Flue gas intelligent data analysis method based on large model ring collection

Abstract

The invention discloses an intelligent data analysis method for annularly-collected flue gas based on a large model, belongs to the technical field of industrial environmental protection and intelligent control, and aims at solving the problems that in an existing annularly-collected flue gas system, energy is wasted seriously, and faults of a valve and a fan cannot be early warned in time. Cleaning, filling, standardizing and storing; s2, on the basis of intelligent analysis of a large model, a data layer-model layer-application layer framework is constructed, smoke concentration changes are predicted, and equipment abnormity is recognized; s3, optimizing the opening degree of the valve and the rotating speed of the fan by adopting a multi-target adaptive particle swarm algorithm, and dynamically adjusting the target weight; s4, triggering early warning through a Tstatistic and a fault probability formula, and providing maintenance suggestions in combination with the knowledge graph; according to the method, potential faults of the valve and the fan can be recognized in advance through the Tstatistical magnitude and the fault probability model, and compared with a traditional post-fault maintenance mode, the fault shutdown time is shortened.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial environmental protection and intelligent control, and particularly relates to an intelligent data analysis method for a large model-based exhaust gas collection system. BACKGROUND

[0002] In the industrial production process, such as copper smelting, steel manufacturing and other industries, the exhaust gas collection system is an important facility for controlling pollutant emissions and ensuring the safety of the working environment. The traditional exhaust gas collection system has many problems. Different operators have large differences in the control of valve opening and fan operation at different times, resulting in poor coordination. This not only causes the exhaust gas in the plant to not be collected in time, endangering the health of employees, but also causes a large amount of energy waste. At the same time, for key equipment such as valves and fans, there is a lack of effective fault monitoring and early warning mechanism, and sudden equipment failure often leads to production interruption, causing huge economic losses. Currently, although there are some technical solutions for equipment fault diagnosis and energy optimization, there are generally problems such as limited data processing capability and poor model generalization. In the face of complex and variable industrial production environment and massive equipment operation data, traditional technology is difficult to achieve accurate fault prediction and efficient energy saving and emission reduction control. With the rapid development of artificial intelligence technology, large model technology shows strong data processing and complex pattern recognition capabilities, providing new ideas and methods to solve the above problems.

[0003] Therefore, it is necessary to provide a large model-based exhaust gas collection system intelligent data analysis method to solve the problems of serious energy waste in the existing exhaust gas collection system and inability to timely warn of valve and fan failures in the prior art. SUMMARY

[0004] The purpose of the present application is to provide a large model-based exhaust gas collection system intelligent data analysis method to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present application provides the following technical solution: a large model-based exhaust gas collection system intelligent data analysis method, comprising the following steps: S1, data acquisition and preprocessing: deploying multiple types of sensors at key positions of the exhaust gas collection system, real-time collecting exhaust gas concentration, temperature, pressure, flow, valve opening, fan speed and fan current; transmitting the collected raw data to a data preprocessing unit, removing outliers by 3σ criterion, filling missing values by linear interpolation method or mean filling method, and adopting Z-score standardization processing to obtain preprocessed data, and the Z-score standardization formula is: ; wherein is the raw data, is the historical data mean, The standard deviation of historical data is used to store real-time data and preprocessed data in a time series database, and fault records and process parameters are stored in a relational database. S2, intelligent analysis based on large model: a large model analysis system with a three-layer architecture of "data layer-model layer-application layer" is constructed, the model layer is fine-tuned based on DeepSeek architecture, including a multi-modal encoder, a time series prediction module and a fault diagnosis module; the preprocessed data of step S1 is input into the large model, the flue gas parameters and the equipment state are fused by the multi-modal encoder, the time series prediction module predicts the future 10-minute flue gas concentration change, and the fault diagnosis module identifies the equipment anomaly in combination with the fault knowledge graph; S3, energy saving and emission reduction control: a multi-objective adaptive particle swarm optimization algorithm is used, with the minimum energy consumption E and the maximum capture efficiency η as the objectives, and the valve opening α (5%-80%) and the fan speed n (500-1450 rpm) as the optimization variables; the target weight is adjusted according to the flue gas concentration prediction result of step S2, and the optimal parameters are output after iterative optimization, and the valve and fan are dynamically adjusted; S4, valve and fan fault reminder: the deviation degree of the current equipment parameters from the normal parameters is calculated by T2 statistic, and the fault probability is calculated by the fault probability formula; when the deviation degree exceeds the threshold or the fault probability is greater than 80%, the early warning is triggered and the fault knowledge graph is matched, and the fault diagnosis report and maintenance suggestion are generated.

