Lime kiln monitoring system based on multi-source data dynamic fusion

Abstract

The invention discloses a lime kiln monitoring system based on multi-source data dynamic fusion, which is characterized by comprising a thermal imager, an infrared sensor, a pressure sensor, a gas analyzer and an edge computing node, the edge computing node is configured with a Raspberry Pi platform and an LSTM neural network accelerator and is used for executing multi-source data preprocessing and feature compression, and the Raspberry Pi platform and the LSTM neural network accelerator are connected with the thermal imager. The system also comprises a working condition analysis system, and the working condition analysis system comprises a ring formation prediction model based on ST-GCN, a combustion optimization model based on DDPG, and an anomaly detection model based on an isolated forest and automatic encoder hybrid model. The invention aims to solve the limitation of a traditional monitoring system, realize real-time monitoring and accurate control, improve the equipment health management and predictive maintenance capability, optimize waste heat recovery and energy utilization, and enhance man-machine interaction and operation convenience.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial process monitoring and intelligent control, and particularly relates to a lime kiln monitoring system based on dynamic fusion of multi-source data. BACKGROUND

[0002] As the core equipment of the metallurgical, chemical, building material and other industries, the running state of the lime kiln directly affects the product quality, energy consumption and environmental emissions. The traditional monitoring system generally adopts single parameter threshold alarm or simple data superposition display mode, which has the problems of low data utilization rate, control lag, and untimely fault discovery. With the development of intelligent manufacturing technology, the limitations of the traditional monitoring system are increasingly prominent, which is specifically manifested in the following technical bottlenecks:

[0003] 1. Insufficient multi-source data fusion and processing capacity.

[0004] The traditional lime kiln monitoring system usually adopts independent sensors to collect temperature, pressure, gas concentration and other parameters, and there is a lack of deep data interaction between each subsystem. For example, the thermocouple can only obtain local temperature point data and cannot reflect the overall temperature field distribution of the kiln; the gas analyzer only monitors the conventional indicators such as CO and CO2, and ignores the real-time tracking of abnormal emissions such as SO2. This "data island" phenomenon leads to the following problems:

[0005] Lack of data heterogeneity processing: different sensors (such as analog signals of thermocouples and digital signals of gas analyzers) have differences in sampling rate, dimension and accuracy, and the traditional system lacks a unified data preprocessing mechanism, which directly transmitted to the central server can cause communication congestion and calculation delay.

[0006] Difficulty in extracting dynamic characteristics: the kiln operation has strong nonlinearity and time-varying characteristics (such as a circle formation period of up to 72 hours), and the traditional static threshold monitoring cannot capture dynamic characteristics such as temperature gradient changes and pressure fluctuation frequencies, resulting in a 8-12 hour lag in circle prediction.

[0007] Weak edge computing capability: most systems rely on cloud for data fusion analysis, but the network bandwidth in industrial field is limited (usually ≤10 Mbps), and transmission of raw data (such as 640x480 pixel matrix of thermal imager) requires a large amount of bandwidth, resulting in a control command delay of 3-5 seconds, which cannot meet the real-time optimization requirements.

[0008] 2. Lack of predictive maintenance and equipment health management.

[0009] The traditional monitoring system adopts a regular maintenance mode, which has the risk of "over-repair" or "repair failure":

[0010] Single vibration monitoring means: only install acceleration sensor to monitor the axial vibration of the kiln body, ignore the radial vibration (accounting for 60%) and torsional vibration (sensitive to gear wear). The measured data of a certain steel plant shows that the traditional system has a false negative rate of 32% for gear box failure.

[0011] Insufficient acoustic feature analysis: abnormal combustion can produce noise in a specific frequency band (such as 2-4kHz frequency band related to ring forming), but the traditional system does not integrate acoustic sensors, resulting in an alarm triggered when the ring thickness is >50mm, and at this time the cleaning needs to be stopped for 48 hours.

