical expression is: ; Among them, the forward length of the window With backward length It depends on the event type. Preset duration of influence Δ The hyperparameters are determined together with the typical response durations observed in historical data analysis; It is the moment the event is triggered, as determined by the runtime state dictionary. The event triggering time is determined by predefined, precisely identifiable event triggering rules (such as "a certain switch signal transition"). Once an actual signal is detected that meets the rules, a unique event triggering time will be automatically marked. .

[0035] It should be noted that the determination of the window length is the result of a combination of mechanism pre-planning and historical data analysis. First, based on the thermal mechanism of the burner and engineering experience, a theoretical influence duration Δ is pre-planned for each type of event. Meanwhile, by analyzing a large amount of historical normal data, the typical actual duration for the wall temperature response to stabilize under this type of event is statistically determined. The final window length will be the larger of the two values ​​to ensure that the entire thermal response process of the event is fully covered.

[0036] Through this joint characterization framework of wall temperature operating status, subsequent models can learn "the spatiotemporal response pattern of the monitoring partition sequence within a specific time window after a specific type of state event occurs", thereby transforming the "single-point threshold problem" of traditional wall temperature alarm into "the consistency problem of wall temperature response to operating status", laying a computable data foundation for subsequent correlation stability modeling and damage identification.

[0037] Step S2: Collect historical wall temperature data of the plasma burner, construct a wall temperature correlation stability baseline generation model based on meta-learning algorithm, input the joint characterization framework of the wall temperature operating state and the historical wall temperature data into the wall temperature correlation stability baseline generation model, output the wall temperature correlation stability baseline, and generate a wall temperature deviation description.

[0038] This step builds upon the joint characterization framework for wall temperature operating status established in step S1. It aims to utilize historical wall temperature data to transform the "response consistency" problem into a computable and measurable correlation pattern, thereby providing a discrimination benchmark for subsequent wall temperature monitoring models that relies on dynamic correlation stability rather than absolute temperature thresholds. The core of this step is to construct a meta-learning-based baseline generation model for wall temperature correlation stability. This model can automatically summarize the normal wall temperature response patterns (i.e., the wall temperature correlation stability baseline) corresponding to various operating status events defined in step S1 from historical wall temperature data and generate wall temperature deviation descriptions to quantitatively describe the degree of deviation.

[0039] Specifically, the calculation in this step is entirely based on the output of step S1. Each event instance in the event sequence S generated in step S1... Each corresponds to an analysis time window. This step first acquires historical wall temperature data, which is obtained by installing monitoring data in each monitoring zone in the monitoring zone sequence P. The temperature sensor on the device collects and stores data at a fixed frequency.

[0040] To enable the baseline generation process to quickly adapt to new equipment or operating conditions, this step employs a meta-learning algorithm to construct a wall temperature-related stability baseline generation model. The learning objective of this wall temperature-related stability baseline generation model is to be able to generate a baseline for event types using only a small amount of data. Generate its corresponding wall temperature-related stability baseline During the meta-training phase, each event type will be... The construction of the wall temperature-related stability baseline is defined as a separate task. For each task, its corresponding historical window data set is divided into a support set. and query set Wall temperature-related stability baseline generation model Taking the support set data as input, its output is the predicted wall temperature-related stability baseline. The optimization objective of the model parameter θ is to minimize the baseline stability of the predicted wall temperature correlation across all meta-training tasks. The stability baseline associated with wall temperature calculated based on a large amount of historical data from this task. The difference between them. This difference is expressed through a composite loss function. To measure this, the specific formula is as follows: ; in, It is a baseline generation model for wall temperature-related stability. Trainable parameters; It is a meta-learning task, corresponding to event type. The task of generating a baseline; It is the meta-task distribution, that is, the probability distribution of the baseline learning tasks corresponding to all different event types; It is the mathematical expectation of the tasks sampled from the meta-task distribution p(T); α and β are the weighting coefficients that balance the losses of each item. , and It is a baseline generation model for wall temperature-related stability. Predicted baseline parameters; , and It is a loss function that calculates the difference between the predicted and actual values ​​of the mean, variance, and correlation coefficient.