[0006] It should be noted that in step S1, the key positions include the copper discharge compartment area of the main plant electric furnace, the top and both sides of the converter, the top and mouth cover area of the anode furnace; the multi-type sensors include sulfur dioxide gas detector, PT100 temperature sensor, MPX5010 pressure sensor, FS4000 flow sensor, valve opening sensor, Hall effect type speed sensor and ACS712 current sensor, and the sampling frequency is 1-10 Hz.

[0007] It should be further noted that in step S2, the time series prediction module adopts the Transformer structure, inputs the past 5-minute preprocessed data, and the prediction error is ≤5%; the fault diagnosis module outputs the fault type and the confidence degree ≥85%, and the fault knowledge graph stores the fault type, reason, performance and maintenance suggestion.

[0008] It should be further noted that in step S3, the velocity formula of the multi-objective adaptive particle swarm optimization algorithm is: ; Wherein: is the particle velocity after updating; is the dynamic inertia weight; is the current particle velocity; is the cognitive learning factor; a social learning factor; a random disturbance factor; a particle individual historical optimal position; a particle current position; a population global optimal position; wherein linearly decreases from 0.9 to 0.4 with iterations, is adaptively adjusted to 1.5-2.5 according to the degree of negative correlation between E and η, is a [0, 1] random number.

[0009] As a preferred embodiment, the target weight adjustment rule in step S3 is: when the flue gas concentration C is greater than 300 mg / m³, the η weight is 0.6 and the E weight is 0.4; when C is less than 100 mg / m³, the E weight is 0.6 and the η weight is 0.4.

[0010] As a preferred embodiment, the T² statistical quantity formula in step S4 is: ; wherein, X is the current parameter vector, μ is the normal parameter mean vector, and S is the normal parameter covariance matrix; The fault probability formula is: ;

[0011] wherein, is a weight matrix, is a bias vector.

[0012] As a preferred embodiment, the early warning in step S4 is divided into three levels: general early warning-no shutdown risk, serious early warning-1-2 hours of possible shutdown, and emergency early warning-shutdown within 30 minutes, and the early warning information is pushed to the mobile phone APP and the monitoring platform.

[0013] As a preferred embodiment, it further includes user interaction and a monitoring platform, which displays a flue gas concentration heat map, a device state panel, an early warning information column, and an energy consumption report, supports remote access through a Web terminal and an APP terminal, and uses HTTPS encrypted transmission.

[0014] As a preferred embodiment, it further includes interface design: the RS-485 channel uses master-slave time-sharing transmission with CRC-16+XOR check added; the Ethernet channel uses the lwIP protocol stack, supports MRP cable redundancy and double-network-port binding.

[0015] As a preferred embodiment, the training data of the large model in step S2 includes at least one year of ring flue gas system operation data, the training set accounts for 80%, the verification set accounts for 20%, and the fine-tuning iteration is greater than or equal to 50 rounds.

[0016] Compared with the prior art, the intelligent data analysis method for flue gas based on a large model provided by the application at least has the following beneficial effects: (1) The application uses a multi-objective self-adaptive particle swarm algorithm, combines a large model for 10-minute short-term prediction of flue gas concentration, and dynamically adjusts the valve opening: when the flue gas concentration is low, the fan speed and valve opening are preferentially reduced to reduce invalid energy consumption; when the flue gas concentration is high, the core is to ensure the trapping efficiency and avoid excessive energy consumption; and based on the coupling calculation of the valve flow characteristic model and the flue gas diffusion model, the application can stably control the flue gas trapping efficiency to be ≥95%, which is much higher than the traditional manual control of 85%-90%.