[0012] In summary, the traditional lime kiln monitoring system has significant shortcomings in data fusion, control accuracy, equipment management, etc., and cannot meet the needs of modern industry for energy saving and consumption reduction, intelligent operation and maintenance. The present application aims to solve the above technical bottlenecks by means of multi-source data dynamic fusion and predictive maintenance, and provides a complete solution for the intelligent upgrading of lime kiln. SUMMARY

[0013] The purpose of the present application is to provide a lime kiln monitoring system based on multi-source data dynamic fusion, aiming to solve the limitations of traditional monitoring systems, achieve real-time monitoring and accurate control, improve equipment health management and predictive maintenance capabilities, optimize waste heat recovery and energy utilization, and enhance human-machine interaction and operation convenience.

[0014] The technical solution of the present application is as follows:

[0015] A lime kiln monitoring system based on multi-source data dynamic fusion is provided, which comprises a perception network formed by deploying a thermal imager on the outer side of the furnace wall, setting an infrared sensor on the top of the combustion chamber, setting a pressure sensor on the inlet / outlet air port, and setting a gas analyzer on the flue. The edge computing node uses Raspberry Pi 4B+FPGA acceleration card, which carries LSTM neural network accelerator to complete time synchronization, dimension unification and outlier elimination of sensor data through edge computing node. The LSTM accelerator compresses the original data dimension to 1 / 20 of the original size, and can retain more than 95% of the key features. The edge computing node further comprises a thermal imaging manifold learning unit, which uses t-SNE algorithm to reduce the 640x480 temperature matrix collected by the thermal imager to 16-dimensional manifold features to retain the topological structure of temperature gradient. The steps are as follows:

[0016] S1. Data preprocessing, data block and feature extraction, dividing the temperature matrix into local blocks, extracting feature vectors containing temperature values and gradients to capture local topological structure:

[0017] S1.1 Block division: divide the 640x480 matrix into smaller blocks (such as 16x16), resulting in 40x30=1200 blocks.

[0018] S1.2 Calculate Gradient: Use Sobel operator to calculate the horizontal and vertical gradient of each pixel, reflecting the direction and intensity of temperature change.

[0019] S1.3 Extract Features: Calculate the following features for each block:

[0020] Mean Temperature: The mean of all pixel temperatures within the block;

[0021] Temperature Standard Deviation: The degree of dispersion of temperature values within the block;

[0022] Mean Horizontal Gradient: The mean of horizontal direction gradients within the block (reflecting the horizontal temperature change);

[0023] Mean Vertical Gradient: The mean of vertical direction gradients within the block (reflecting the vertical temperature change);

[0024] S2. Apply t-SNE Dimensionality Reduction: Input the extracted features into the t-SNE model, reducing the dimension to a 16-dimensional manifold space:

[0025] The following are the configurations of key parameters and their roles:

[0026] n_components = 16: Indicates that the target output dimension is 16-dimensional, i.e., mapping high-dimensional data to a 16-dimensional space;

[0027] perplexity = 30-50: Adjusted according to data size, the larger the value, the more global structure is concerned, the smaller the value, the more local structure is concerned;

[0028] learning_rate = 200-500: To avoid oscillation phenomenon in the optimization process;

[0029] n_iter = 1000-5000: Iteration number is set between 1000 and 5000, ensuring that the model can fully converge;

[0030] S3. Verify Topology Preservation Effect:

[0031] Verify whether the temperature gradient structure is preserved after dimensionality reduction by the following methods:

[0032] Method 1: Reverse Visualization

[0033] Further reduce the 16-dimensional features to 2D through PCA or UMAP, and observe whether they present a pattern consistent with the original temperature gradient.

[0034] Method 2: Similarity Analysis

[0035] Calculate the similarity ranking of high-dimensional features (original blocks) and low-dimensional features, and check whether the local neighborhood is consistent.