[0041] After training through meta-learning, when faced with a new device, it is only necessary to collect a small number of event instances (i.e., event instances) under normal operation of that device. and its analysis time window As the support set, the input wall temperature-related stability baseline generation model This allows for the rapid generation of a suitable wall temperature-related stability baseline. .

[0042] Wall temperature-related stability baseline For event types The formal representation of this is as follows, which describes the normal spatiotemporal response pattern: ; in, It is based on the event type. The baseline for wall temperature-related stability; It is the j-th key time point, located within the analysis time window. On the boundaries or feature points of each response segment within the internal division; It is the correlation coefficient matrix between the temperature changes of any two partitions in the monitored partition sequence P within a specific time interval; Is Monitor the expected temperature vector of all partitions in the partition sequence P at all times. ; yes The covariance matrix at time t, with diagonal elements as Off-diagonal elements can characterize spatial covariance, where It is the variance of the temperature across all zones, which together define the "set of permissible response patterns" in both time and space dimensions.

[0043] Finally, based on the established wall temperature-related stability baseline This step also requires generating a wall temperature deviation description for each event instance. This is done through quantitative analysis of the time window. Wall temperature response and wall temperature-related stability baseline This is achieved by measuring the distance at key time points. Maharanobis distance The calculation formula is as follows: ; in, At a critical juncture The distance to Maharanobis; It is in historical data at key time points The actual recorded wall temperature vector; It is the baseline of wall temperature-related stability. In China The expected temperature vector at time t; It is the baseline of wall temperature-related stability. In China The inverse of the covariance matrix at time t.

[0044] Meanwhile, by analyzing the distance to Maharanobis Correlation coefficient matrix in stability baseline associated with wall temperature Differences in these metrics can detect disruptions in spatial coordination. Combining these quantitative indicators, deviations can be categorized into specific descriptions such as "excessively short hysteresis" (manifested as D(t) decreasing prematurely during the hysteresis response phase), "abnormal diffusion sequence" (manifested as a spatial correlation coefficient pattern inconsistent with the baseline), or "missing recovery" (manifested as D(t) remaining persistently high during the recovery phase).

[0045] It should be noted that the analysis of the distance to Maharanobis... Correlation coefficient matrix in stability baseline associated with wall temperature The difference is determined by first calculating the Mahalanobis distance between the real-time wall temperature vector and the baseline mean vector. To determine whether the temperature distribution at the current moment has deviated significantly overall, if... If the threshold is exceeded, an abnormal signal is triggered; then, the correlation coefficient matrix in the spatial correlation structure derived from real-time data is compared with that in the wall temperature correlation stability baseline. The quantitative differences between them (e.g., calculating the angle between the principal component directions or the matrix norm distance) are used to confirm that when the difference exceeds a preset threshold, it can be determined that the spatial cooperative relationship has been broken.

[0046] These structured descriptions of wall temperature deviations, together with the baseline of wall temperature correlation stability, constitute a clear and computable input and judgment standard for the wall temperature monitoring model to perform "correlation failure identification" in the subsequent step S3, completely transforming the early warning logic from "temperature exceeding the limit" to "correlation instability".

[0047] Step S3: Construct a wall temperature monitoring model based on a graph convolutional long short-term memory network. Train the wall temperature monitoring model using the wall temperature correlation stability baseline and wall temperature deviation description to obtain a trained wall temperature monitoring model. Collect real-time wall temperature data and input it into the trained wall temperature monitoring model. Output the correlation damage degree and the ranking of the contributing factors.

