Industrial boiler coking prediction method and system based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based industrial boiler coking prediction method and system. The method comprises the steps of infrared thermal image time sequence acquisition, multi-spectral dynamic threshold segmentation, acoustic impedance feature extraction, multi-modal working condition fusion, space-time attention prediction modeling and adaptive threshold decision triggering. According to the method, the local threshold values are adaptively calculated on different spectral bands such as short waves, medium waves and long waves, response characteristics of all temperature zones can be considered, the false detection and missing detection risks caused by the fixed threshold values are remarkably reduced, and high-temperature hot spots are extracted more accurately and reliably; acoustic impedance characteristics obtained through calculation quantitatively reflect the structure and viscosity change of a coking layer by using attenuation and time delay information of sound waves in an ash layer, a direct and fine physical basis is provided for subsequent rheological property inference, and the defect that mechanical changes of materials are difficult to capture through pure temperature monitoring is effectively overcome.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial boiler coking prediction, and particularly relates to an industrial boiler coking prediction method and system based on artificial intelligence. BACKGROUND

[0002] During the operation of an industrial boiler, coking is a key problem that seriously affects the safe, stable and economic operation of the boiler. Coking not only reduces the thermal efficiency of the boiler and increases fuel consumption, but also can cause local overheating of the furnace and even equipment damage or safety accidents. Therefore, accurately predicting the coking trend of an industrial boiler is of great significance to ensuring the safe operation of the boiler and improving production efficiency.

[0003] At present, the industrial boiler coking prediction technology has made some progress, but there are still many deficiencies. On the one hand, in terms of temperature monitoring, the traditional method mostly uses a fixed threshold to process the furnace temperature data to identify hot spots that may have coking risks. However, the temperature distribution in the boiler furnace is complex and dynamically changes, and different temperature zones have different response characteristics to coking. The fixed threshold cannot take into account the actual situation of each temperature zone, resulting in false positives and false negatives in the extraction of high-temperature hot spots, and cannot accurately and reliably identify the real coking risk area. On the other hand, existing coking prediction methods often only rely on temperature monitoring data, ignoring the important factor of material mechanics changes during coking. The formation of the coking layer is not only related to temperature, but also closely related to the structure and viscosity changes of the ash layer. Simply relying on temperature monitoring cannot fully and deeply understand the actual state of the coking layer, and cannot provide sufficient physical basis for subsequent coking trend prediction, thereby affecting the accuracy of the prediction results.

[0004] Therefore, there is an urgent need for an industrial boiler coking prediction method and system based on artificial intelligence to solve the above problems. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art, and provides a new technical solution of an industrial boiler coking prediction method and system based on artificial intelligence.

[0006] According to a first aspect of the present application, an industrial boiler coking prediction method based on artificial intelligence is provided, comprising the following steps: performing equal-interval time-series scanning on the inner wall of the furnace to obtain a temperature time-series data set; performing local statistical feature and deep belief network adaptive adjustment segmentation processing on the temperature time-series data set to obtain a hot spot coordinate sequence; measuring the deposition layer sound velocity attenuation through pulse response in the hot spot coordinate sequence coverage area to obtain an acoustic impedance feature vector; The acoustic impedance feature vector is fused with real-time working condition parameters such as coal supply, air volume and load level to input a deep feature fusion network to obtain a fusion feature vector; The fusion feature vector is input into a prediction model combining graph convolution and time sequence attention mechanism to obtain a coking trend prediction sequence in N future sampling periods. Multi-scale deviation analysis is performed on the prediction sequence and the same load reference sequence, and the early warning threshold is adjusted through online incremental learning to obtain a hierarchical alarm trigger signal; the hierarchical alarm trigger signal is mapped and fused with the prediction sequence to generate a final coking risk level and quantitative prediction result.

[0007] Preferably, the temperature time series data set is subjected to local statistical feature and deep belief network adaptive adjustment segmentation processing to obtain a hotspot coordinate sequence, including: adopting a multi-spectral dynamic threshold segmentation algorithm to perform local statistical feature and deep belief network adaptive adjustment segmentation processing on the temperature time series data set to obtain a temperature matrix, performing local variance and local mean statistics on the temperature matrix, and combining a deep belief network adaptive optimization threshold to obtain a hotspot coordinate sequence.

[0008] Preferably, the formula of the multi-spectral dynamic threshold segmentation algorithm is: ; Wherein, is the adaptive threshold value of the mth waveband pixel (u, v) for subsequent binary segmentation; is the average temperature value of the mth waveband pixel (u, v) in a local domain with a window size of w x w; is the temperature standard deviation of the mth waveband pixel (u, v) in the same local domain; is a waveband gain coefficient obtained by deep belief network adaptive learning, used to amplify or reduce the threshold offset to adapt to the thermal distribution characteristics of different wavebands.