[0017] (2) The application can identify potential faults (such as valve jamming and fan motor overload) of the valve and fan in advance through T2 statistics (measuring the degree of deviation of equipment parameters from the normal state) and a fault probability model, which reduces the downtime compared with the traditional “post-fault maintenance” mode; and after the fault warning is triggered, the system generates a diagnosis report (such as “excessive fan current → reason: inlet filter screen blockage → suggestion: clean the ash”) in a short time through the fault knowledge graph (associated with fault type, reason, and maintenance suggestion), without the need for maintenance personnel to investigate the fault reason on site, thereby shortening the fault handling time; at the same time, the knowledge graph can accumulate maintenance cases (such as storing 120+ fault records for 3 years), form an enterprise-level equipment health management database, and improve the confidence of subsequent fault diagnosis.

[0018] (3) The large model of the application is fine-tuned based on the DeepSeek architecture, and the training data covers the flue gas operation data of multiple industries such as copper smelting and steel manufacturing (such as flue gas parameter characteristics of different smelting equipment such as electric furnaces, converters, and blast furnaces), which can quickly adapt to different industrial scenarios through the fusion of “flue gas parameters + equipment state + process labels” by a multi-modal encoder.

[0019] (4) The user interaction and monitoring platform of the application provides a visual interface of “flue gas concentration heat map + equipment state panel + energy consumption report”, supports remote access (HTTPS encrypted transmission) of computer Web end and mobile phone APP end, and the system defaults to the control mode of “large model automatic optimization + parameter automatic delivery”, which only needs manual intervention (needs secondary authorization) in special working conditions (such as sensor failure).

[0020] (5) The application has strong compatibility through the design of RS-485 and Ethernet double interfaces: the RS-485 channel supports master-slave time-sharing transmission (anti-interference, error rate <1%), which can adapt to existing sensors of enterprises (such as PT100 temperature sensors, Hall effect speed sensors); the Ethernet channel adopts a lightweight lwIP protocol stack (data packet is simplified to 512 bytes), supports MRP cable redundancy (single network line breaks are self-healed within 50 ms), can be integrated with existing SCADA (Supervisory Control And Data Acquisition, data acquisition and monitoring control system) and MES (Manufacturing Execution System, manufacturing execution system) of enterprises, without the need to replace existing hardware devices, reducing the initial investment of enterprise technological transformation. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 It is a software architecture schematic diagram of the intelligent data analysis method for flue gas based on a large model ring set according to the application; Figure 2 It is a data acquisition system diagram of the intelligent data analysis method for flue gas based on a large model ring set according to the application; Figure 3 It is a flowchart of the intelligent data analysis method for flue gas based on a large model ring set according to the application; Figure 4 It is a user interaction and monitoring platform UI interface schematic diagram of the intelligent data analysis method for flue gas based on a large model ring set according to the application. DETAILED DESCRIPTION

[0022] The application will be further described below in combination with examples.

[0023] Please refer to Figures 1-4 The application provides an intelligent data analysis method for flue gas based on a large model ring set, including the following steps: S1, data acquisition and preprocessing: deploying multiple types of sensors at key positions of the ring set flue gas system, real-time collecting flue gas concentration, temperature, pressure, flow, valve opening, fan speed and fan current; transmitting the collected raw data to a data preprocessing unit, removing outliers through the 3σ criterion, filling missing values by linear interpolation method or mean filling method, and adopting Z-score standardization processing to obtain preprocessed data, and the Z-score standardization formula is: ; Among them is the raw data, is the historical data mean, is the historical data standard deviation; a time series database is used to store real-time data and preprocessed data, and a relational database is used to store fault records and process parameters; S2, intelligent analysis based on a large model: a large model analysis system with a "data layer-model layer-application layer" three-layer architecture is constructed, the model layer is fine-tuned based on a DeepSeek architecture, including a multi-modal encoder, a time series prediction module, and a fault diagnosis module; the preprocessed data of step S1 is input into the large model, the flue gas parameters and the equipment state are fused by the multi-modal encoder, the time series prediction module predicts the change of flue gas concentration in the next 10 minutes, and the fault diagnosis module identifies equipment abnormalities in combination with the fault knowledge graph.

