A method and system for predicting a collapse risk of an electrostatic precipitator

Abstract

The application discloses a kind of electric precipitator collapse risk prediction method and system, using sensor monitoring each risk index of electric precipitator;The collected data are filtered, and the influence of the fluctuation of electric precipitator inlet pressure on equipment amplitude is filtered out;BP neural network model is constructed, as the basis for identifying electric precipitator ontology collapse risk;Whether displacement overrun phenomenon occurs is calculated, and the threshold value of the alarm condition triggered by BP model is adjusted according to the result;Electric precipitator collapse risk is predicted using BP neural network model.This method can reduce the inspection workload of electric precipitator ontology, save the labor cost in thermal power generation, and improve the safety and risk controllability in the process of electric power production, prolong the service life of equipment.

Description

Technical Field

[0001] This invention relates to the field of thermal power plant equipment maintenance, specifically to a method and system for predicting the collapse risk of electrostatic precipitators. Background Technology

[0002] Electrostatic precipitators (ESPs) are essential equipment in thermal power plants, serving as a primary means of cleaning flue gas pollution. They remove particles from the flue gas emitted from the furnace, significantly reducing the amount of dust released into the atmosphere and thus mitigating the environmental pollution caused by thermal power plants. However, with the widespread use of ESPs in thermal power plants, some unavoidable problems have emerged. During long-term operation, issues such as poor ash conveying (especially in ESPs retrofitted for ultra-low emissions) and untimely ash cleaning lead to ash accumulation and corrosion within the ESP. Ash accumulation causes excessive load on the ESP, while corrosion reduces the structural strength and support capacity, potentially leading to structural collapse and safety accidents.

[0003] In the context of comprehensive energy conservation and emission reduction, thermal power units have undertaken more peak-shaving tasks in order to meet the requirements of fully absorbing non-fossil energy sources such as wind power and photovoltaics. They not only have to respond quickly to the rise and fall of operating loads within a wide load range, but also face problems such as uncontrollable coal quality caused by co-firing and long-term deviation of boiler operating parameters from normal values. These problems pose a huge challenge to the safe and stable operation of electrostatic precipitators.

[0004] Currently, the monitoring systems for electrostatic precipitators (ESPs) in power plants are mainly DCS systems. These systems primarily monitor the operating status of the ESPs themselves, such as dust concentration and humidity, gas pressure and temperature in the pipelines, and flue gas flow rate—parameters related to production activities. However, current ESP monitoring systems do not include the risk of ESP collapse within their monitoring scope. But as thermal power plants increasingly undertake peak-shaving and co-firing tasks, the uncontrollable risks in the production process are gradually increasing. Therefore, incorporating the risk of ESP collapse into production management is imperative.

[0005] The problem with existing technology is:

[0006] (1) Manual inspection has blind spots and wastes manpower costs;

[0007] (2) The safety risks of electrostatic precipitators are uncontrollable during production and use;

[0008] (3) Electrostatic precipitators operate under harsh conditions for a long time, which affects the service life of the equipment. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention provides a method for predicting the collapse risk of electrostatic precipitators. An improved wavelet transform threshold noise reduction method is used to filter the information collected by sensors, thereby filtering out the influence of electrostatic precipitator inlet pressure fluctuations on the equipment amplitude and retaining only the influence of the electrostatic precipitator inlet pressure on the sensors, thus ensuring the validity of the sensor data. A BP neural network model is established using intermediate variables (tilt angle offset and electrostatic precipitator center of gravity downward displacement distance) that can reflect the influence of various operating conditions on the electrostatic precipitator body. An external intelligent controller is used as the control system of this invention.

[0010] This invention provides a method for predicting the collapse risk of an electrostatic precipitator, comprising the following steps:

[0011] S1 uses sensors to collect data and monitor the risk indicators of the electrostatic precipitator;

[0012] S2, the collected data is filtered to filter out the impact of the inlet pressure fluctuation of the electrostatic precipitator on the equipment amplitude;

[0013] S3. Construct a BP neural network model as the basis for identifying the risk of collapse of the electrostatic precipitator body;

[0014] S4, determine whether the displacement exceeds the limit, and adjust the threshold of the alarm condition triggered by the BP model according to the result;

[0015] S5 uses a constructed BP neural network model to predict the collapse risk of electrostatic precipitators.

[0016] Furthermore, the sensor collects data, including the following steps:

[0017] S101 uses an accelerometer to measure the deflection angle of the load-bearing column;

[0018] S102 uses a settlement sensor to measure electrotropic displacement;

[0019] S103 uses a displacement sensor to measure the displacement of the load-bearing structure.