[0036] Also includes working condition analysis system, the working condition analysis system includes ST-GCN-based ring forming prediction model, DDPG-based combustion optimization model and isolated forest + automatic encoder hybrid model-based anomaly detection model, wherein:

[0037] The thermal imaging local entropy feature, the pressure fluctuation power spectral density and the gas concentration rate of change are input into the ring forming prediction model, and the ring forming position and ring forming thickness prediction are output, and the implementation method is:

[0038] 1. Input feature processing

[0039] Thermal imaging local entropy feature: calculate the Shannon entropy of local temperature field through sliding window, quantify the degree of disorder,

[0040] H = -Σ i p i log p i

[0041] Pi: probability density of temperature interval i

[0042] Pressure fluctuation power spectral density: extract the main frequency band energy ratio after FFT transformation

[0043]

[0044] Gas concentration rate of change: CO / CO2 concentration difference term

[0045]

[0046] 2. Construction of space-time graph

[0047] Spatial graph: divide the kiln body into N nodes, and the edge weight is determined by the thermal conductivity coefficient k ij

[0048]

[0049] σ: temperature difference sensitivity parameter

[0050] Time edge: connect the hidden state of the same node at different times

[0051] 3. ST-GCN layer propagation formula

[0052] Graph convolution operation:

[0053]

[0054] A k : order adjacency matrix, D k : degree matrix, Learnable weight

[0055] ​Temporal convolution: causal convolution to handle temporal dependencies

[0056] h t = tanh(W*X t-T+1：t +b)

[0057] 4. Output layer

[0058] Position prediction: fully connected layer outputs 3D coordinates (x, y, z)

[0059]

[0060] Thickness prediction: sigmoid activation function constrains thickness range

[0061]

[0062] 5. Loss function

[0063] Joint loss function:

[0064]

[0065] λ1, λ2: task weights, λ3: regularization coefficient

[0066] Input infrared flame pulsation frequency, O2 / CO concentration ratio, furnace pressure gradient into the combustion optimization model, output the best air coefficient and fuel flow recommended value, the implementation method is:

[0067] 1. State space definition:

[0068] State vector:

[0069] F IR : infrared flame pulsation main frequency (Hz)

[0070] r O2 / CO : oxygen-carbon ratio measured value

[0071] Furnace pressure gradient (Pa / m)

[0072] 2. Action space definition

[0073] Continuous action: air coefficient α and fuel flow Fa = [α, F] ∈ [α min , α max ] × [F min , F max ]

[0074] 3. Strategy network (Actor)

[0075] Two-layer fully connected network outputs action mean

[0076] μ(s|θ μ ) = tanh(W2·ReLU(W1s + b1) + b2)

[0077] 4. Critic

[0078] Q-value estimation adopts a double stream structure:

[0079] Q(s, a) = ReLU(W 1s s + W 1a a + b1)·W2 + b2

[0080] 5. Reward function design

[0081] Multi-objective reward:

[0082]

[0083] ηth: thermal efficiency, NO x : nitrogen oxide emission concentration

[0084] w i : weight coefficient, Δ: action change amount penalty

[0085] 6. Training process

[0086] Experience replay buffer storage (S t , a t , r t , s t+1 )

[0087] Target network soft update:

[0088] θ' <- τθ + (1-τ)θ'

[0089] A multi-sensor fusion vector is input into an anomaly detection model, and an abnormal type is output, the abnormal type including leakage, coking, flameout and confidence, and a method is implemented:

[0090] 1. Feature fusion

[0091] After multi-sensor data splicing, standardization is performed:

[0092]

[0093] 2. Isolation forest module

[0094] Construct T iTree, and calculate an abnormal score:

[0095]

[0096] E(h(x)): sample path length expectation

[0097] c(n): average path length correction term

[0098] Threshold setting: s>0.6 is determined as potential anomaly

[0099] 3. Autoencoder module

[0100] Encoder:

[0101] h = ReLU(W e z + b e )

[0102] Decoder:

[0103]

[0104] Reconstruction error:

[0105]

[0106] 4. Joint decision

[0107] Final anomaly type determined by Softmax classifier:

[0108]

[0109] Confidence fusion:

[0110] Confidence = β·s(x) + (1-β)·max k P(y=k|z)

[0111] Further, the thermal imager adopts a non-refrigeration vanadium oxide detector, is equipped with a window and an air purging device, identifies initial characteristics of the ring by a temperature field distribution, and achieves the effect of early warning of the ring; and a thermal erosion model of the refractory material is established to evaluate the service life of the refractory material.

