Transform time sequence prediction and anomaly detection method and device for boiler pipeline and storage medium

Abstract

The invention relates to the technical field of boiler pipeline anomaly detection, in particular to a boiler pipeline Transform time sequence prediction and anomaly detection method and device and a storage medium, a causal data set with causal attributes is formed through multi-source time sequence data causal preprocessing, and a boiler pipeline Transform time sequence prediction and anomaly detection algorithm is established based on time domain and frequency domain statistical characteristics and trend characteristics in the causal data set. The method comprises the following steps: generating causal features, taking the causal features as input, injecting a boiler system causal graph into a Transform architecture to generate attention bias, constructing a causal priority model injected with the attention bias, processing the real-time causal features through the trained causal priority model, and outputting a multi-step prediction result within a specified time in the future; through combination of causal logic and accurate prediction capability of a Transformer model, the change trend of key operation parameters of a boiler pipeline is pre-judged in advance, effective signals are screened through causal attention bias and a gating mechanism, and accurate identification of anomaly detection is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of boiler pipeline anomaly detection, in particular to a boiler pipeline Transformer time series prediction and anomaly detection method, device and storage medium. BACKGROUND

[0002] The boiler system has multiple variables such as fuel, air volume, temperature, pressure, flow, and flue gas interacting and time delay coupling; anomalies often spread along the chain of "combustion -> heating surface -> steam system -> safety interlock", which is easy to cause high safety risk. Pure attention and correlation methods are easily affected by mixed factors and false correlations, and it is difficult to maintain stable interpretation under working condition drift, sensor drift and data missing; time delay and topology misalignment will cause false attribution.

[0003] And form a causal diagram by injecting P&ID, control logic and mechanism knowledge, injecting bias and mask of attention, combining front door / back door adjustment and counterfactual gating, improving robustness and auditability, and emphasizing "explainable, reviewable, and accountable" in the safety field. Causal path and counterfactual evidence provide evidence for the audit chain. SUMMARY

[0004] The present application provides a boiler pipeline Transformer time series prediction and anomaly detection method, device and storage medium, which establishes a Transformer model injecting causal weights to predict data with real causal relationship, thereby solving the problems in the background art.

[0005] To achieve the above-mentioned purpose, the technical scheme of the present application is as follows: A boiler pipeline Transformer time series prediction and anomaly detection method, executed by a computer device, comprises the following steps: S1, multi-source time series data causalization preprocessing, collecting multi-modal data in the boiler pipeline through sensors, and dividing and labeling the data according to the pre-established boiler system causal diagram; then, data preprocessing is performed on the multi-modal data to form a causal data set with causal attributes; S2, constructing a causal matching feature sequence, based on the time domain and frequency domain statistical features and trend features in the causal data set, dividing multi-scale sliding windows through self-encoder and causal logic, so that the feature sequence and the data in the causal data set are causally associated, and generating causal features; S3, taking the causal features as input, injecting the boiler system causal diagram into the Transformer architecture to generate attention bias, filtering effective causal signals through the gating mechanism, and constructing a causal priority model injecting attention bias; S4, the time sequence prediction and anomaly detection output, through the completed causal priority model, processes real-time causal features, outputs multi-step prediction results within a specified future time, and generates an anomaly score, synchronously outputs an abnormal root cause to a path explanation of causal transmission and an operation suggestion.

[0006] Optionally, the specific step process of S1 is as follows: S11, through existing multi-type sensors, core multi-modal data of a boiler pipeline are collected, and the multi-modal data include vibration data, pressure data, flow data and temperature data; S12, based on a pre-established boiler system causal graph, the multi-modal data are grouped, multi-modal data existing in a direct causal chain are grouped into the same data bucket, and a causal time delay parameter between variables in the bucket is labeled; meanwhile, in combination with a sensor self-checking result and historical data integrity, a reliability of each variable data is labeled; S13, statistical outliers whose numerical values exceed 3 times of a standard deviation of a mean value and outliers in data distribution identified by a DBSCAN algorithm are screened and removed, and after the removal, the removed data are replaced by a median value or linear interpolation of a corresponding variable; S14, missing time length classification processing is performed, linear interpolation is used to fill in less than or equal to 3 sampling points, Kalman filtering prediction is used to fill in 3-10 sampling points, and more than 10 sampling points are replaced by matching a same variable data mode under a similar historical working condition; S14, for temperature and pressure variables conforming to a normal distribution, Z-score standardization processing is used, and for flow and vibration data, Min-Max normalization processing is used, so that the multi-modal data are uniformly scaled to an interval of [0, 1]; S15, PTP protocol is used for time synchronization of sensors and a data collection unit, based on a unified time reference, each piece of data is time stamped, data with time stamp disorder is removed, data with different sampling rates are uniformly resampled to the same time grid, and a preprocessed causal data set is output.