[0048] This step uses the wall temperature-related stability baseline generated in step S2. Using standardized wall temperature deviation descriptions as core inputs and training objectives, a wall temperature monitoring model with a deep fusion attention mechanism is constructed based on a graph convolutional-long short-term memory network model to achieve intelligent identification and cause localization of associated damage.

[0049] In the construction of the wall temperature monitoring model, a core architecture combining Graph Convolutional Network (GCN) and Long Short-Term Memory Network (LSTM) is adopted, and a three-level attention mechanism is embedded to focus on key features. The input of the wall temperature monitoring model is directly built upon the output of the previous step: for each state event defined by step S1... and the time of occurrence The wall temperature monitoring model receives a response window of the wall temperature from real-time monitoring. As the object of analysis, the data within this wall temperature response window is the wall temperature state vector sequence T(t) defined in step S1. Simultaneously, the wall temperature monitoring model receives the data generated in step S2, corresponding to this event type. Wall temperature-related stability baseline As a reference benchmark, this wall temperature-related stability baseline includes the expected temperature vector of the normal response. Covariance Matrix Statistical characteristics, etc.

[0050] The wall temperature monitoring model first processes the spatial features of the wall temperature response window using a graph convolutional network layer. The graph convolutional network layer is constructed based on the monitoring partition set P={ defined in step S1. , ,…, Its adjacency matrix A is constructed based on the spatial adjacency relationships between monitoring partitions. For example, if the monitoring partitions are... and If they are physically adjacent, then =1, otherwise 0. The propagation formula for the graph convolutional network layer is: ; in, It is the feature matrix of all nodes in the l-th layer, with dimensions M×F (M is the number of monitoring partitions, and F is the feature dimension). It is an adjacency matrix with added self-connections. , It is the identity matrix, used to introduce self-loops, because the identity matrix... All elements on the main diagonal are 1, which can add a connecting edge from itself to each node in the graph (i.e. each monitoring partition), thus forming a "self-loop". It is an adjacency matrix The degree matrix, whose diagonal elements Calculated from the adjacency matrix, representing the number of connections for each monitoring partition; It is the trainable weight matrix of the l-th layer, which is obtained through backpropagation optimization during training; It is the ReLU activation function. This process enables the features of each monitoring zone to aggregate information from its neighboring monitoring zones, thereby capturing the spatial diffusion pattern of wall temperature.

[0051] Subsequently, the wall temperature monitoring model introduces a Convolutional Attention Module (CBAM) to adaptively optimize the spatial features extracted by the graph convolutional network layers. This CBAM performs channel attention and spatial attention sequentially. Channel attention first compresses spatial information through global average pooling and max pooling, and then generates channel attention weight vectors through a shared multilayer perceptron. The calculation formula is as follows: ; Where F is the feature map output by GCN; AvgPool and MaxPool are global average pooling and max pooling operations, respectively, used to compress the spatial dimension of the feature map; MLP is a two-layer fully connected network with shared parameters, whose weights are learned during training; σ is the Sigmoid function. The output... It is the channel attention weight vector, used to recalibrate the importance of each feature channel.

[0052] Spatial attention is applied to the feature map optimized by channel attention. By concatenating the results of average pooling and max pooling along the channel dimension, and then performing convolution calculations through a convolution kernel, a spatial attention weight map is generated. : ; Here, [;] represents a concatenation operation; It is a convolutional layer with a kernel size of k×k, whose parameters are learned during training; σ is the sigmoid function. The output shows the spatial attention weights. Spatial monitoring zones are used to emphasize those that are more critical to the analysis of the current response.

[0053] After spatial feature optimization, the sequence features are input into a Long Short-Term Memory (LSTM) network to capture temporal dynamics. The LTM network outputs the hidden state at each time step t. The complete hidden state sequence is denoted as { , ,…, Building upon this, the wall temperature monitoring model further introduces a temporal attention mechanism, assigning different temporal attention weights to the entire hidden state sequence to focus on critical time periods where abnormal responses may occur. Temporal attention weights The calculation method is as follows: ; Here, 'score' is a similarity calculation function, such as using the dot product. b is a learnable query vector obtained through training; This represents the attention weight at time step t; a larger value indicates a more critical time point. Ultimately, this results in the comprehensive feature vector of the entire wall temperature response window. All hidden states are represented based on temporal attention weights. The weighted sum, i.e. .