[0112] Further, the infrared sensor adopts a 2.7-3.5 μm and 4.2-4.8 μm dual-wavelength detector, has a response time of 50 ms, calculates a flame temperature field and combustion product distribution through a dual-wavelength radiation intensity ratio, and detects a combustion frequency (1-10 Hz) in real time, thereby providing feedback for fuel injection control, effectively penetrating the flame smoke, and solving the problem that a traditional single-wavelength sensor is easily blocked.

[0113] Further, the pressure sensor is a ceramic capacitive pressure sensor, has a range of 0-10 kPa and an overload capacity of 200% FS, is made of a stainless steel isolation diaphragm, adjusts the rotation speed of the air blower through an inlet / outlet pressure difference, triggers a PID parameter self-tuning when a fluctuation exceeds ±200 Pa, has a 200% FS overload protection, and is suitable for a kiln pressure impact working condition.

[0114] Further, the electrochemical gas analysis array is configured with CO, CO2, O2, SO2 electrochemical sensors, with a detection lower limit of 1ppm, the sampling pipeline is heated to 120 DEG C, the combustion completeness can be calculated by C0 / CO2 ratio, the combustion efficiency can be evaluated, and the SO2 concentration can be compared with the environmental protection limit value in real time to monitor the emission compliance; the heating pipeline avoids water vapor condensation.

[0115] Further, the edge computing node adopts a Raspberry Pi 4B computing platform, carries a Broadcom BCM2711 quad-core ARM Cortex-A72 processor, is configured with 8GB LPDDR4 memory and 64GB eMMC storage, and is integrated with an LSTM neural network acceleration card based on FPGA, the acceleration card contains 128 computing units, supports half-precision floating point operation, and has a peak computing power of 2.4TOPS, is used for executing data cleaning, time domain alignment, feature dimension reduction and state prediction, can perform real-time feature extraction, completes data cleaning, time domain alignment, feature dimension reduction and other preprocessing operations, and has a power consumption of only 15W, compared with a traditional industrial computer, has a unit power consumption computing power of 160GOPS / W, and meets the deployment requirements of an industrial site.

[0116] Further, the dynamic data fusion module adopts a bidirectional LSTM encoding layer synchronous attention weighting layer, the input data dimension is a spliced vector of a thermal imaging matrix, an infrared spectrum, a pressure time sequence and a gas concentration, a temperature field-burning state-pressure fluctuation-gas composition correlation model is established, and key indicators such as a kiln ring index and a combustion efficiency coefficient are output.

[0117] In addition, the human-computer interaction terminal is equipped with a 12.3-inch touch screen (1280x800), supports multi-layer superposition of a thermal imaging pseudo-color map, a trend curve and a three-dimensional model, displays a temperature field distribution through color mapping, automatically marks an abnormal area as red, visualizes the state, and can provide decision support information such as a maintenance plan and parameter adjustment suggestions.

[0118] In addition, the system further comprises a waste heat recovery intelligent scheduling device, fuses thermal imaging temperature field and flue gas composition data, dynamically adjusts a heat exchanger start-stop strategy, can adjust a coal economizer power according to a flue gas temperature gradient, optimizes system energy efficiency, and can automatically start a bypass waste heat boiler when the SO2 concentration exceeds a standard, and protects equipment.

[0119] In addition, the system further comprises a predictive maintenance device, integrates a vibration sensor and an acoustic monitoring unit, identifies bearing failure through a vibration spectrum feature, forms early warning of failure, establishes a device degradation curve, generates a dynamic maintenance work order, and evaluates system working conditions at any time.

[0120] As described above, due to the adoption of the technical solutions, the present application has the following beneficial effects:

[0121] 1. The thermal imaging manifold learning unit in the edge computing node of the lime kiln monitoring system based on multi-source data dynamic fusion uses the t-SNE algorithm to reduce the 640*480 high-resolution temperature matrix collected by the thermal imager to 16-dimensional manifold features, while retaining the topological structure of the temperature gradient, effectively reducing the data dimension, reducing the computational complexity, and retaining the key temperature gradient information, providing more effective feature representation for subsequent data fusion, analysis and prediction, and helping to improve the accuracy and efficiency of the system.