[0007] Optionally, a pre-established boiler system causal graph is established as follows: S121, based on a research target, variables strongly related to a core link are selected through expert experience, a variable and an observation mode are defined, and a basic node set of a causal graph is formed; S122, judgments of a direct causal relationship between variables by field experts are integrated, an initial directed edge is drawn, a causal flow direction of the directed edge is limited based on physical laws such as thermal engineering and fluid mechanics, and the directed edge is supplemented from a causal chain refined from historical operation cases; S123, based on the analysis of observation data with PC algorithm and FCI algorithm, identify potential weak causal relationship through conditional independence test, eliminate false correlation edges, score the causal structure with GES algorithm and NOTEARS algorithm, select the highest scoring structure, optimize the existence of edges; By human intervention key variables, record system response, verify the authenticity of causal relationship, correct algorithm bias; S124, integration and conflict processing: If the algorithm result is contradictory to the physical model or expert knowledge, the prior knowledge is used as the criterion; For high confidence new relationship discovered by the algorithm, it is included in the causal diagram and labeled with reliability; S125, for the determined causal edge, quantize the core attribute, calculate the causal delay of the nodes connected by the directed edge based on time series data, determine the correlation strength 0-1 with the average causal effect, the higher the value, the stronger the correlation; Combine the prior knowledge reliability and data support, label the reliability of each edge, and output the causal diagram of the boiler system. Optionally, the specific step process of S2 is as follows: S21, input the causal data set in S1, extract the time domain feature mean, trend slope and peak value, frequency domain feature spectrum energy and main frequency for multi-modal data respectively, and output the basic feature set; S22, input the basic feature set into the autoencoder, give high weight to the core causal chain related features combined with the causal graph of S1, strengthen the causal correlation representation, and output the deep causal features; S23, multi-scale window alignment, according to the causal delay labeled by S1, divide the sliding window according to 10 seconds, 10 minutes and 24 hours, adjust the window sliding step length, match the deep causal feature sequence with the causal delay, and output the deep causal feature with window; S24, check the feature coverage of the deep causal feature with window against the causal graph, eliminate redundant features, complete missing features, ensure that the deep causal feature with window is consistent with the causal correlation, and output the verified deep causal feature; S25, unify the feature dimension of the verified deep causal feature, and output the causal matching feature sequence.

[0008] Optionally, the specific process steps of S3 are as follows: S31, initialize the model infrastructure, input the causal matching feature sequence in S25, and determine the Transformer infrastructure parameters: Configure 8 heads of self-attention, total dimension 512, cross-modal cross-attention module, use Pre-LN architecture and GELU activation feedforward network, hidden layer dimension 2048, build model framework, and input the causal matching feature sequence as the initial input feature of the model. S32, Causal attention bias injection, based on the boiler system causal graph of S1 and the causal label of S2 features, the causal correlation strength between variables, i.e. the directed edge and the time delay parameter, is converted into attention bias. The variables with direct correlation in the causal graph are given high bias values, and the variables with no causal correlation are given low bias values. The self-attention and cross-attention calculation are embedded to guide the model to focus on effective causal signals; Where, the calculation formula of the high bias value is as follows:

[0009] Where, represents the edge weight from variable node to in the causal graph, v represents the node, and i and j are the index subscripts of the nodes, which are determined by expert priori and data-driven; is the causal time delay parameter, and the value range is set according to the process path; represents the maximum edge weight in the current window or global configuration, which is used for normalization; represents the maximum time delay; and are learning parameters, which are determined according to the characteristics of the attention mechanism; is a Sigmoid activation function, which is used to map the bias to interval stable weighting; The calculation formula of the low bias value is as follows:

[0010] Where: represents the causal correlation strength threshold, which is determined according to the statistical distribution of the edge weight of the causal graph and the domain priori; represents the buffer interval width, which is determined according to the density of the edge weight distribution and the fault tolerance demand of weak correlation under actual working conditions; represents the basic bias value, the value range is , which is determined according to the demand of the model for irrelevant signal suppression and the demand for weak effective signal reservation.

[0011] S33, Causal gating mechanism configuration, based on the signal reliability labeled in S1 and the causal correlation strength obtained from the causal matching feature sequence in S2, the causal gating mechanism is designed to calculate the gating weight. The more reliable and causally correlated features, the higher the gating weight. S34, construct a comprehensive loss function including prediction loss, detection loss and causal consistency loss; S35 uses the causal matching feature sequences of S2 to divide the training set and validation set in a 7:3 ratio, and iteratively trains the model using a comprehensive loss function, employing cosine annealing learning. When the MAPE predicted by the validation set is ≤5% and the causal consistency loss is stable, training stops, and a Transformer model with causal reasoning ability is output.