[0020] Furthermore, the collected data undergoes filtering processing, including the following steps:

[0021] S201, selecting the Haar function as the wavelet basis, decomposing to J layers, and performing discrete wavelet transform on all noisy signals f(t) acquired by the accelerometer to obtain a series of wavelet decomposition coefficients d. j,k :

[0022] d j，k = <f(t),ψ j，k (t)>

[0023] Where ψj , k (t) is L 2 An orthonormal basis of (R), <.> denotes inner product operation, d j , k Let denot be the wavelet coefficient at position k on scale j.

[0024] S202, Select a suitable initial threshold and process it to obtain the improved threshold:

[0025] The general threshold (sqtwolog) is selected as the initial threshold, and the initial threshold T corresponding to the wavelet coefficients of the J-th layer is... j for

[0026]

[0027] Where σ = median(W) HH ) / 0.6475, W HH These are the orthogonal wavelet coefficients of the noise in the high-frequency subband, and median indicates taking the median value;

[0028] The initial threshold is improved using the correlation coefficient, resulting in an improved threshold T. i :

[0029]

[0030] Where α represents the weight coefficient, which is between 0 and 1;

[0031] S203, Select a suitable threshold function and apply it to the series of wavelet coefficients d obtained from S201. j,k The wavelet coefficients are then processed to obtain the estimated wavelet coefficients.

[0032] A hard threshold function is selected as the threshold function, and the estimated wavelet coefficients are obtained after processing with the threshold function.

[0033]

[0034] Where T represents the improved threshold;

[0035] S204, using the estimated wavelet coefficients from S203 Wavelet reconstruction is performed to obtain the denoised signal f′(t), which represents the tilt angle and displacement output by the sensor at a certain moment under the influence of only the inlet pressure of the electrostatic precipitator.

[0036] Furthermore, the BP neural network model is constructed, including the following steps:

[0037] S301, Initialize the parameters in the neural network;

[0038] S302, obtain the correspondence between the downward shift of the center of gravity and the ash accumulation mass, and input the sample;

[0039] S303, activate forward propagation. The sample is introduced from the input layer, processed layer by layer by the hidden layers, and then passed to the output layer to obtain the error value between the output of the output layer and the actual output. Then, the back propagation of the error begins.

[0040] S304, activate backpropagation, calculate the error terms of the output unit and the hidden unit according to the loss function E(θ), and update the parameters such as weights and biases in the neural network;

[0041] S305, output samples, determine whether training is successful. If training is successful, store the current weights and predict the downward shift distance ΔL of the center of gravity in the range of 0 to 2Δm based on the weights. Output the corresponding curve of ΔL and Δm, and output the time required for Δm to reach the critical alarm value. If training fails, determine whether the termination condition is met. If not, repeat steps S303, S304, and S305 until the termination condition is met.

[0042] Further, the steps to determine whether displacement exceeds the limit are as follows:

[0043] The relationship between the horizontal displacement S collected by the displacement sensor and the length d of the load-bearing column and the tilt angle α of the load-bearing column is calculated using the following expression:

[0044] S>d*tanα

[0045] If the above expression is true, then the displacement is determined to be excessive.

[0046] Furthermore, the risk of electrostatic precipitator collapse is predicted using a BP neural network model, including the following steps:

[0047] S501, check if the displacement has been judged to be excessive. If it has been judged to be excessive, reduce the critical alarm value of Δm to 0.8*7‰d.

[0048] S502, based on the curve of the downward shift distance ΔL of the center of gravity as a function of Δm, outputs the corresponding relationship curve of the downward shift distance ΔL of the center of gravity in a wider range (0-2 Δm);

[0049] S503, based on the corresponding relationship curve and the current data collected by the sensor, determines the time length between the Δm value corresponding to ΔL under the current working condition and the critical alarm value, that is, predicts the time length until the current working condition continues to operate until the collapse danger state.

[0050] The present invention also provides a system for predicting the collapse risk of an electrostatic precipitator, comprising:

[0051] Sensors, including accelerometers, displacement sensors, and settlement sensors, are used to collect risk data of the electrostatic precipitator, including the deflection angle of the load-bearing column, vertical displacement, and displacement offset of the load-bearing structure.

[0052] The noise reduction module is used to filter out the amplitude fluctuations caused by pressure fluctuations at the inlet of the electrostatic precipitator.

[0053] The neural network module, including a BP neural network, is used to generate curves corresponding to risk factors and predict collapse risk.