[0054] During the training phase, the wall temperature monitoring model uses the data generated in step S2 for supervised learning. The training data includes historical wall temperature data and the corresponding wall temperature correlation stability baseline. And a description of wall temperature deviation. A composite loss function is used to simultaneously optimize both the association destruction classification and deviation type prediction tasks. Composite Loss Function Defined as: ; in, It is the binary cross-entropy loss for association disruption classification, used to measure the difference between the model's predicted degree of association disruption and the true label; The multi-label cross-entropy loss describes the prediction of wall temperature deviation and measures the difference between the deviation type predicted by the model and the true category; λ is a hyperparameter used to balance the weights of the two losses and is determined through cross-validation. Specifically, the binary cross-entropy loss... The calculation formula is: ; Where N is the number of training samples; These are real labels (0 for normal, 1 for abnormal), derived from the deviation judgment based on Mahalanobis distance in step S2; This is the probability of associated damage predicted by the wall temperature monitoring model. Multi-label cross-entropy loss. The calculation formula is: ; in, It is the number of categories that deviate from the type (such as "too short hysteresis", "abnormal diffusion", etc.); It is the true label (0 or 1) of sample i in category c, which comes from the wall temperature deviation description classification in step S2; This is the class probability predicted by the wall temperature monitoring model. This is achieved by minimizing the composite loss function. The model parameters were optimized to accurately identify both association disruption and deviation type.

[0055] The final output layer of the wall temperature monitoring model consists of two parallel branches. The first branch, through a fully connected layer and a sigmoid activation function, maps the integrated feature vector to a scalar, namely the degree of correlation disruption. The calculation formula is as follows: ; in, It is the degree of associated damage predicted by the wall temperature monitoring model; It is the comprehensive feature vector output by the preceding network layers; and σ represents the trainable weights and bias parameters of the output branch, which are optimized through training; σ is the Sigmoid function. The larger the value of ∈[0,1], the more severe the deviation of the current real-time wall temperature response from the baseline of step S2.

[0056] The second branch is responsible for ranking the contributions of causes, and its core lies in evaluating the "state-event chain". State-event chain C = { , ,..., } refers to multiple consecutive operational state events occurring within the same analysis time window or a short adjacent time period defined in step S1 (such as a fuel injection mode switch immediately following plasma power adjustment). The wall temperature monitoring model uses its internally learned attention weights to quantify the contribution of each state event chain. The contribution score Score(C) of the state event chain C is calculated as follows: ; in, It is a state event chain, containing multiple running state events that occur consecutively within a short period of time; e is an event in the state event chain. It is the moment when event e occurs; It corresponds to a specific time point. The time attention weights are obtained from the time attention mechanism; It is the spatial attention weight vector of the monitoring partition that affects event e, obtained from the CBAM module; The maximum value of this vector represents the intensity of event e's influence on the most critical monitoring zone. A higher contribution score indicates a greater likelihood that the event chain will disrupt the correlation. Finally, the wall temperature monitoring model sorts all identified event chains in descending order of their contribution scores, outputting this sorted list as the causal contribution ranking.

[0057] Through this design, step S3 not only achieves high-precision anomaly identification, but also elevates the early warning to the "correlation mechanism diagnosis" level through interpretable attention mechanisms and state event chain analysis, providing operation and maintenance personnel with a clear direction for handling.

[0058] Step S4: Based on the degree of associated damage, generate a risk level for wall temperature monitoring; based on the risk level and the ranking of contributing factors, generate an early warning signal for plasma burner wall temperature monitoring.