[0012] Optionally, the calculation process for the gating association strength is as follows: Let the confidence level of the variable nodes in S1 be . ∈[0,1], signal health The result of causal edge weight normalization is ; Normalization and transformation:

[0013]

[0014] Path consistency score:

[0015] in, for The representative delay, 0, used to adjust weights, set according to the model training objective; Noise and Degradation Metrics:

[0016] in, For variable nodes Noise measurement, As a measure of causal magnitude; raw gating score The calculation process is as follows:

[0017] in, These are hyperparameters used to adjust the contribution weights of each factor. ; The gating weights are calculated as follows:

[0018] in, For variables Single-node gating weights, for arrive The gating weights of causal edges, which combine the availability of the nodes at both ends with the causal relationship, are used for feature weighting.

[0019] Optionally, the specific step flow of S4 is as follows: S41, input the real-time collected boiler multi-modal data, and generate real-time causal matching feature sequences consistent with the format during model training through S2 step processing; S42, input the real-time causal matching feature sequences into the trained Transformer model, and the model outputs multi-step prediction results of future time through causal attention and gate mechanism reasoning, including point prediction values and interval prediction, to provide baseline values for anomaly detection; S43, calculate the prediction error by comparing the prediction results of S4.2 with the real-time data; combine the conditional Mahalanobis distance constructed by S1 causal graph to generate an anomaly score; if the score exceeds a preset threshold, it is determined to be abnormal, and a preliminary abnormality identification is output; The preset threshold can be set according to the historical data statistical distribution, the causal correlation strength of S1 causal graph, the boiler safe operation specification, and the optimal threshold value of model verification.

[0020] S44, based on the causal graph of S1 and the causal reasoning logic built in the model, trace the abnormal root cause to the path of causal transmission for the abnormality determined in S4.3, and generate an understandable root cause explanation; link historical fault handling cases to output executable operation suggestions; S45, output the prediction results, anomaly score, root cause explanation and operation suggestion to the control system or on-duty terminal; and store the inference data of this time.

[0021] Optionally, S4 further includes: S5, based on the anomaly score, locate the core root cause through counterfactual reasoning, periodically perform counterfactual verification on the model, output the verification result, and ensure that the response direction of the causal relationship is correct, while storing the causal evidence chain of each step of reasoning; S6, collect the execution feedback data of the operation suggestion in S4, combine the verification result in S5, update the boiler system causal graph and anomaly detection threshold; optimize the model parameters through incremental learning to dynamically adapt the causal logic to the actual working conditions.

[0022] In another aspect, the application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0023] In another aspect, the application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0024] Compared with the prior art, the application has the following beneficial effects: 1、The present application realizes the early prediction of the change trend of the key operating parameters of the boiler pipeline by combining the causal logic with the precise prediction ability of the Transformer model, achieves the active avoidance of potential safety hazards and the escape from passive response to faults, realizes the accurate identification of abnormal detection through the filtering of effective signals and the elimination of false associations by the causal attention bias and the gating mechanism, and achieves the reduction of the frequency of invalid alarms and the reduction of the burden of the on-duty personnel verification; 2、The present application realizes the rapid tracing and clear explanation of the abnormal root cause and transmission path through the causal graph of the boiler system and the causal reasoning logic built in the model, escapes from the dependence on expert experience, and greatly shortens the fault handling cycle; 3、The present application realizes the dynamic adaptation of the model to the working condition changes such as coal replacement and equipment aging through the counterfactual verification and incremental learning mechanism, achieves the need for large-scale retraining, reduces the cost of subsequent operation and maintenance, realizes the landability of the model output operation suggestion by fusing the physical law of boiler operation and the actual working condition on site, and achieves the reduction of unplanned shutdown and the guarantee of long-term stable operation of the unit. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a method step flowchart of the present application; Figure 2 is a specific step flowchart of S3 in the method of the present application; Figure 3 is a specific calculation flowchart of the gating correlation strength in S33 of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme of the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments.

[0027] The embodiments of the present application will be further described in detail below in combination with the drawings and examples. The following examples are used to illustrate the present application, but cannot be used to limit the scope of the present application.