[0009] Preferably, the fusion feature vector is input into a prediction model combining graph convolution and time sequence attention mechanism to obtain a coking trend prediction sequence in N future sampling periods, including: the prediction model adopts Chebyshev polynomial approximation for graph convolution, and adaptively adjusts the attention weight of the time sequence attention mechanism according to the sampling period to obtain a coking trend prediction sequence in N future sampling periods.

[0010] Preferably, the formula of the Chebyshev polynomial approximation is: ; Wherein, H (l) ∈R n×dl: the first layer node feature matrix, n is the number of nodes (i.e. the number of hotspots), d1 is the feature dimension; L~ = 2L / max - I: the normalized graph Laplacian, max is the largest eigenvalue, and I is the identity matrix; Tk(L~): Chebyshev polynomial of the k-th order, defined recursively as: T0(L~) = I, T1(L~) = L~, Tk(L~) = 2L~Tk-1(L~) - Tk-2(L~); k = 1, 2, …, K dl×dl+1 : the learned weight matrix corresponding to the k-th order polynomial; K: the order of the polynomial, which determines the size of the receptive field, typically taking values between 2 and 4, balancing computational efficiency and depth of information fusion; Efficient approximation of the normalized graph Laplacian L~ using Chebyshev polynomials is used to realize multi-order neighborhood information fusion on the hotspot coordinate graph and enhance spatial structure perception.

[0011] Preferably, multi-scale deviation analysis is performed on the predicted sequence and the same load reference sequence, and the early warning threshold is adjusted online for incremental learning to obtain a hierarchical alarm trigger signal, including: using an early warning threshold online incremental learning decision maker to dynamically adjust the high, medium, and low three-level alarm thresholds based on the distribution of the latest K=5 trend indices using the Bayesian update algorithm.

[0012] Preferably, the formula of the Bayesian update algorithm is: ; Where, τi is the i (high / medium / low) level alarm threshold, which is used to map the trend index to an alarm signal; Dt is the set of coking trend index samples collected in time step t; p(τi|Dt-1) is the threshold posterior distribution of the new round after Dt; p(Dt|τi) is the likelihood estimate of the trend index falling into the corresponding level given the threshold τi, which is obtained from the empirical distribution or Gaussian approximation model; The i-th level alarm threshold τi is updated online, and the adaptive adjustment of the high, medium, and low three-level thresholds is realized by fusing the latest period trend index distribution Dt and the previous posterior, thereby improving the early warning accuracy and robustness.

[0013] Preferably, the inner wall of the furnace is scanned at equal intervals to obtain a temperature time series dataset, including: using an infrared thermal imaging sensor to scan the inner wall of the furnace at equal intervals to obtain a temperature time series dataset, and the spatial resolution of the infrared thermal imaging sensor is not less than 1024x768, and the scanning period is between 0.5 and 2 seconds.

[0014] Preferably, the acoustic impedance feature vector is fused with real-time working condition parameters such as coal supply, air volume, load level into a deep feature fusion network to obtain a fusion feature vector, including: a deep feature fusion network adopting a multi-layer Transformer structure, fusing each modal information through a cross attention mechanism to output a fusion feature vector with a dimension of 256.

[0015] The application also provides an industrial boiler coking prediction system based on artificial intelligence, comprising: An infrared thermal image time sequence acquisition module is configured to perform equal-interval time sequence scanning on the inner wall of the furnace to obtain a temperature time sequence dataset; A multi-spectral dynamic threshold segmentation module is configured to perform segmentation processing on the temperature time sequence dataset through local statistical features and deep confidence network adaptive adjustment to obtain a hotspot coordinate sequence; An acoustic impedance feature extraction module is configured to measure the deposition layer sound velocity attenuation through pulse response in the hotspot coordinate sequence coverage area to obtain an acoustic impedance feature vector; A multi-modal working condition fusion module is configured to fuse the acoustic impedance feature vector with real-time working condition parameters such as coal supply, air volume, load level into a deep feature fusion network to obtain a fusion feature vector; A spatio-temporal attention prediction modeling module is configured to input the fusion feature vector into a prediction model combining graph convolution and time sequence attention mechanism to obtain a coking trend prediction sequence in the next N sampling periods; An adaptive threshold decision triggering module is configured to perform multi-scale deviation analysis on the prediction sequence and the load reference sequence and adjust the early warning threshold through online incremental learning to obtain a hierarchical alarm triggering signal; and map and fuse the hierarchical alarm triggering signal and the prediction sequence to generate a final coking risk level and quantitative prediction result.