[0054] The intelligent controller is used to calculate and transmit the results to the DCS system via communication protocols and upload them to the plant-level monitoring information system (SIS).

[0055] Furthermore, the intelligent controller is an external device, ensuring the security and independence of the DCS system, and the intelligent controller can be disconnected at any time.

[0056] Furthermore, the acceleration sensor and displacement sensor are installed at the top of the load-bearing column, and the settlement sensor is installed at the bottom ash outlet of the ash hopper.

[0057] The advantages of this invention are:

[0058] (1) Reduce the workload of inspection of the electrostatic precipitator body and save labor costs in thermal power generation;

[0059] (2) It has improved the safety and risk controllability in the power production process;

[0060] (3) Real-time monitoring and timely cleaning of the accumulated dust in the electrostatic precipitator extends the service life of the equipment. Attached Figure Description

[0061] Figure 1 Flowchart of the electrostatic precipitator collapse risk prediction system;

[0062] Figure 2 A flowchart illustrating the process of filtering data using wavelet transform thresholding for noise reduction.

[0063] Figure 3 A flowchart for establishing a BP neural network model. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Due to the unique nature of thermal power production processes, applying existing vibration monitoring technology to predict the risk of electrostatic precipitator collapse requires a series of improvements and overcoming the following technical challenges.

[0066] 1. Sensor noise reduction

[0067] The vibration of the electrostatic precipitator body is mainly generated in its own flue gas purification process. The magnitude of this vibration depends on the influence of parameters such as the flue gas temperature, pressure, and flow rate of the working medium at the inlet of the electrostatic precipitator.

[0068] Because the pressure fluctuations at the inlet of the electrostatic precipitator are irregular changes influenced by a variety of unstable factors, the signals collected by the data acquisition device are also more prone to fluctuations, thus affecting the effectiveness of the collected data.

[0069] This invention employs an improved wavelet transform threshold noise reduction method to filter the information collected by the sensor, thereby filtering out the influence of the electrostatic precipitator inlet pressure fluctuation on the equipment amplitude and retaining only the influence of the electrostatic precipitator inlet pressure on the sensor, thus ensuring the validity of the sensor-collected data.

[0070] 2. Establish a mathematical model for the risk prediction system.

[0071] To accurately predict the collapse risk of an electrostatic precipitator (ESP) and ensure the effectiveness and scientific rigor of the risk prediction system, it is necessary to comprehensively consider the impact of various factors on the ESP. Given that these influencing factors mostly change with the operating conditions, the mathematical model for the risk prediction system should consider and reconstruct the impact of as many factors as possible on the ESP collapse.

[0072] This invention establishes a BP neural network model using intermediate variables that reflect the impact of various operating conditions on the electrostatic precipitator: tilt angle offset and the downward displacement of the electrostatic precipitator's center of gravity. Simply put, the model follows the principle that the more dust accumulates, the heavier the electrostatic precipitator becomes, and the greater the downward displacement of its center of gravity. Because the supporting columns are rigid materials, they do not compress due to changes in load; the change is only reflected in the tilt angle of the supporting columns. Therefore, the more dust accumulates, the greater the offset (tilt angle) of the supporting columns, and the greater the risk of collapse of the electrostatic precipitator.

[0073] 3. System usability and security

[0074] To balance practicality and security, an external intelligent controller was adopted as the control system of this invention. Using this approach, complex calculations can be performed within the external intelligent controller, while the calculation results are transmitted to the DCS system and uploaded to the plant-level monitoring information system (SIS) via a communication protocol.

[0075] The advantages of this approach are that it satisfies the requirement of integrating the invention into the DCS system, ensuring the system's practicality, while also guaranteeing the DCS system's security and independence. Furthermore, the add-on system can be disconnected at any time, and corresponding parameters and strategies can be modified, while avoiding the risks associated with offline DCS configuration.

[0076] like Figure 1 The diagram shown is a system flowchart for predicting the collapse risk of electrostatic precipitators. The steps of the entire process are as follows:

[0077] S1 uses sensors to collect data on the deflection angle of the electrostatic precipitator's supporting column, the vertical displacement of the main body, and the displacement offset of the supporting structure.

[0078] This system uses three types of sensors: an acceleration sensor to measure the deflection angle of the load-bearing column, a settlement sensor to measure the vertical displacement of the electrostatic precipitator body, and a displacement sensor to measure the displacement of the load-bearing structure and thus determine whether the load-bearing structure has deformed. The types and quantities of sensors in this embodiment are shown in the table below:

[0079]

[0080] The acceleration sensor and displacement sensor are fixed to the top of the load-bearing column using steel cable ties, with one acceleration sensor and one displacement sensor fixed to each load-bearing column; the settlement sensor is fixed at the bottom of the ash hopper, with one settlement sensor fixed to each ash hopper.