[0028] The present application provides a boiler pipeline Transformer time series prediction and abnormal detection method, which is executed by a computer device and comprises the following steps: S1, multi-source time series data causalization preprocessing, collecting multi-modal data in the boiler pipeline through a sensor, and marking the data according to a pre-established boiler system causal graph; then, data preprocessing is performed on the multi-modal data to form a causalized data set with causal attributes; S2, construct a causal matching feature sequence, based on the time domain and frequency domain statistical features and trend features in the causal data set, divide the multi-scale sliding window through the self-encoder and the causal logic, so that the feature sequence is causally associated with the data in the causal data set, and a causal feature is generated; S3, taking the causal feature as input, injecting the boiler system causal graph into the attention bias in the Transformer architecture, filtering the effective causal signal through the gating mechanism, and constructing a causal priority model with attention bias injection; S4, time series prediction and anomaly detection output, through the completed causal priority model, process real-time causal features, output multi-step prediction results within a specified time in the future, and generate anomaly scores, simultaneously output the path explanation and operation suggestions of the abnormal root cause to the causal transmission, as shown in Figure 1 .

[0029] Embodiment: A coal-fired unit boiler piping system as a core pressure-bearing component of the unit, its operating state is directly related to the safety and stability of the unit and the energy utilization efficiency. In the long-term operation process, the system faces many technical bottlenecks: for example: after the abnormal working condition occurs, the root cause positioning is highly dependent on the experience of field experts, and there is a lack of systematic causal tracing logic, which leads to a long fault troubleshooting period and easily causes fault expansion risk; The time series prediction accuracy of key operating parameters is insufficient, it is difficult to predict potential risks in advance, and the unit operation is always in a passive state of responding to faults.

[0030] In order to solve the above technical bottlenecks and improve the safety and intelligent control level of the boiler piping operation, the time series prediction and anomaly detection work is carried out by using the technical scheme.

[0031] Preparation: Equipment configuration: 12 sets of three-axis MEMS acceleration sensors, pressure sensors, electromagnetic flowmeters, and PT100 temperature sensors are deployed inside the boiler piping, covering the core piping areas such as the boiler economizer and the superheater; Historical data: collect normal or abnormal working condition data for nearly 3 months; Causal graph construction: based on the research target, filter the variables with strong correlation through expert experience, define the variables and observation methods, and form the basic node set of the causal graph; Integrate the judgments of field experts on the direct causal relationship between variables, draw the initial directed edges, limit the causal flow direction of the directed edges based on physical laws such as thermal engineering and fluid mechanics, and refine the causal chain from historical operation cases to supplement the directed edges; Based on the analysis of observation data using PC algorithm and FCI algorithm, potential weak causal relationships are identified through conditional independence test, false correlation edges are removed, GES algorithm and NOTEARS algorithm are used to score the causal structure, and the highest scoring structure is selected to optimize the existence of edges. By artificially intervening key variables, recording system responses, verifying the authenticity of causal relationships, and correcting algorithm bias, if the algorithm results contradict physical models or expert knowledge, the prior knowledge is used as the reference. For high-confidence new relationships discovered by the algorithm, they are included in the causal graph and labeled with reliability. For the determined causal edges, the core attributes are quantified, the causal time delay of the nodes connected by the directed edges is calculated based on time series data, the correlation strength is determined as 0-1 using the average causal effect, and the higher the value, the stronger the correlation. By combining the reliability of prior knowledge and the data support, the reliability of each edge is labeled, and the causal graph of the boiler system is output. Finally, a boiler system causal graph containing multiple core variables such as feedwater flow, drum water level, and superheater wall temperature is constructed, and a causal graph containing multiple causal edges is drawn, with key causal chain time delays, such as "feedwater flow → drum water level" T = 30s, "flame center offset → wall temperature anomaly" T = 5min, correlation strength (0.7-0.95), and reliability (0.85-0.98).

[0032] Detection starts: Data collection: real-time collection of 4 types of multi-modal data such as vibration, pressure, flow, and temperature through sensors, and synchronous recording of sensor self-checking status; Causal bucketing and labeling: according to the causal graph, "feedwater flow - drum water level - superheater wall temperature" and other direct causal chain data are classified into data buckets, and the time delay parameters and data reliability of each bucket are labeled, such as normal sensor data reliability 0.95 and temporary fault sensor data reliability 0.5; Data cleaning: use the 3σ principle to remove outliers in pressure data that are more than ±3 times the standard deviation from the mean, use the DBSCAN algorithm to identify outliers in vibration data, and replace them with the median; short-term missing data (≤3 sampling points) is filled with linear interpolation, and long-term missing data, i.e., >10 sampling points, is replaced with matching historical similar working condition data; Standardization and time synchronization: temperature and pressure are standardized using Z-score, and flow and vibration are normalized to the [0,1] interval using Min-Max; time synchronization is achieved through PTP protocol, and the output causal data set is resampled to 1 step per second.