[0016] One technical effect of the application is: (1) The application can take into account the response characteristics of each temperature zone by adaptively calculating local thresholds in different spectral bands such as short wave, medium wave and long wave, significantly reduces the false detection and missed detection risks caused by fixed thresholds, and makes the extraction of high-temperature hotspots more accurate and reliable; the acoustic impedance feature calculated using the attenuation and time delay information of sound waves in the ash layer quantitatively reflects the structure and viscosity changes of the coking layer, provides a direct and detailed physical basis for the subsequent rheological property inference, and effectively makes up for the defect that pure temperature monitoring is difficult to capture the mechanical changes of materials.

[0017] (2) The application incorporates the cleaning effect feedback into the model iteration, continuously corrects the parameters of the segmentation, fusion and prediction network modules, makes the system gradually adapt to diversified working conditions and equipment differences in actual application, realizes the dynamic closed loop from "priori-prediction-feedback-relearning", and guarantees the continuous improvement of prediction performance in long-term operation. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flow chart of an industrial boiler coking prediction method based on artificial intelligence according to the present application; Figure 2 A block diagram of an industrial boiler coking prediction system based on artificial intelligence according to the present application. DETAILED DESCRIPTION

[0019] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. Note that the relative arrangement, numerical expressions, and numerical values of components and steps set forth in these embodiments are not limiting to the scope of the present application unless otherwise specifically stated.

[0020] Embodiments of the present application will be described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar components or components having the same or similar functions are denoted throughout by the same or similar reference numerals. The embodiments described below by reference to the accompanying drawings are exemplary and are for the purpose of explaining the present application only and are not to be construed as limiting the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without making creative efforts are within the scope of protection of the present application.

[0021] Embodiment 1: Referring to Figure 1 As shown in the drawings, an industrial boiler coking prediction method based on artificial intelligence includes: The inner wall of the furnace is scanned at equal intervals in time to obtain a temperature time series dataset; In step S1, an infrared thermal imaging sensor is used to scan the inner wall of the furnace at equal intervals in time to obtain a temperature time series dataset, the spatial resolution of the infrared thermal imaging sensor is not less than 1024x768, and the scanning period is between 0.5 and 2 seconds; Wherein, infrared thermal image time series acquisition: refers to continuous scanning of the inner wall of the furnace by an infrared imaging device at a set time interval to obtain a temperature distribution image at each time.

[0022] Temperature time series dataset: a time series matrix set composed of multiple thermal image images, providing original data for subsequent high temperature area detection; Further, device positioning: fix the infrared camera on the outside of the furnace wall or at the observation hole to ensure that the field of view covers the typical hot spot area; Parameter configuration: set the spatial resolution (such as 1024x768), frame rate (such as 25Hz), and sampling period (such as 0.5s-2s); Data acquisition: use camera SDK or industrial bus (Ethernet / IP, ModbusTCP, etc.) to receive and store the scanning results each time; Data pre-processing: Distortion correction, noise filtering (e.g. median filtering, de-extreme value) and pixel temperature calibration (based on the gray-scale-temperature mapping table provided by the camera manufacturer) are performed on the original image.

[0023] In which commercial infrared thermal imager such as FLIR series is used, combined with industrial control system; combined with FPGA / SoC real-time preprocessing; data is uploaded to the host computer or edge computing device through OPC UA, MQTT and other protocols, and the temperature time series data set is obtained.

[0024] The temperature time series data set is subjected to local statistical feature and deep belief network adaptive adjustment segmentation processing to obtain a hotspot coordinate sequence. In step S2, the temperature time series data set is subjected to local statistical feature and deep belief network adaptive adjustment segmentation processing by using a multi-spectral dynamic threshold segmentation algorithm, to obtain a temperature matrix, and the temperature matrix is subjected to local variance and local mean statistics, and combined with deep belief network adaptive optimization threshold, to obtain a hotspot coordinate sequence. In which, multi-spectral: refers to processing different wavebands in the infrared spectrum such as short wave (SWIR), medium wave (MWIR) and long wave (LWIR), in order to take into account the response differences in different temperature intervals.

[0025] Dynamic threshold segmentation: based on local statistical features and deep belief network adaptive adjustment threshold, rather than fixed threshold, to improve the accuracy of high temperature area identification.

[0026] Hotspot coordinate sequence: all abnormal high temperature pixel coordinate sets extracted after segmentation.