[0122] 2. The present application effectively solves the problems of low data utilization, control lag, and delayed fault discovery in traditional lime kiln monitoring systems, and realizes the comprehensive improvement of lime kiln operation efficiency, product quality and energy utilization efficiency, creating significant economic and environmental benefits for enterprises.

[0123] 3. The present application realizes real-time monitoring of the running state of the lime kiln through multi-source data dynamic fusion technology, including temperature distribution, flame shape and other key parameters, so as to realize accurate control and optimize the combustion process, significantly improving the energy utilization efficiency.

[0124] 4. The present application adjusts the start-stop strategy of the waste heat recovery equipment dynamically, fully utilizes the kiln flue gas waste heat, reduces energy waste, reduces production cost, and reduces environmental pollution. BRIEF DESCRIPTION OF DRAWINGS

[0125] The present application will be described by way of example and with reference to the accompanying drawings, in which:

[0126] Fig. 1 is a structural schematic diagram of the lime kiln monitoring system based on multi-source data dynamic fusion of the present application.

[0127] Fig. 2 is a workflow diagram of the present application. DETAILED DESCRIPTION

[0128] All features disclosed in this specification, or the steps of any method or process disclosed, can be combined in any manner, except where mutually exclusive.

[0129] Any feature disclosed in this specification, unless stated otherwise, can be replaced by any other equivalent or similar feature. That is, unless stated otherwise, each feature is one example only of a number of equivalent or similar features.

[0130] The features and performance of the present application will be further described in detail below in conjunction with the embodiments.

[0131] As shown in Figs. 1-2 is a deployment diagram of the lime kiln monitoring system

[0132] 1. Implementation method of reducing 640x480 temperature matrix to 16-dimensional manifold features using t-SNE algorithm to preserve temperature gradient topology:

[0133] The edge computing node further includes a thermal imaging manifold learning unit that uses the t-SNE algorithm to reduce the 640x480 temperature matrix collected by the thermal imager to 16-dimensional manifold features to preserve the topology of the temperature gradient, and the steps are:

[0134] S1. Data preprocessing, data blocking and feature extraction, dividing the temperature matrix into local blocks and extracting feature vectors containing temperature values and gradients to capture local topology:

[0135] S1.1 Block division: divide the 640x480 matrix into smaller blocks (e.g. 16x16), resulting in 40x30 = 1200 blocks.

[0136] S1.2 Gradient calculation: use the Sobel operator to calculate the horizontal and vertical gradients of each pixel, reflecting the direction and intensity of temperature change.

[0137] S1.3 Feature extraction: calculate the following features for each block:

[0138] Average temperature: the average temperature of all pixels in the block;

[0139] Temperature standard deviation: the degree of dispersion of temperature values in the block;

[0140] Average horizontal gradient: the average of the horizontal gradient in the block (reflecting the horizontal temperature change);

[0141] Average vertical gradient: the average of the vertical gradient in the block (reflecting the vertical temperature change);

[0142] S2. Apply t-SNE dimensionality reduction, input the extracted features into the t-SNE model, and reduce to 16-dimensional manifold space:

[0143] The following are the configurations of key parameters and their roles:

[0144] n_components = 16: indicates that the target output dimension is 16-dimensional, i.e. mapping high-dimensional data to a 16-dimensional space;

[0145] perplexity = 30-50: adjust according to data size, the larger the value, the more global structure is concerned, the smaller the value, the more local structure is concerned;

[0146] learning_rate = 200-500: to avoid oscillation in the optimization process;

[0147] n_iter = 1000 ~ 5000: The number of iterations is set between 1000 and 5000 to ensure that the model can converge fully;

[0148] S3. Verify the effect of topology preservation:

[0149] Verify whether the temperature gradient structure is preserved after dimensionality reduction by the following method:

[0150] Method 1: Reverse visualization:

[0151] Reduce the 16-dimensional feature to 2D through PCA or UMAP, and observe whether it presents a pattern consistent with the original temperature gradient.

[0152] Method 2: Similarity analysis:

[0153] Calculate the similarity ranking of high-dimensional features (original blocks) and low-dimensional features, and check whether the local neighborhood is consistent.