[0024] Specifically, the data layer is used for data storage and calling, the time series database stores real-time data (retained for 3 months), the relational database stores historical fault data (retained for 3 years), and supports quick retrieval by "smelting stage" and "equipment number"; The model layer is used for multi-modal fusion, trend prediction, and fault diagnosis, and is fine-tuned based on a DeepSeek architecture, including: 1. Multi-modal encoder: fuse flue gas parameters (concentration, temperature), equipment state (speed, current), process labels (such as "electric furnace smelting"); 2. Time series prediction module: Transformer structure, input past 5 minutes preprocessed data, predict flue gas concentration change in next 10 minutes (prediction error ≤5%); 3. Fault diagnosis module: combine fault knowledge graph to construct Prompt, output fault type and confidence (≥85%); The application layer is used for instruction output and report generation, output valve opening / fan speed optimization value, fault warning level (general / severe / emergency), ton product energy consumption report.

[0025] The training data of the large model includes at least 1 year of circulating flue gas system operation data, the training set accounts for 80%, the validation set accounts for 20%, and the fine-tuning iteration is ≥50 rounds.

[0026] S3, energy saving and emission reduction control: a multi-objective adaptive particle swarm optimization algorithm is used, with the minimum energy consumption E and the maximum capture efficiency η as the target, and the valve opening α (5%-80%) and the fan speed n (500-1450 rpm) as the optimization variable; adjust the target weight according to the flue gas concentration prediction result of step S2, output the optimal parameters after iterative optimization, and dynamically adjust the valve and the fan; The velocity formula of the multi-objective adaptive particle swarm optimization algorithm is: ; Wherein: is the velocity of the particle after updating; is the dynamic inertia weight; is the current velocity of the particle; is the cognitive learning factor; is the social learning factor; Random disturbance factor; The individual optimal position of particles in history; The current position of particles; The global optimal position of the population; Wherein Linearly decreases from 0.9 to 0.4 with iterations, Adaptively adjusts to 1.5-2.5 according to the degree of negative correlation between E and η, Is a [0, 1] random number.

[0027] The target weight adjustment rule is: when the flue gas concentration C is greater than 300 mg / m³, the η weight is 0.6 and the E weight is 0.4; when C is less than 100 mg / m³, the E weight is 0.6 and the η weight is 0.4.

[0028] S4, valve fan fault warning: calculate the deviation degree of the current equipment parameter from the normal parameter through T²statistical quantity, calculate the fault probability by using the fault probability formula; when the deviation degree exceeds the threshold value or the fault probability is greater than 80%, trigger the early warning and match the fault knowledge graph, generate the fault diagnosis report and maintenance suggestion.

[0029] Wherein the T²statistical quantity formula is: ; Wherein, X is the current parameter vector, μ is the normal parameter mean vector, and S is the normal parameter covariance matrix; The fault probability formula is: ; Wherein, is a weight matrix, is a bias vector.

[0030] The early warning is divided into three levels: general early warning-no shutdown risk, serious early warning-1-2 hours of possible shutdown, and emergency early warning-shutdown within 30 minutes, and the early warning information is pushed to the mobile phone APP and the monitoring platform.

[0031] Further as Figure 4 Illustrated, it is worth to be particularly explained that it also includes user interaction and monitoring platform, the platform shows flue gas concentration thermal map, equipment state panel, early warning information column and energy consumption report, supports Web end and APP end remote access, and adopts HTTPS encrypted transmission.