[0027] Further, spectral separation: the original image is separated into multiple frequency bands by infrared camera hardware or post-processing algorithm; Local statistical calculation: the mean μ and standard deviation σ are calculated within the w×w range of each pixel (u, v); Multi-spectral dynamic threshold segmentation: the formula is used to calculate; In which, is the adaptive threshold value of the mth waveband pixel (u, v) for subsequent binary segmentation; : the average temperature value of the mth waveband pixel (u, v) in the local field with window size w×w; : the temperature standard deviation of the mth waveband pixel (u, v) in the same local field; : the waveband gain coefficient learned adaptively by the deep belief network, used to amplify or reduce the threshold offset to adapt to the thermal distribution characteristics of different wavebands, in which Adaptive thresholding of temperature matrix for m-th spectral band at pixel (u, v) is calculated, which can automatically adjust with local statistical features and network learning, accurately identifying high-temperature hot spots in different bands; Parallel computing of local statistics and network inference on GPU / FPGA; Using TensorRT, ONNX Runtime, etc. to accelerate deep belief network inference; Using OpenCV and CUDA to realize image segmentation and connected domain detection; Mark the area with pixel value higher than T as a hotspot; extract and denoise to form a hotspot coordinate sequence.

[0028] Measure the sound velocity attenuation of the deposition layer in the hotspot coordinate sequence coverage area to obtain the acoustic impedance feature vector; In step S3, the acoustic impedance: the product of medium density ρ and sound velocity v Z=ρv, reflects the comprehensive influence of the ash layer on the propagation of sound waves.

[0029] Acoustic impedance feature vector: a set of impedance values measured by the sound wave conduction sensor for multiple hot spot areas, used for subsequent viscosity prediction.

[0030] Further, the sensor array arrangement: uniformly arrange multi-channel ultrasonic transducers in the hotspot coverage area; Pulse response measurement: send a narrow pulse signal and record the time delay Δt of the received signal; Sound velocity calculation: calculate the actual sound velocity v=d / Δt according to the known ash layer thickness d; Impedance calculation: combine or estimate the ash layer density ρ to obtain the impedance Z=ρv.

[0031] Among them, ultrasonic sensors (1MHz-5MHz) are selected, and high-speed ADCs (50MHz-100MHz) are used; real-time processing of echo signals on embedded MCUs / DSPs; Use wireless sensor network (WSN) to aggregate multi-point data to edge node to obtain acoustic impedance feature vector.

[0032] Fuse the acoustic impedance feature vector with real-time working condition parameters such as coal supply, air volume, and load level to input into a deep feature fusion network to obtain a fusion feature vector; In step S4, multi-modal working conditions: refer to the collection of acoustic features and multiple sensor data such as coal supply, air volume, and load level; Fusion feature vector: the comprehensive representation generated by the deep learning model after fusing each modal signal, used for more accurate spatiotemporal prediction; Among them, data standardization: normalize or Z-score standardize different dimension data (acoustic impedance, flow, power). Feature concatenation: Concatenate multiple features in channel dimension or time dimension to form initial fusion input. Deep network inference: Extract interaction information through multi-layer Transformer or cross-attention module, and output fixed-dimension vector.

[0033] Further, implemented in deep learning frameworks such as PyTorch, TensorFlow, etc.; deployed on GPU clusters or Tensor Processing Units (TPUs); Deployed on edge AI devices using model compression and quantization techniques to obtain fusion feature vectors.

[0034] Input the fusion feature vector into a prediction model combining graph convolution and time series attention mechanism to obtain a sequence of coking trend predictions for the next N sampling periods; In step S5, the prediction model uses Chebyshev polynomial approximation for graph convolution and adaptively adjusts the attention weight of the time series attention mechanism according to the sampling period to obtain a sequence of coking trend predictions for the next N sampling periods; The formula of the Chebyshev polynomial approximation is: ; Where H (l) ∈R n×dl : The l-th layer node feature matrix, n is the number of nodes (i.e. hotspots), and dl is the feature dimension; L~=2L / λmax-I: Normalized graph Laplacian, λmax is its maximum eigenvalue, and I is the identity matrix; Tk(L~): Chebyshev polynomial of order k, defined recursively as: T0(L~)=I, T1(L~)=L~, Tk(L~)=2 L~ T k-1 (L~)-T k-2 (L~); θk∈R dl×dl+1 : The learned weight matrix corresponding to the k-th order polynomial; K: Polynomial order, determines the receptive field size, typically 2-4, balancing computational efficiency and information fusion depth; Use Chebyshev polynomials to efficiently approximate the normalized graph Laplacian L~ to realize multi-order neighborhood information fusion on the hotspot coordinate graph and enhance spatial structure perception ability; Where GNN: Construct a spatial topology graph for hotspot coordinates to realize spatial neighborhood information propagation.

[0035] Temporal attention: learnable weights are assigned to different time steps to focus on the feature changes at key time points.

[0036] Prediction sequence: the estimated value sequence of the coking trend indicator in the next N sampling periods.

[0037] Graph construction: taking hotspot coordinates as nodes, and establishing edges according to spatial distance or physical connection relationship; Chebyshev convolution: using the formula ; Aggregated multi-order neighborhood information; Temporal attention: adaptive window τ is used to weight and sum the historical features to form the temporal context representation; Prediction output: generated by a fully connected layer or a sequence decoder.