[0154] 2. Usage steps of the working condition analysis system:

[0155] S1. Time and location reference unification:

[0156] S1.1 Hardware synchronization: Deploy PPS (Pulse Per Second) signal distribution module on edge computing node, thereby realizing: frame-level synchronization of thermal imager and infrared sensor, and ensuring synchronization error <2ms, alignment of sampling time of pressure / gas sensor and start time of thermal imaging frame;

[0157] S1.2 Orientation registration: Align thermal imager pixel coordinates with kiln three-dimensional physical coordinates, and align infrared sensor field of view angle with key areas of combustion chamber.

[0158] S2. Data preprocessing

[0159] S2.1 Data standardization

[0160]

[0161] S2.2 Dynamic feature extraction

[0162] S2.2.1 Thermal imaging feature extraction: temperature gradient field (Sobel operator calculation), local entropy feature (detecting ring area);

[0163] S2.2.2 Infrared feature extraction: pressure fluctuation power spectral density (PSD), surge frequency feature;

[0164] S2.2.3 Gas feature extraction: CO / CO2 concentration ratio (combustion completeness index), O2 concentration change rate (excess air coefficient).

[0165] S3. Input working condition analysis model library, output data.

[0166] S3.1 Ringing prediction model:

[0167] Input features: thermal imaging local entropy features, pressure fluctuation power spectral density, gas concentration rate of change.

[0168] Output: Ringing position + thickness prediction.

[0169] S3.2 Combustion optimization model:

[0170] Input features: infrared flame pulsation frequency, O2 / CO concentration ratio, furnace pressure gradient.

[0171] Output: Optimal air coefficient + recommended fuel flow value.

[0172] S3.3 Abnormality detection model:

[0173] Input features: multi-sensor feature fusion vector.

[0174] Output: Abnormal type (leakage / coking / flameout) + confidence.

[0175] 2. Sensor network deployment:

[0176] Thermal imager installation: 4 FLIR A655sc thermal imagers (resolution 640x480, NETD 50mK) are evenly distributed on the outside of the kiln rotary section, covering a field of view angle of 120° arc surface, equipped with sapphire glass window and compressed air purging device, purging frequency 5 times / hour, to prevent dust adhesion.

[0177] Infrared sensor arrangement: LandCyclops L series dual-wavelength infrared sensor is installed at the center of the combustion chamber top, 2.7μm and 4.5μm wavelength channels are independently calibrated, quartz lens group is set in the optical path, focusing diameter ≤50mm, to ensure flame penetration monitoring.

[0178] Pressure sensor configuration: 2 Druck PTX5072 ceramic capacitive differential pressure sensors are installed at the flanges of the air inlet and exhaust outlet, using 316L stainless steel diaphragm, range 0-10kPa, overload protection threshold 20kPa, connected to the field bus through M12 aviation plug.

[0179] Gas analyzer integration: sampling probe is welded at the horizontal section of the flue, ABB AO2020 electrochemical gas analyzer is built-in, equipped with 120℃ heating sampling pipeline, CO / CO2 / O2 / SO2 detection lower limit is 1ppm / 0.5ppm / 0.1% / 1ppm respectively, response time T90≤10 seconds.

[0180] 3. Edge computing node configuration:

[0181] Hardware platform: Raspberry Pi 4B computing module (BCM2711 quad-core A72 @ 1.5GHz, 8GB LPDDR4-3200) is adopted, and self-made FPGA acceleration card (Xilinx Artix-7 XC7A100T) is mounted, which are interconnected through PCIe 2.0 x 1 interface.

[0182] Software environment: Python 3.9 runtime is deployed, PyTorch 1.12 mobile version framework is integrated, LSTM model is optimized by TensorRT 8.5, and half-precision floating point (FP16) inference delay is less than or equal to 180 ms.

[0183] Preprocessing flow:

[0184] Data cleaning: 3σ principle is adopted to eliminate abnormal pixel points of thermal imager (such as temperature mutation caused by smoke shielding);

[0185] Time domain alignment: based on NTP protocol, the time stamps of each sensor are synchronized, and the error is less than or equal to 10 ms;

[0186] Feature dimension reduction: principal component analysis (PCA) is used to compress thermal imaging data from 307200 dimensions (640x480) to 1536 dimensions, retaining 98.2% of the variance.