[0033] By combining causal logic with the precise prediction ability of the Transformer model, the trend of changes in key operating parameters of boiler pipelines is predicted in advance, and potential safety hazards are actively avoided, thus breaking away from passive response to failures. S2, construct a causal matching feature sequence: S21, input the causal data set in S1, and extract the time domain feature mean, trend slope and peak value, the frequency domain feature spectrum energy and main frequency for multi-modal data, and output the basic feature set; S22, input the basic feature set into the autoencoder, give high weight to the features related to the core causal chain based on the causal graph of S1, strengthen the causal correlation representation, and output the deep causal features; S23, multi-scale window alignment, according to the causal time delay labeled in S1, divide the sliding window into 10 seconds, 10 minutes and 24 hours, adjust the window sliding step, make the deep causal feature sequence match the causal time delay, and output the deep causal feature with window; S24, check the feature coverage of the deep causal feature with window against the causal graph, eliminate redundant features, complete missing features, ensure that the deep causal feature with window is consistent with the causal correlation, and output the verified deep causal feature; S25, unify the feature dimension of the verified deep causal feature, and output the causal matching feature sequence.

[0034] Through the causal attention bias and gating mechanism to filter effective signals and eliminate false associations, accurate identification of anomalies is achieved, and the frequency of invalid alarms is reduced, and the burden of the on-duty personnel is reduced; S3, causal priority Transformer model building and training, as shown in Figure 2 ; S31, initialize the model basic architecture, input the causal matching feature sequence in S25, and determine the Transformer basic architecture parameters: Configure 8 heads of self-attention, total dimension 512, cross-modal cross-attention module, use Pre-LN architecture and GELU activation feedforward network, hidden layer dimension 2048, build model framework, and input the causal matching feature sequence as the initial input feature of the model; S32, inject causal attention bias, based on the boiler system causal graph of S1 and the causal labels of S2 features, convert the causal correlation strength between variables, i.e. directed edges and time delay parameters, into attention bias, assign high bias value to variables with direct correlation in the causal graph, and assign low bias value to variables without causal correlation, embed self-attention and cross-attention calculation, and guide the model to focus on effective causal signals; Among them, the variables with high bias value The calculation formula is as follows:

[0035] Wherein, represents the edge weight from the variable node to in the causal graph, v represents the node, i and j are the index subscripts of the nodes, and the expert prior and data-driven joint determination; is a causal delay parameter, and the value range is set according to the process path; represents the maximum value of the edge weight in the current window or global configuration, and is used for normalization; represents the maximum value of the delay; and are learning parameters, which are determined according to the characteristics of the attention mechanism; is a Sigmoid activation function, which is used to map the bias to a stable weight interval; a low bias value The calculation formula is as follows:

[0036] Wherein: represents the causal correlation strength threshold, which is determined according to the statistical distribution of the edge weight of the causal graph and the field prior; represents the buffer interval width, which is determined according to the density of the edge weight distribution and the fault tolerance demand of weak correlation under actual working conditions; represents the basic bias value, and the value range is , which is determined according to the suppression demand of irrelevant signals and the reservation demand of weak effective signals of the model.

[0037] Through the causal graph of the boiler system and the built-in causal reasoning logic of the model, the rapid tracing and clear explanation of the abnormal root cause and transmission path are realized, the dependence on expert experience is eliminated, and the effect of greatly shortening the fault handling cycle is achieved; S33, causal gating mechanism configuration, based on the signal reliability marked in S1 and the causal correlation strength obtained from the causal matching feature sequence in S2, the causal gating mechanism is designed to calculate the gating weight. The more reliable and causally related the features are, the higher the gating weight is. S34, construct a comprehensive loss function including prediction loss, detection loss and causal consistency loss; S35 uses the causal matching feature sequences of S2 to divide the training set and validation set in a 7:3 ratio, and iteratively trains the model using a comprehensive loss function, employing cosine annealing learning. When the MAPE predicted by the validation set is ≤5% and the causal consistency loss is stable, training stops, and a Transformer model with causal reasoning ability is output.

[0038] like Figure 3 As shown, the calculation process for the gating association strength is as follows: Let the confidence level of the variable nodes in S1 be . ∈[0,1], signal health The result of causal edge weight normalization is ; Normalization and transformation:

[0039]

[0040] Path consistency score:

[0041] in, for The representative delay, 0, used to adjust weights, set according to the model training objective; Noise and Degradation Metrics:

[0042] in, For variable nodes Noise measurement, As a measure of causal magnitude; raw gating score The calculation process is as follows:

[0043] in, These are hyperparameters used to adjust the contribution weights of each factor. ; The gating weights are calculated as follows:

[0044] in, For variables Single-node gating weights, for arrive The gating weights of causal edges, which combine the availability of the nodes at both ends with the causal relationship, are used for feature weighting.