[0038] Further, use DGL, PyG, etc. Graph neural network library; accelerate inference on time series dedicated hardware (such as NVIDIA Jetson Orin); containerize (Docker / Kubernetes) the model and deploy it on the cloud to obtain the prediction sequence.

[0039] Multi-scale deviation analysis is performed on the prediction sequence and the same load reference sequence, and the early warning threshold is adjusted online and incrementally learned to obtain a hierarchical alarm trigger signal; the hierarchical alarm trigger signal is mapped and fused with the prediction sequence to generate the final coking risk level and quantitative prediction result; In step S6, the early warning threshold online incremental learning decision maker dynamically adjusts the high, medium and low three-level alarm thresholds based on the distribution of the last K=5 trend indexes through the Bayesian update algorithm; The formula of the Bayesian update algorithm is: ; Where τi is the i(th) (high / medium / low) alarm threshold, which is used to map the trend index to an alarm signal.

[0040] Dt is the set of coking trend index samples collected in time step t; p(τi|Dt-1) is the posterior distribution of the threshold after Dt. p(Dt|τi) is the likelihood estimate of the trend index falling into the corresponding level given the threshold τi, which is obtained from the empirical distribution or Gaussian approximation model; The i(th) alarm threshold τi is updated online and incrementally, and the adaptive adjustment of the high, medium and low three-level thresholds is realized by fusing the latest period trend index distribution Dt and the previous posterior, which improves the early warning accuracy and robustness; Trend Index: A comprehensive index reflecting the deviation between the prediction sequence and the historical benchmark after the benchmark deviation analysis.

[0041] Hierarchical alarm trigger signal: According to the comparison result of trend index and online updating threshold τ, generate low / medium / high alarm instruction.

[0042] Further, deviation calculation For each prediction time t, calculate δt(s)=(Et-Ft(s)) / Ft(s); Where s represents different scales (seconds, minutes, hours), Et is the predicted value, and Ft(s) is the benchmark value.

[0043] The multi-scale deviation vector is composed of the summary {δt(s)}.

[0044] Threshold online update Using the latest batch of deviation vectors as data Dt, adjust the threshold distribution using the Bayesian formula or online gradient update method: ; Extract the expectation or quantile of the posterior distribution as the new threshold τi′.

[0045] Alarm signal determination: If a deviation component δt(s) exceeds τi′ for K consecutive periods, the i-th level alarm signal is generated.

[0046] Mapping fusion output Define the fusion function Rt=α (i / N)+(1-α) E^t ; Where i is the alarm level (1-3), N is the maximum level number, E^t is the normalized predicted value, and α is the weight.

[0047] Rt is the final risk level and quantitative prediction result.

[0048] Where, streaming computing and statistics Real-time deviation calculation and threshold updating are implemented in streaming frameworks such as Apache Flink, SparkStreaming, etc. Use built-in window functions (tumbling, sliding) to manage data aggregation of different scales.

[0049] Online learning library Use TensorFlow Probability, PyMC3, or self-developed lightweight modules to calculate Bayesian posterior; For gradient increments, you can use online Python learning tools such as River and scikit-multiflow.

[0050] Alarm and Convergence Services The decision-making logic is encapsulated as a microservice, which receives prediction input and outputs risk results via REST API or gRPC interface; The fusion function can be completed within the same service or it can be sent to edge devices, where the PLC / DCS performs local mapping and command issuance.

[0051] Visualization and Alerts Use monitoring platforms such as Grafana and Kibana to display RtR_tRt in real time and trigger SMS / email / DCS process alarms when the threshold is exceeded; It supports seamless integration with existing industrial protocols (OPCUA, ModbusTCP).

[0052] As can be seen from the above, by adaptively calculating the local threshold in different spectral bands such as shortwave, medium wave, and long wave, the response characteristics of each temperature region can be taken into account, significantly reducing the risk of false detection and false negative detection caused by fixed threshold, making the extraction of high temperature hotspots more accurate and reliable. By utilizing the attenuation and time delay information of sound waves in the ash layer, the calculated acoustic impedance characteristics quantitatively reflect the structural and viscosity changes of the coking layer, providing a direct and precise physical basis for subsequent inference of rheological properties, and effectively making up for the deficiency that simple temperature monitoring is difficult to capture material mechanical changes.

[0053] Example 2: Prediction of coking in industrial boilers under medium load: Operating conditions overview: Rated heat load: 50MW; Coal feed rate: 12t / h; Sampling cycle: 1s, continuous data collection for 10min; S1 Infrared Thermal Imaging Time-Sequence Acquisition Sensor: Spatial resolution 1024×768, frame rate 25Hz.

[0054] Scanning cycle: one scan every 1 second, 600 frames in 10 minutes, forming a temperature time series dataset.