[0187] 4. Dynamic data fusion module implementation:

[0188] Network structure: bidirectional LSTM encoding layer is adopted, and the input data dimension is [thermal imaging matrix (1536 dimensions), infrared spectrum (128 dimensions), pressure time sequence (32 dimensions), and gas concentration (4 dimensions)] spliced vector.

[0189] 5. Human-computer interaction terminal development:

[0190] Hardware configuration: 12.3-inch Siemens SIMATIC HMI KTP1200 Basic PN is adopted, with a resolution of 1280x800, a capacitive touch screen, and support for multi-point touch and glove operation mode.

[0191] Interface design:

[0192] Main screen: thermal imaging pseudo-color map (color scale range 800-1350℃) covers the three-dimensional model of the kiln, and the abnormal area (temperature gradient > 5℃ / cm) is marked with red outline;

[0193] Trend screen: 24-hour historical curves of 8 parameters such as combustion chamber pressure, CO concentration, and ring index are displayed synchronously;

[0194] Alarm screen: display by severity level (red- emergency shutdown, orange- operation intervention, yellow- attention monitoring), support alarm confirmation and handling record.

[0195] 6. Intelligent scheduling strategy for waste heat recovery:

[0196] Control logic:

[0197] When the standard deviation of flue gas temperature field is greater than 15℃, start the heat exchanger power regulation mode, dynamically adjust the economizer inlet water temperature set value from 80℃ to 60-90℃ through Modbus TCP;

[0198] Real-time monitoring of SO2 concentration, when the sampling value is greater than 50mg / m 3 for 3 consecutive times, automatically open the bypass valve (opening degree 100%), isolate the waste heat boiler.

[0199] Energy saving effect: the actual measurement data of a certain cement plant shows that the average daily energy saving is 12.3%, and the corrosion rate of the waste heat boiler is reduced from 0.8mm / year to 0.3mm / year.

[0200] 7. Predictive maintenance implementation process:

[0201] Vibration monitoring: install PCB Piezotronics 356A15 acceleration sensor in the vertical / horizontal direction of the kiln body riding wheel bearing seat, sampling frequency 1024Hz, and focus on 0.4 times frequency (bearing wear characteristic), 1 times frequency (unbalance) components in frequency spectrum analysis.

[0202] Maintenance decision:

[0203] When the vibration RMS value exceeds 5mm / s, a yellow warning is generated, suggesting to shorten the lubrication period;

[0204] When the 0.4 times frequency amplitude exceeds 1 times frequency by 50%, a red warning is generated, and planned shutdown maintenance is scheduled.

[0205] Application effect: after the implementation of a certain production line, the number of unplanned shutdowns is reduced from an average of 12 times per year to 3 times, and the bearing inventory turnover rate is increased by 40%.

[0206] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A lime kiln monitoring system based on dynamic fusion of multi-source data, characterized in that: The system includes a thermal imager, an infrared sensor, a pressure sensor, a gas analyzer, and an edge computing node. The edge computing node is configured with a Raspberry Pi platform and an LSTM neural network accelerator to perform multi-source data preprocessing and feature compression. The edge computing node also includes a thermal imaging manifold learning unit, which uses the t-SNE algorithm to reduce the dimension of the temperature matrix collected by the thermal imager to a 16-dimensional manifold feature to preserve the topological structure of the temperature gradient. The steps are as follows: S1. Data preprocessing, data segmentation and feature extraction: Divide the temperature matrix into local blocks and extract feature vectors containing temperature values ​​and gradients to capture local topological structures. S2. Apply t-SNE dimensionality reduction and input the extracted features into the t-SNE model to reduce the dimensionality to a 16-dimensional manifold space. S3. Verify the topology preservation effect; The system also includes an operating condition analysis system, which includes a ring formation prediction model based on ST-GCN, a combustion optimization model based on DDPG, and an anomaly detection model based on an isolation forest + autoencoder hybrid model, wherein: The loop prediction model is fed with the local entropy characteristics of thermal imaging, the power spectrum density of pressure fluctuations, and the rate of change of gas concentration, and outputs the prediction of the loop position and thickness. Input infrared flame pulsation frequency, O2 / CO concentration ratio, and furnace pressure gradient into the combustion optimization model, and output the recommended values ​​of optimal air coefficient and fuel flow rate; The multi-sensor fusion vector is input into the anomaly detection model, and the anomaly type is output, where the anomaly type includes leakage, coking, flameout, and confidence.