[0045] Through the counterfactual verification and incremental learning mechanism, dynamic adaptation of the model to changes in working conditions such as coal replacement and equipment aging is realized, so that large-scale retraining is not required, and the cost of subsequent operation and debugging is reduced. S4, time series prediction and anomaly detection output; S41, input real-time collected boiler multi-modal data, after S2 step processing, generate real-time causal matching feature sequence consistent with model training format; S42, input real-time causal matching feature sequence into trained Transformer model, model infer through causal attention and gating mechanism, output multi-step prediction result of future time, including point prediction value and interval prediction, provide baseline value for anomaly detection; S43, calculate prediction error by comparing prediction result of S4.2 with real-time data; combine conditional Mahalanobis distance constructed by S1 causal graph, fuse to generate anomaly score, score exceeding preset threshold is determined as anomaly, output preliminary anomaly identification; The preset threshold can be set according to historical data statistical distribution, causal correlation strength of S1 causal graph, boiler safe operation specification and model verification optimal threshold.

[0046] S44, based on S1 causal graph and model built-in causal inference logic, trace back to the path of causal transmission for anomaly determined by S4.3, generate understandable root cause explanation; link historical fault handling cases, output executable operation suggestion; S45, deliver prediction result, anomaly score, root cause explanation and operation suggestion to control system or duty terminal; at the same time, store this time inference data.

[0047] By fusing boiler operation physical law and actual working condition, realize the landing of model output operation suggestion, reduce unplanned shutdown and ensure long-term stable operation of the unit.

[0048] S5, based on anomaly score, locate core root cause through counterfactual reasoning, periodically perform counterfactual verification on model, output verification result, ensure correct response direction of causal relationship, at the same time, store causal evidence chain of each step of reasoning; S6, collect execution feedback data of operation suggestion in S4, combine verification result in S5, update boiler system causal graph and anomaly detection threshold; optimize model parameters through incremental learning, make causal logic dynamically adapt to actual working condition.

[0049] In another aspect, the application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute steps of the above method.

[0050] In still another aspect, the present application also discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to perform the steps of the above method.

[0051] In still another aspect, the present application also discloses a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to perform the steps of the above method.

[0052] It can be understood that the system, device and storage medium provided by the embodiments of the present application correspond to the method provided by the embodiments of the present application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding part in the above method.

[0053] In the above embodiments, the system, device and storage medium provided by the embodiments of the present application can be realized by software, hardware, firmware or any combination thereof. When realized by software, the system, device and storage medium provided by the embodiments of the present application can be realized in the form of a computer program product in whole or in part. The computer program product comprises one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the flow or function described in the embodiments of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another computer readable storage medium.

[0054] For example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media.

[0055] The available medium can be a magnetic medium (for example, floppy disk, hard disk, magnetic tape), an optical medium (for example, DVD) or a semiconductor medium (for example, Solid State Disk (SSD)) and the like.

[0056] It should be noted that in this paper, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.

[0057] Also, the term "comprising" or "including" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0058] Each of the embodiments in the specification is described in a relevant manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the description of the method embodiments.

[0059] The embodiments of the present application are given for example and description, although the embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A boiler pipeline Transformer time series prediction and anomaly detection method, characterized in that, The following steps are performed by a computer device: S1, multi-source time series data causalization preprocessing, collecting multi-modal data in the boiler pipeline through sensors, and dividing and labeling the data according to a pre-established boiler system causal diagram; Further data preprocessing is performed on the multi-modal data to form a causalized data set with causal attributes; S2, constructing a causal matching feature sequence, based on the time domain and frequency domain statistical features and trend features in the causalized data set, dividing multi-scale sliding windows through a self-encoder and causal logic, making the feature sequence correspond to the causal association of the data in the causalized data set, and generating causal features; S3, taking the causal features as input, injecting a boiler system causal diagram into a Transformer architecture to generate attention bias, filtering effective causal signals through a gating mechanism, and constructing a causal priority model with injected attention bias; S4, time series prediction and anomaly detection output, processing real-time causal features through the completed causal priority model, outputting multi-step prediction results within a specified time in the future, and generating an anomaly score, synchronously outputting an abnormal root cause to a path explanation of causal transmission and operation suggestions.