[0055] Data acquisition: The data is automatically stored into 600 matrices using the camera SDK, each matrix being 1024×768 pixels in size.

[0056] S2 Multi-band Dynamic Threshold Segmentation Window size: w=31; Number of spectral bands: m=3 (long wave, medium wave, short wave).

[0057] Deep belief network training: 1000 frames of labeled hotspot images were used to learn κ(1,2,3)=1.2,1.0,0.9.

[0058] Local mean at pixel (u,v) in the first spectral segment of a certain frame =650°C, standard deviation σu,v(1)=20°C, then the threshold Tu,v(1) = 650 + 1.2 × 20 = 674°C; After binary segmentation, the hotspot coordinate sequence was extracted, totaling 1200 coordinate points.

[0059] S3 Acoustic Impedance Feature Extraction Sensor array: 8-channel ultrasonic transducer, center frequency 1MHz.

[0060] Measurement method: impulse response method, sampling rate 50MHz.

[0061] Calculation Example: For a certain hotspot area, the measured sound velocity in the sediment layer decreases from the original 340 m / s to 310 m / s, and the layer thickness is 2 mm. What is the acoustic impedance? Z=ρv=2500kg / m 3 ×310m / s=7.75×105kg / (m 2 •s) The acoustic impedance eigenvector (dimension 8×1) is obtained.

[0062] S4 Multimodal Working Condition Fusion Input: Vector with coal feed rate of 12t / h, air volume of 18000m3 / h, and load of 50MW.

[0063] Fusion Network: 3-layer Transformer, outputting a 256-dimensional fused feature vector.

[0064] C = [7.75, 7.80, ...] × 10^5, after normalization, is concatenated with the working condition as input, and the network outputs a fused feature vector.

[0065] S5 Spatiotemporal Attention Prediction Modeling Graph construction: Number of nodes n=1200, edges are connected according to spatial proximity (k=5).

[0066] The Chebyshev order K=3, and the weights {θ0,θ1,θ2,θ3} are obtained through training.

[0067] Temporal attention: Adaptive window length τ=10.

[0068] Output: The predicted coking trend sequence E={e1,…,e60} within the next N=60s.

[0069] S6 Adaptive Threshold Decision Trigger Initial threshold priors: high-level τ1~N(0.8,0.05), medium-level τ2~N(0.5,0.05), low-level τ3~N(0.2,0.05).

[0070] Sample set Dt: Sample distribution of coking trend index over the last 60 seconds.

[0071] After the Bayesian update, the posterior values ​​τ1′=0.82, τ2′=0.52, and τ3′=0.22 are obtained. The coking trend index sample > τ1′ for three consecutive sampling periods triggers an advanced alarm.

[0072] The technical effects are compared in Table 1 below: Table 1

[0073] As can be seen from the above, by deeply integrating acoustic features with multi-source operating condition signals such as coal feed rate, air volume, and heat load, and realizing mutual correction and compensation between information through the cross-attention mechanism, the coupling relationship between equipment status and operating parameters can be captured simultaneously, so that the prediction model still has strong generalization ability and robustness under complex operating conditions. Graph convolution based on Chebyshev multinomials effectively aggregates multi-order neighborhood information among spatial hotspots; the temporal attention mechanism focuses on key time windows, emphasizing moments of state abrupt change or trend acceleration. The synergy of these two mechanisms enables the model to simultaneously and accurately characterize the spatial distribution evolution and temporal dependence of the coagulation process, thereby significantly improving the predictive foresight and accuracy.

[0074] Example 3: Coking prediction under high load fluctuation conditions: Operating conditions overview: Heat load range: fluctuating between 40–60MW; Coal feed rate: 11–13 t / h; Air volume: 16000–20000 m³ / h; Sampling period: 0.5s, continuous sampling for 20min.

[0075] S1 Infrared Thermal Imaging Time-Sequence Acquisition The sensor is the same as in Example 2, with a frame rate of 50Hz, and a total of 60,000 frames in 20 minutes to form a dataset.

[0076] S2 Multi-band Dynamic Threshold Segmentation Window w=21, spectral band m=4 (add mid-to-near infrared band).

[0077] The training yielded κ(1..4) = 1.3, 1.1, 0.95, 0.9.

[0078] Calculation example: Shortband μ=620°C, σ=25°C, threshold T = 620 + 0.95 × 25 = 643.75°C; Extract hotspot coordinates B≈15000 points.

[0079] S3 Acoustic Impedance Feature Extraction 12-channel transducer with a center frequency of 1.5MHz.

[0080] The sound velocity decreased from 345 m / s to 295 m / s, the layer thickness was 3 mm, and the density was 2600 kg / m³. Z = 2600 × 295 = 7.67 × 10⁵ kg / (m²) 2 •s); The feature vector has a dimension of 12×1.