[0032] Develop a user interaction interface to real-time display the running state of the flue gas system, flue gas concentration distribution, equipment running parameters, energy consumption data and fault early warning information, etc. Users can conveniently check the system running condition through the interface, manually adjust the control parameters, query historical data and fault records, etc. Realize remote monitoring function, allow the management personnel to access the system through the computer or mobile device in any place with network connection, master the running condition of the ring smoke gas system in real time, and handle the equipment failure and abnormal situation in time.

[0033] Further, it is worth specifically mentioning that it also includes interface design: RS-485 channel adopts master-slave time-sharing transmission, adds CRC-16+XOR check; the Ethernet channel adopts lwIP protocol stack, supports MRP cable redundancy and double network port binding.

[0034] Specifically includes the following contents: 1, RS-485 upgrade channel Master-slave time-sharing transmission: The central coordinator acts as the 485 master station, polling the slave node according to the topological order, and sending upgrade instructions.

[0035] The slave node only responds after receiving the exclusive address code (such as 0xA5+node ID), avoiding bus conflict.

[0036] Anti-interference enhancement: Double check is added to the data frame: CRC-16 (frame check) + XOR check (byte level).

[0037] Baud rate adaptation: initial 115200bps, if the error rate >10%, automatically reduce to 57600bps.

[0038] Cascade relay management: The repeater identifies the target node segment number, only forwards the data of the owned section, and reduces the bus load.

[0039] 2, Ethernet port upgrade channel Protocol stack optimization: Lightweight TCP / IP protocol stack (such as lwIP) is adopted, and the data packet is simplified to 512 bytes, which is suitable for resource-limited MCU.

[0040] Upgrade instruction uses UDP broadcast to discover nodes, and subsequent transmission switches to TCP point-to-point connection.

[0041] Industrial Ethernet fault tolerance: Support cable redundancy (MRP protocol), self-healing within 50ms when single network cable breaks.

[0042] Key node deployment double network port binding, bandwidth aggregation + fault switching.

[0043] In summary: the present application combines a large model for 10-minute short-term prediction of flue gas concentration through a multi-objective adaptive particle swarm algorithm, dynamically adjusts the valve opening: when the flue gas concentration is low, the fan speed and valve opening are reduced first to reduce invalid energy consumption; when the flue gas concentration is high, the core is to ensure the trapping efficiency and avoid excessive energy consumption; and based on the coupling calculation of the valve flow characteristic model and the flue gas diffusion model, the present application can stably control the flue gas trapping efficiency to be ≥95%, which is much higher than the traditional manual control of 85%-90%; the present application can identify potential faults of the valve and the fan (such as valve jamming and fan motor overload) in advance through T2 statistics (measure the degree of deviation of equipment parameters from the normal state) and a fault probability model, reduce the downtime compared with the traditional "fault maintenance" mode; and after the fault warning is triggered, the system generates a diagnosis report (such as "excessive fan current → reason: inlet filter screen blockage → suggestion: clean the accumulated dust") through the fault knowledge graph (associated fault type, reason, maintenance suggestion) in a short time, without the need for maintenance personnel to investigate the fault reason on site, shorten the fault handling time; at the same time, the knowledge graph can accumulate maintenance cases (such as 120+ fault records stored for 3 years), form an enterprise-level equipment health management database, and improve the subsequent fault diagnosis confidence.

[0044] The large model of the present application is based on the DeepSeek architecture fine-tuning, and the training data covers multi-industry flue gas operation data of copper smelting, steel manufacturing, etc. (such as flue gas parameter characteristics of different smelting equipment such as electric furnace, converter, blast furnace, etc.), which can quickly adapt to different industrial scenes through multi-modal encoder fusion of "flue gas parameters + equipment state + process label".