2. The boiler pipeline transformer time series forecasting and anomaly detection method of claim 1, wherein: The specific step process of S1 is as follows: S11, collecting core multi-modal data of the boiler pipeline through existing multiple types of sensors, the multi-modal data including vibration data, pressure data, flow data and temperature data; S12, grouping the multi-modal data based on the pre-established boiler system causal diagram, grouping the multi-modal data with direct causal chains into the same data bucket, and labeling the causal time delay parameters between variables in the bucket; at the same time, combining the sensor self-checking results and the historical data integrity to label the reliability of each variable data; S13, filtering out statistical outliers whose values exceed 3 times the standard deviation of the mean and outliers in the data distribution identified by the DBSCAN algorithm, and replacing the filtered data with the median or linear interpolation of the corresponding variable; S14, processing according to missing time length, less than or equal to 3 sampling points are filled with linear interpolation, 3-10 sampling points are filled with Kalman filter prediction, and more than 10 sampling points are replaced by matching the same variable data pattern under similar working conditions in history; S14, for variables with normally distributed temperature and pressure, Z-score standardization is used, and for variables with fixed value range of flow and vibration data, Min-Max normalization is used, so that the multi-modal data are uniformly scaled to the [0, 1] interval; S15, PTP protocol is used for time synchronization of sensors and data acquisition units, based on a unified time reference, each data is time stamped, time stamped data is removed, data with different sampling rates is uniformly resampled to the same time grid, and the preprocessed causalized data set is output.

3. The boiler tube Transformer time series forecasting and anomaly detection method of claim 2, wherein: The pre-established boiler system causal diagram is established as follows: S121, based on the research target, variables strongly related to the core link are selected through expert experience, variables and observation methods are defined, and a basic node set of the causal diagram is formed; S122, integrate the judgment of the direct causal relationship between variables by field experts, draw the initial directed edge, limit the causal flow direction of the directed edge based on physical laws such as thermodynamics and fluid mechanics, and supplement the directed edge with causal chains refined from historical operation cases; S123, based on the PC algorithm and the FCI algorithm, analyze the observation data, identify potential weak causal relationships through conditional independence test, remove false correlation edges, score the causal structure using the GES algorithm and the NOTEARS algorithm, and select the highest scoring structure to optimize the existence of edges; Verify the authenticity of the causal relationship by manually intervening in the key variables, recording the system response, and correcting the algorithm bias; S124, integration and conflict processing: If the algorithm result conflicts with the physical model or expert knowledge, the prior knowledge is used as the reference; For high-confidence new relationships discovered by the algorithm, they are included in the causal graph and labeled with reliability; S125, for the determined causal edges, quantize the core attributes, calculate the causal time delay of the nodes connected by the directed edges based on time series data, determine the correlation strength 0-1 using the average causal effect, and the higher the value, the stronger the correlation; Combine the prior knowledge reliability and data support to label the reliability of each edge, and output the boiler system causal graph.

4. The boiler tube Transformer time series forecasting and anomaly detection method of claim 3, wherein: The specific step process of S2 is as follows: S21, input the causal data set in S1, extract the time domain feature mean, trend slope and peak value, frequency domain feature spectrum energy and main frequency for multi-modal data respectively, and output the basic feature set; S22, input the basic feature set into the autoencoder, give high weight to the features related to the core causal chain based on the causal graph in S1, strengthen the causal correlation representation, and output the deep causal features; S23, multi-scale window alignment, according to the causal time delay labeled in S1, divide the sliding window according to 10 seconds, 10 minutes and 24 hours, adjust the window sliding step to match the deep causal feature sequence with the causal time delay, and output the deep causal features with windows; S24, check the feature coverage of the deep causal features with windows against the causal graph, remove redundant features, complete missing features, and ensure that the deep causal features with windows are consistent with the causal correlation, and output the verified deep causal features; S25, unify the feature dimension of the verified deep causal features, and output the causal matching feature sequence.