[0081] S4 Multimodal Working Condition Fusion Input feature vector and real-time coal feed, air volume, and load sequence (length 40). The Transformer has 4 layers and outputs a fused feature vector of 512 dimensions.

[0082] S5 Spatiotemporal Attention Prediction Modeling Graph nodes n≈15000, k=8; K=4; Temporal attention window τ=20; Predict the coking trend over the next 120 seconds.

[0083] S6 Adaptive Threshold Decision Trigger Priori τ1~N(0.85,0.04), τ2~N(0.55,0.04), τ3~N(0.25,0.04); After the update, τ1′=0.88, τ2′=0.58, τ3′=0.28; 3-cycle over-threshold triggering.

[0084] The technical effects are compared in Table 2 below:

[0085] As can be seen from the above, by continuously updating the latest coking trend index and historical threshold distribution using Bayesian methods, the alarm limits can be automatically fine-tuned according to operational fluctuations. This not only allows for timely detection of potential risks but also avoids frequent false alarms caused by normal fluctuations in operating conditions, thereby improving the sensitivity and stability of the early warning system. By incorporating the feedback of furnace cleaning effect into model iteration, the parameters of each network module are continuously corrected in terms of segmentation, fusion, and prediction. This allows the system to gradually adapt to diverse working conditions and equipment differences in practical applications, achieving a dynamic closed loop from "priority - prediction - feedback - relearning" and ensuring continuous improvement in prediction performance during long-term operation.

[0086] Example 4: An artificial intelligence-based industrial boiler coking prediction system includes: The infrared thermal imaging time-series acquisition module is used to perform equally spaced time-series scanning of the inner wall of the furnace to obtain a temperature time-series dataset; The multi-band dynamic threshold segmentation module is used to perform segmentation processing on temperature time series datasets by adaptive adjustment of local statistical features and deep belief networks to obtain hotspot coordinate sequences; The acoustic impedance feature extraction module is used to measure the sound velocity attenuation of the sediment layer through impulse response in the area covered by the hot spot coordinate sequence to obtain the acoustic impedance feature vector. The multimodal operating condition fusion module is used to fuse acoustic impedance feature vectors with real-time operating condition parameters such as coal feed rate, air volume, and load level into a deep feature fusion network to obtain fused feature vectors. The spatiotemporal attention prediction modeling module is used to combine the convolution of the input graph with the fused feature vector with the temporal attention mechanism to obtain the coagulation trend prediction sequence within the next N sampling periods. The adaptive threshold decision triggering module is used to perform multi-scale deviation analysis on the predicted sequence and the benchmark sequence with the same load, and to adjust the early warning threshold online by incremental learning to obtain a graded alarm triggering signal; the graded alarm triggering signal is mapped and fused with the predicted sequence to generate the final coking risk level and quantitative prediction result.

[0087] Its beneficial effects are similar to those of an AI-based industrial boiler coking prediction method, including the ability to detect coking trends in advance, reducing unexpected downtime and unplanned maintenance; with the help of intelligent boiler cleaning strategy recommendations, it can also optimize the timing and parameters of boiler cleaning, reducing energy and material waste; through multi-level adaptive and online learning, the overall operation is more stable and safe, which can extend the service life of boilers and related auxiliary equipment, and improve the economy and environmental protection of industrial production.

[0088] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting coking in industrial boilers based on artificial intelligence, characterized in that, Includes the following steps: The inner wall of the furnace was scanned at equal intervals to obtain a temperature time series dataset. The temperature time series dataset is segmented using local statistical features and a deep belief network for adaptive adjustment to obtain a hotspot coordinate sequence. The acoustic impedance characteristic vector is obtained by measuring the sound velocity attenuation of the deposition layer through impulse response in the area covered by the hot spot coordinate sequence. The acoustic impedance feature vector is fused with real-time operating parameters such as coal feed rate, air volume, and load level, and then input into a deep feature fusion network to obtain the fused feature vector. The prediction model, which combines the input graph convolution of the fused feature vector with the temporal attention mechanism, yields a coking trend prediction sequence for the next N sampling periods. Multi-scale deviation analysis is performed on the predicted sequence and the baseline sequence with the same load, and the early warning threshold is adjusted by online incremental learning to obtain a graded alarm trigger signal; The graded alarm trigger signal is mapped and fused with the prediction sequence to generate the final coking risk level and quantitative prediction result.

2. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 1, characterized in that, The segmentation process of performing local statistical features and adaptive adjustment of a deep belief network on the temperature time series dataset to obtain a hotspot coordinate sequence includes: performing local statistical features and adaptive adjustment of a deep belief network on the temperature time series dataset using a multi-spectral dynamic threshold segmentation algorithm to obtain a temperature matrix; performing local variance and local mean statistics on the temperature matrix; and combining the deep belief network to adaptively optimize the threshold to obtain a hotspot coordinate sequence.

3. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 2, characterized in that, The formula for the multi-band dynamic threshold segmentation algorithm is as follows: ； in, The adaptive threshold at pixel (u,v) in the m-th band is used for subsequent binary segmentation; The average temperature value of the m-th band pixel (u,v) within a local area of ​​window size w×w; : The temperature standard deviation of the m-th band pixel (u,v) within the same local area; : Band gain coefficients obtained through adaptive learning by a deep belief network.

4. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 1, characterized in that, The prediction model that combines the fused feature vector input graph convolution with the temporal attention mechanism to obtain the coking trend prediction sequence for the next N sampling periods includes: the prediction model uses Chebyshev polynomial approximation for graph convolution, and adaptively adjusts the attention weight of the temporal attention mechanism according to the sampling period to obtain the coking trend prediction sequence for the next N sampling periods.

5. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 4, characterized in that, The formula for the Chebyshev polynomial approximation is: ； Among them, H (l) ∈R n×dl : The feature matrix of the l-th layer nodes, where n is the number of nodes and dl is the feature dimension; L~=2L / λmax-I: The normalized graph Laplacian, where λmax is its largest eigenvalue and I is the identity matrix; Tk(L~): Chebyshev polynomial, order k, defined recursively as: T0(L~)=I, T1(L~)=L~, Tk(L~)=2L~Tk-1(L~)-Tk-2(L~); θk∈ Rdl×dl+1 : The learnable weight matrix corresponding to the k-th order polynomial; K: Polynomial order, which determines the size of the receptive field. Typical values ​​are 2–4, balancing computational efficiency and information fusion depth.

6. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 1, characterized in that, The step of performing multi-scale deviation analysis on the predicted sequence and the baseline sequence with the same load, and adjusting the early warning threshold online through incremental learning to obtain graded alarm trigger signals, includes: using an online incremental learning decision-maker for early warning thresholds based on the most recent K=5 trend index distributions, and dynamically adjusting through a Bayesian update algorithm to obtain high, medium, and low alarm thresholds.

7. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 6, characterized in that, The formula for the Bayesian update algorithm is: ； Where τi is the i-th level alarm threshold, used to map the trend index into an alarm signal; Dt is the sample set of coking trend index collected within time step t; p(τi∣Dt-1) is the threshold posterior distribution of the new round after Dt; p(Dt|τi) is the likelihood estimate of the trend index falling at the corresponding level when given a threshold τi, obtained by empirical distribution or Gaussian approximation model.

8. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 1, characterized in that, The step of performing time-series scanning of the inner wall of the furnace at equal intervals to obtain a temperature time-series dataset includes: using an infrared thermal imaging sensor to perform time-series scanning of the inner wall of the furnace at equal intervals to obtain a temperature time-series dataset, wherein the spatial resolution of the infrared thermal imaging sensor is not less than 1024×768 and the scanning period is between 0.5 and 2 seconds.

9. The method for predicting coking in industrial boilers based on artificial intelligence according to claim 1, characterized in that, The process of fusing the acoustic impedance feature vector with real-time operating parameters such as coal feed rate, air volume, and load level into a deep feature fusion network to obtain a fused feature vector includes: using a deep feature fusion network with a multi-layer Transformer structure, and outputting the fused feature vector with a dimension of 256 after fusing information from each modality through a cross-attention mechanism.

10. An artificial intelligence-based industrial boiler coking prediction system, characterized in that, include: The infrared thermal imaging time-series acquisition module is used to perform equally spaced time-series scanning of the inner wall of the furnace to obtain a temperature time-series dataset; The multi-band dynamic threshold segmentation module is used to perform segmentation processing on temperature time series datasets by adaptive adjustment of local statistical features and deep belief networks to obtain hotspot coordinate sequences; The acoustic impedance feature extraction module is used to measure the sound velocity attenuation of the sediment layer through impulse response in the area covered by the hot spot coordinate sequence to obtain the acoustic impedance feature vector. The multimodal operating condition fusion module is used to fuse the acoustic impedance feature vector with real-time operating condition parameters such as coal feed rate, air volume, and load level into a deep feature fusion network to obtain the fused feature vector. The spatiotemporal attention prediction modeling module is used to combine the convolution of the input graph with the fused feature vector with the temporal attention mechanism to obtain the coagulation trend prediction sequence within the next N sampling periods. An adaptive threshold decision triggering module is used to perform multi-scale deviation analysis on the predicted sequence and the same load benchmark sequence and adjust the early warning threshold through online incremental learning to obtain a graded alarm triggering signal. The graded alarm trigger signal is mapped and fused with the prediction sequence to generate the final coking risk level and quantitative prediction result.

Images