2. The lime kiln monitoring system according to claim 1, characterized in that: The thermal imager is installed on the outside of the furnace wall of the lime kiln and uses an uncooled vanadium oxide microbolometer detector with a resolution of not less than 640×480 pixels and a frame rate of ≥30Hz. It is used to collect real-time temperature distribution data on the kiln surface and transmit the original thermal image data to the edge computing node through the Ethernet interface.

3. The lime kiln monitoring system according to claim 1, characterized in that: The infrared sensor uses dual-wavelength infrared detection technology, with a wavelength range covering 2.7-3.5μm and 4.2-4.8μm, a measurement accuracy of ±0.5℃, and a response time of ≤50ms. It is deployed on the top of the kiln combustion chamber to monitor the flame morphology and temperature gradient of the combustion zone. The data is communicated with the edge computing node via the RS-485 bus.

4. The lime kiln monitoring system according to claim 1, characterized in that: The pressure sensor is a ceramic capacitive differential pressure sensor with a measuring range of 0-10kPa and an overload capacity of 200% FS. It is installed at the air inlet and exhaust port of the kiln to detect the pressure difference in the furnace in real time and transmit the pressure data to the edge computing node via the Modbus RTU protocol.

5. The lime kiln monitoring system according to any one of claims 1 to 4, characterized in that: It also includes a gas analyzer equipped with an electrochemical sensor array, which includes CO, CO2, O2 and SO2 detection units. The sampling pipeline is made of polytetrafluoroethylene and is equipped with a heating and insulation device. The analysis data is transmitted to the edge computing node via the CAN bus.

6. The lime kiln monitoring system according to claim 1, characterized in that: The edge computing node uses a Raspberry Pi 4B computing platform, equipped with a Broadcom BCM2711 quad-core ARM Cortex-A72 processor, configured with 8GB LPDDR4 memory and 64GB eMMC storage, and integrated with an FPGA-based LSTM neural network accelerator card. The accelerator card contains 128 computing units, supports half-precision floating-point operations, and has a peak computing power of 2.4TOPS, which is used to perform data cleaning, time domain alignment, feature dimensionality reduction and state prediction.

7. The lime kiln monitoring system according to claim 1, characterized in that: It also includes a dynamic data fusion module, which runs on the edge computing node and adopts an improved attention mechanism LSTM network. The improved attention mechanism LSTM network structure includes an input layer, a bidirectional LSTM encoding layer, an attention weighted layer and a fully connected output layer. The input data dimension is the splicing vector of the thermal imaging temperature matrix, the infrared spectrum vector, the pressure time series data and the gas concentration sequence, and the output is the kiln ring index and combustion efficiency.

8. The lime kiln monitoring system according to claim 1, characterized in that: The system is also equipped with a human-computer interaction terminal, an industrial-grade touch screen with a resolution of 1280×800, supports multi-layer overlay display, and simultaneously displays thermal imaging pseudo-color images, sensor numerical curves, LSTM prediction results and three-dimensional visualization kiln models. It has historical data playback, alarm threshold configuration and remote diagnosis interface functions.

9. The lime kiln monitoring system according to claim 1, 6, 7 or 8, characterized in that: It also includes an intelligent scheduling device for waste heat recovery, which integrates the temperature field distribution and flue gas composition analysis results in multi-source data to dynamically adjust the start and stop strategies of the waste heat recovery equipment.

10. The lime kiln monitoring system according to claim 9, characterized in that: It also includes a predictive maintenance device, including an integrated vibration sensor composed of the above sensors. The integrated vibration sensor and acoustic monitoring unit support equipment fault warning and health status assessment, and generate customized maintenance plans.

Images