Claims

1. A method for intelligent data analysis of flue gas based on a large-scale model, characterized by: Includes the following steps: S1. Deploy multiple types of sensors at key locations in the flue gas collection system to collect real-time data on flue gas concentration, temperature, pressure, flow rate, valve opening, fan speed, and fan current; transmit the collected raw data to the data preprocessing unit. S2. Construct a large model analysis system with a three-layer architecture of "data layer - model layer - application layer". The model layer is based on the DeepSeek architecture and is finely tuned, including a multimodal encoder, a time series prediction module and a fault diagnosis module. Input the preprocessed data from step S1 into the large model. The multimodal encoder fuses flue gas parameters and equipment status. The time series prediction module predicts the flue gas concentration change in the next 10 minutes. The fault diagnosis module combines the fault knowledge graph to identify equipment abnormalities. S3. A multi-objective adaptive particle swarm optimization algorithm is adopted, with the objectives of minimizing energy consumption E and maximizing capture efficiency η. The optimization variables are valve opening α (5%-80%) and fan speed n (500-1450rpm). The objective weights are adjusted according to the flue gas concentration prediction results in step S2. After iterative optimization, the optimal parameters are output, and the valve and fan are dynamically adjusted. S4. Calculate the deviation of the current equipment parameters from the normal parameters using the T² statistic, and calculate the failure probability using the failure probability formula; when the deviation exceeds the threshold or the failure probability is >80%, trigger an early warning and match the fault knowledge graph to generate a fault diagnosis report and maintenance suggestions.

2. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The key locations mentioned in step S1 include the copper tapping area of ​​the main plant electric furnace, the top and side areas of the converter furnace, and the top and furnace mask areas of the anode furnace. The various types of sensors include a sulfur dioxide gas detector, a PT100 temperature sensor, an MPX5010 pressure sensor, an FS4000 flow sensor, a valve opening sensor, a Hall effect speed sensor, and an ACS712 current sensor, with a sampling frequency of 1-10Hz.

3. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The time-series prediction module in step S2 adopts a Transformer structure, inputs preprocessed data from the past 5 minutes, and has a prediction error of ≤5%; the fault diagnosis module outputs the fault type and confidence level of ≥85%, and the fault knowledge graph stores the fault type, cause, manifestation, and maintenance suggestions.

4. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The speed formula for the multi-objective adaptive particle swarm algorithm in step S3 is: ； in: The velocity of the particles after the update; For dynamic inertia weights; This represents the particle's current velocity. For cognitive learning factors; As a social learning factor; Random disturbance factor; This represents the best historical position of an individual particle. The particle's current position; The optimal position for the entire group; in As the iterations decreased linearly from 0.9 to 0.4, The value is adaptively adjusted to 1.5-2.5 based on the degree of negative correlation between E and η. It is a random number in the range [0,1].

5. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The target weight adjustment rule in step S3 is as follows: when the flue gas concentration C>300mg / m³, the weight of η is 0.6 and the weight of E is 0.4; when C<100mg / m³, the weight of E is 0.6 and the weight of η is 0.

4.

6. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The formula for the T² statistic mentioned in step S4 is: ； Where X is the current parameter vector, μ is the normal parameter mean vector, and S is the normal parameter covariance matrix; The formula for the failure probability is: ； in, This is the weight matrix. This is the bias vector.

7. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The warnings mentioned in step S4 are divided into three levels: general warning - no risk of downtime, severe warning - possible downtime in 1-2 hours, and emergency warning - downtime within 30 minutes. The warning information is pushed to the mobile APP and monitoring platform.

8. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: It also includes a user interaction and monitoring platform, which displays a flue gas concentration heat map, equipment status panel, early warning information bar and energy consumption report, supports remote access via web and APP, and uses HTTPS encrypted transmission.

9. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: It also includes interface design: the RS-485 channel adopts master-slave time-division transmission and adds CRC-16+XOR check; the Ethernet channel adopts the lwIP protocol stack and supports MRP cable redundancy and dual network port binding.

10. The intelligent data analysis method for flue gas based on a large-scale model as described in claim 1, characterized in that: The training data for the large model mentioned in step S2 includes at least one year of operation data of the flue gas collection system, with the training set accounting for 80%, the validation set accounting for 20%, and the fine-tuning iterations ≥50 rounds.

Images