[0059] This step follows the correlation destruction level output from step S3. The goal of ranking the contributions of various factors is to transform these intelligent reasoning results into actionable hierarchical early warning and closed-loop response instructions. The core of this step lies in establishing a decision-making logic with "degree of associated damage" as quantitative input and "risk level" and "intervention strategy" as output. By introducing reproducibility consistency criteria and strategy self-verification constraints, the system can effectively control false alarms and avoid response risks while ensuring early warning, ultimately achieving a safe closed loop from risk perception to control execution.

[0060] First, based on the degree of association disruption output in step S3 Risk quantification and classification are performed, with risk level determined by a comprehensive risk score R that integrates the degree of damage, duration, and type of deviation. The calculation formula is as follows: ; in, The degree of correlation disruption is the core input, which comes directly from the model output in step S3. is the duration of destruction factor, which is obtained by counting the number of consecutive abnormal response windows; p is the deviation type factor, which comes from the output vector of the multi-label classification branch in step S3, and its element values ​​represent the confidence of specific abnormal patterns such as "recovery missing". and It is a normalization function used to normalize the destruction duration factor. And the deviation type factor p is mapped to the interval [0,1], so that it is consistent with The magnitudes are matched to facilitate weighted summation; α, β, and γ are weighting coefficients, set based on historical operating data and expert experience. Based on the numerical range of the comprehensive risk score R, risks are divided into three levels: weak, medium, and high, corresponding to different warning levels and response requirements.

[0061] Specifically, the comprehensive risk score R is transformed into a graded criterion: when the comprehensive risk score R is low, it is judged as a weak risk warning, used to indicate that there may be short-term disturbances; when the comprehensive risk score R is medium, it is judged as a medium risk warning; when the comprehensive risk score R is high, it is judged as a high risk warning.

[0062] To filter out false alarms caused by occasional disturbances, this step introduces the "reproducibility consistency criterion." This criterion requires that the warning trigger must satisfy the condition that the associated destructive characteristics show a consistent or aggravating trend within a consecutive Q event triggering windows. The formula for calculating its consistency index H is: ; in, Q is the degree of association disruption of the q-th window, which is directly taken from the real-time output of step S3; Q is the criterion window size, which is the number of consecutive event windows to be checked. It is the threshold for determining an anomaly; I( ) is an indicator function; This represents the change in the degree of disruption to the association between adjacent windows; q is the window index. When the consistency index H exceeds a preset threshold, a warning for the corresponding risk level is triggered.

[0063] Finally, by integrating all information, a clear plasma burner wall temperature monitoring and early warning signal is generated and output. This signal is a structured output containing the following core elements: the final determined risk level (weak, medium, high), the most critical trigger for this warning (directly taken from the top-ranked item in the trigger contribution ranking output in step S3), and the degree of associated damage in the current window. This warning signal provides operations and maintenance personnel with clear information on the risk level, the main cause of the anomaly, and its severity, thereby supporting rapid and accurate decision-making.

[0064] To ensure the safety of the treatment, this step incorporates a "strategy self-verification constraint" mechanism: Before the strategy is executed, the system quickly simulates the predicted wall temperature response after the strategy is implemented, based on the wall temperature correlation stability baseline established in step S2, and calculates the predicted deviation from the wall temperature correlation stability baseline. If the simulation results show an increased deviation (e.g., an increase in the predicted Mahalanobis distance), the system will automatically downgrade the treatment strategy to a more conservative version or suggest a delay in execution, while continuously monitoring changes in the actual degree of correlation damage.