5. The boiler tube Transformer time series forecasting and anomaly detection method of claim 4, wherein: The specific process steps of S3 are as follows: S31, initialize the model infrastructure, input the causal matching feature sequence in S25, and determine the Transformer infrastructure parameters: Configure 8 heads of self-attention, total dimension 512, cross-modal cross-attention module, use Pre-LN architecture and GELU activated feedforward network, hidden layer dimension 2048, build the model framework, and input the causal matching feature sequence as the initial input feature of the model; S32, causal attention bias injection, based on the boiler system causal graph of S1 and the causal label of S2 features, the causal correlation strength between variables, i.e. directed edges and time delay parameters, is converted into attention bias, variables with direct correlation in the causal graph are assigned high bias values, and variables without causal correlation are assigned low bias values, embedded in self-attention and cross-attention calculation, guiding the model to focus on effective causal signals; wherein the calculation formula for giving a high bias value is as follows: wherein, represents the edge weight from variable node to in the causal graph, v represents the node, i and j are the index subscripts of the nodes, and the expert prior and data-driven joint determination is performed; is a causal delay parameter, and the value range is set according to the process path. denotes the maximum edge weight in the current window or global configuration for normalization; denotes the maximum delay; and is a learning parameter determined according to the characteristics of the attention mechanism. Sigmoid activation function for mapping bias to Interval stable weighting; Assigning low bias values The calculation formula is as follows: wherein: represents a causal correlation strength threshold, which is determined according to a distribution of edge weights of the causal graph and a priori in the field; represents a buffer interval width, which is determined according to a density of the distribution of edge weights and a fault tolerance requirement for weak correlation in an actual working condition; represents a basic bias value, and the value range is , which is determined according to a suppression requirement of irrelevant signals and a reservation requirement of weak effective signals of the model; S33, causal gating mechanism configuration, based on the signal reliability labeled in S1 and the causal correlation strength obtained from the causal matching feature sequence in S2, design the causal gating mechanism to calculate the gating weight, the more reliable and causally correlated features, the higher the gating weight; S34, construct a comprehensive loss function including prediction loss, detection loss and causal consistency loss; S35, divide the training set and validation set in a ratio of 7 to 3 using the causal matching feature sequence of S2, and iteratively train the model using the comprehensive loss function, and use the cosine annealing learning method for training; when the validation set prediction MAPE≤5%, the causal consistency loss is stable, stop training, and output the Transformer model with causal reasoning capability.

6. The boiler pipeline transformer time series forecasting and anomaly detection method of claim 5, wherein: The calculation process of the gating correlation strength is as follows: Let the credibility of the variable nodes in S1 be ∈ [0, 1], signal health The result of the causal edge weight normalization is ; Normalization and transformation: Path consistency Score: wherein, is a representative latency, 0, is an adjustment weight, set according to the model training objective; Noise and degradation metrics: wherein is a noise metric for the variable node is a noise metric for the variable node is a causal magnitude metric Gating raw scores The calculation proceeds as follows: wherein, is a hyperparameter used to adjust the contribution weight of each factor, ; The calculation of the gating weight is as follows: wherein, is a variable single-node gating weight, is to causal edge gating weight, fusing the availability and causal relevance of both end nodes for feature weighting.

7. The boiler pipeline transformer time series forecasting and anomaly detection method of claim 6, wherein: The specific step process of S4 is as follows: S41, input the real-time collected boiler multi-modal data, process it through S2 step, and generate real-time causal matching feature sequence consistent with the format during model training; S42, input the real-time causal matching feature sequence into the trained Transformer model, and the model reasons through the causal attention and gating mechanism, and outputs the multi-step prediction result of the future time, including point prediction value and interval prediction, which provides a benchmark value for anomaly detection; S43, calculate the prediction error by comparing the prediction result of S4.2 with the real-time data; combine the conditional Mahalanobis distance constructed by S1 causal graph to generate an anomaly score; if the score exceeds the preset threshold, it is determined to be abnormal, and the preliminary abnormality is output; The preset threshold can be set according to the historical data statistical distribution, the causal correlation strength of S1 causal graph, the boiler safe operation specification and the optimal threshold value of model verification; S44, based on the causal graph of S1 and the causal reasoning logic built into the model, trace the abnormal root cause to the path of causal transmission for the anomaly determined in S4.3, and generate an understandable root cause explanation; link historical fault handling cases to output executable operation suggestions; S45, the prediction result, anomaly score, root cause explanation and operation suggestion are sent to the control system or on-duty terminal; at the same time, the reasoning data of this time is stored.

8. The boiler pipeline transformer time series forecasting and anomaly detection method of claim 1, wherein: S4 also includes: S5, based on the anomaly score, locate the core root cause through counterfactual reasoning, periodically perform counterfactual verification on the model, output the verification result, ensure that its response direction to the causal relationship is correct, and at the same time store the causal evidence chain of each step of reasoning; S6, collecting the execution feedback data of the operation suggestion in S4, combining the checking result in S5, updating the boiler system causal diagram and the abnormality detection threshold; optimizing the model parameters through incremental learning, so that the causal logic dynamically adapts to the actual working condition.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program, when executed by a processor, causes the processor to perform the steps of the method of any one of claims 1 to 8. 10.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-9. The computer program, when executed by the processor, causes the processor to perform the steps of the method of any one of claims 1 to 8.

Images