[0065] This technical solution divides the plasma burner wall into monitoring zones, establishes an operational status dictionary, and constructs a joint characterization framework for wall temperature operation status, achieving precise spatiotemporal data alignment. Based on a meta-learning algorithm, a suitable correlation stability baseline can be quickly generated using only a small amount of new equipment data. A wall temperature monitoring model is constructed using a graph convolutional-long short-term memory network fusion attention mechanism to deeply mine spatiotemporal correlation features, accurately identify correlation failures, and locate the core causes. Risk levels are generated based on the degree and duration of correlation failures, and false alarms are filtered through a reproducibility consistency criterion, outputting structured early warning signals. This solution effectively addresses the pain points of traditional methods, such as isolated monitoring and the difficulty of constructing small sample baselines, improving early warning accuracy and operational targeting, and ensuring the safe and stable operation of equipment.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. A method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence, characterized in that, Includes the following steps: Step S1: Divide the plasma burner wall to obtain a monitoring zone sequence, establish a state operation dictionary, and construct a joint characterization framework for wall temperature operation state based on the monitoring zone sequence and the state operation dictionary; Step S2: Collect historical wall temperature data of the plasma burner, construct a wall temperature correlation stability baseline generation model based on meta-learning algorithm, input the joint characterization framework of the wall temperature operating state and the historical wall temperature data into the wall temperature correlation stability baseline generation model, output the wall temperature correlation stability baseline, and generate a wall temperature deviation description. Step S3: Construct a wall temperature monitoring model based on a graph convolutional-long short-term memory network. Train the wall temperature monitoring model using the wall temperature correlation stability baseline and wall temperature deviation description to obtain a trained wall temperature monitoring model. Collect real-time wall temperature data and input it into the trained wall temperature monitoring model. Output the correlation damage degree and cause contribution ranking. Step S4: Based on the degree of associated damage, generate a risk level for wall temperature monitoring; based on the risk level and the ranking of contributing factors, generate an early warning signal for plasma burner wall temperature monitoring.

2. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 1, characterized in that, The process of obtaining the monitoring partition sequence and establishing a status operation dictionary includes: Guided by the thermal risk mechanism of the burner wall, it is first divided into several monitoring zones along the axial and circumferential directions. Each monitoring zone has a clear thermal response semantic, and the collection of all these monitoring zones constitutes the monitoring zone sequence. ; After completing the spatial dimension division, the equipment operating status needs to be standardized and semantically processed to define the key event types affecting the wall temperature. Therefore, an operating status dictionary needs to be established. ; The spatial structure and operational status dictionary defined by the monitoring partition sequence P. Based on the defined event semantics, the continuous data generated in actual operation is standardized into two independent time series aligned under a unified time base, where the first independent time series is the wall temperature state vector sequence {T( ),T( ),…,T( )},in Is At any given time, the wall temperature state vector is composed of temperature readings from all M monitoring zones; the second independent time sequence is the operational state event sequence. .

3. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 2, characterized in that, The framework for jointly characterizing the wall temperature operating state includes: The core operation of constructing a joint characterization framework for wall temperature operating states is to convert the wall temperature state vector sequence {T( ),T( ),…,T( The event sequence S is precisely time-aligned with the runtime event sequence S, for each event instance in the runtime event sequence S. The joint characterization framework for wall temperature operating status will automatically extract a corresponding analysis time window from the wall temperature state vector sequence. .

4. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 3, characterized in that, The method for constructing a wall temperature-related stability baseline generation model based on a meta-learning algorithm includes: A meta-learning algorithm is used to construct a baseline generation model for wall temperature-related stability. Wall temperature-related stability baseline generation model Taking the support set data as input, its output is the predicted wall temperature-related stability baseline. The optimization objective of the model parameter θ is to minimize the baseline of stability associated with the predicted wall temperature. Stability baseline associated with wall temperature The difference between them is determined by the composite loss function. To measure; After training through meta-learning, when faced with new equipment, only a small number of event instances from the equipment's normal operation need to be collected as a support set and input into the wall temperature correlation stability baseline generation model. This allows for the rapid generation of a suitable wall temperature-related stability baseline. .

5. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 4, characterized in that, The description of the generated wall temperature deviation includes: Based on the established wall temperature-related stability baseline Generate a wall temperature deviation description for each event instance and calculate the analysis time window. Wall temperature response and wall temperature-related stability baseline At critical junctures Maharanobis distance ; By analyzing the distance to Mahalanobis Correlation coefficient matrix in stability baseline associated with wall temperature The differences are detected by examining the disruption of spatial coordination, and the deviations are categorized into specific descriptions.

6. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 5, characterized in that, The wall temperature monitoring model constructed based on graph convolutional-long short-term memory network includes: In the construction of the wall temperature monitoring model, a core architecture combining graph convolutional networks and long short-term memory networks is adopted, and a three-level attention mechanism is embedded to achieve focus on key features; The wall temperature monitoring model first processes the spatial features of the wall temperature response window through a graph convolutional network layer; the adjacency matrix A of the graph convolutional network layer is constructed based on the spatial adjacency relationship between the monitoring zones. Subsequently, the wall temperature monitoring model introduces a convolutional attention module to adaptively optimize the spatial features extracted by the graph convolutional network layers. This convolutional attention module performs channel attention and spatial attention sequentially. Channel attention first compresses spatial information through global average pooling and max pooling, and then generates channel attention weight vectors through a shared multilayer perceptron. ; Spatial attention is applied to the feature map optimized by channel attention. By concatenating the results of average pooling and max pooling along the channel dimension, and then performing convolution calculations through a convolution kernel, a spatial attention weight map is generated. ; After spatial feature optimization, the sequence features are input into a Long Short-Term Memory (LSTM) network to capture temporal dynamics. The LSM network outputs the hidden state at each time step t. Furthermore, a temporal attention mechanism is introduced, assigning different temporal attention weights to the entire hidden state sequence. ; Ultimately, the comprehensive feature vector of the entire wall temperature response window All hidden states are represented based on temporal attention weights. The weighted sum.

7. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 6, characterized in that, The output yields the degree of correlation disruption, including: By using a fully connected layer and a sigmoid activation function, the integrated feature vector is mapped to a scalar, representing the degree of association disruption. The calculation formula is as follows: ； in, It refers to the degree of disruption to the association; It is a comprehensive feature vector; and σ represents the trainable weights and bias parameters; σ is the Sigmoid function.

8. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 7, characterized in that, The output yields the ranking of the contributing factors, including: The evaluation of the "state event chain" refers to multiple consecutive operational state events occurring within the same analysis time window or a short adjacent time period. The contribution score (C) of the state event chain C is calculated using the following formula: ； in, It is a state event chain; 'e' is an event in the state event chain. It is the moment when event e occurs; It corresponds to a specific time point. Time attention weights; It is the spatial attention weight vector that affects the monitoring partition of event e; Take the maximum value of the vector; Finally, the wall temperature monitoring model sorts all identified state event chains in descending order of their contribution scores and outputs this sorted list as the cause contribution ranking.

9. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 8, characterized in that, The risk levels for generating wall temperature monitoring include: Based on the degree of association damage Risk quantification and classification are performed, with risk level determined by a comprehensive risk score R that integrates the degree of damage, duration, and type of deviation. The calculation formula is as follows: ； in, It refers to the degree of disruption to the association; It is the duration of destruction factor; p is the deviation type factor; and It is the normalization function; α, β, γ are the weighting coefficients; Based on the range of the comprehensive risk score R, the risk level is divided into three levels: weak, medium, and high, which correspond to different warning levels and response requirements. When the consistency index H exceeds the preset threshold, the warning for the corresponding risk level is triggered.

10. The method for monitoring and early warning of plasma burner wall temperature based on artificial intelligence according to claim 9, characterized in that, The consistency index is calculated as follows: ； in, is the degree of association disruption in the q-th window; Q is the size of the criterion window; It is the threshold for determining an anomaly; I( ) is an indicator function; q represents the change in the degree of disruption of the association between adjacent windows; q is the window index.

Images