[0081] S2 performs filtering on the collected data.

[0082] like Figure 2 The diagram shows a flowchart of the wavelet transform threshold denoising method for filtering data. This method filters the information acquired by the sensor. The aim is to filter out the influence of electrostatic precipitator inlet pressure fluctuations on the equipment amplitude, retaining only the influence of the electrostatic precipitator inlet pressure on the sensor.

[0083] The basic principle of wavelet transform threshold denoising is that after a noisy signal is decomposed by wavelets, the energy of the useful signal is concentrated in a few wavelet coefficients with large amplitudes, while the energy of the noise component is contained in a large number of wavelet coefficients with smaller amplitudes. Therefore, a suitable threshold can be selected to process the wavelet coefficients. Wavelet coefficients with absolute values ​​less than the threshold are set to zero, while wavelet coefficients with absolute values ​​greater than or equal to the threshold are retained or shrunk. By processing the wavelet coefficients by setting them to zero or shrinking them (threshold function), the wavelet coefficients of the signal can be estimated. Wavelet reconstruction can be performed using the estimated wavelet coefficients to achieve the purpose of denoising.

[0084] Using a wavelet signal f1(t) (original signal) under a steady state at a certain moment, which is only affected by the inlet pressure P of the electrostatic precipitator, and the wavelet signal f2(t) (noisy signal) after the inlet pressure of the electrostatic precipitator changes by Δp and reaches P, the two signals will show obvious correlation characteristics in the output signal.

[0085] When f1(t) and f2(t) are processed in the same wavelet decomposition method, f1(t) and f2(t) are decomposed into wavelet coefficients containing multiple frequency bands. As more signal correlations (non-noise) are formed within the wavelet coefficients, higher correlations will be obtained, resulting in an increase in the correlation coefficient. However, when the useful signal decreases and the noise increases, the correlation degree of the wavelet coefficients will decrease.

[0086] The specific steps are as follows: Figure 2 As shown:

[0087] S201, selecting the Haar function as the wavelet basis, decomposing to J layers, and performing discrete wavelet transform on all noisy signals f(t) acquired by the accelerometer to obtain a series of wavelet decomposition coefficients d. j，k

[0088] The noisy signal f(t) describes the tilt angle (or displacement) output by the sensor at a certain moment under the dual influence of the change in inlet pressure of the electrostatic precipitator and the inlet pressure value of the electrostatic precipitator. Its discrete wavelet transform mathematical expression is as follows:

[0089]

[0090] Where, ψ j，k (t)=2 j / 2 ψ(2 j -k), <·> indicates inner product operation.

[0091] When ψ j，k (t) is L 2 If (R) is an orthogonal basis, then in the above transformation, the orthogonal wavelet basis signal f(t) can be simplified to:

[0092]

[0093] Where, d j，k = <f(t),ψ j，k (t)> is the discrete wavelet transform of f(t), also known as the wavelet coefficient at position k at scale j.

[0094] The wavelet basis is chosen to be the Haar function. The Haar function is an orthogonal wavelet function with compact support used in wavelet analysis, and it is also the simplest wavelet function. Its definition is as follows:

[0095]

[0096] S202, Select an appropriate initial threshold T j The initial threshold is processed to obtain the improved threshold;

[0097] The general threshold (sqtwolog) is selected as the initial threshold, and the initial threshold T corresponding to the wavelet coefficients of the J-th layer is... j It can be represented as

[0098] Where, σ = median(W) HH ) / 0.6475, W HH These are the orthogonal wavelet coefficients of the noise in the high-frequency subband, with median indicating the median value.

[0099] Improve the initial threshold using the correlation coefficient.

[0100] Correlation coefficient:

[0101]

[0102] Among them, C 1，2 (p) is the covariance formed by d1 and d2 under time delay P, and σ1 and σ2 are the variance values ​​belonging to d1 and d2.

[0103] Improved threshold T i :

[0104]

[0105] Where α represents the weight coefficient, which is between 0 and 1, and T j This is the initial threshold.

[0106] S203, Select a suitable threshold function For the series of wavelet coefficients d obtained from ① j，k The wavelet coefficients are then processed to obtain the estimated wavelet coefficients.

[0107] We choose a hard threshold function as the threshold function, and its expression is:

[0108]

[0109] Where, d j，k These are the original wavelet coefficients. These are the estimated wavelet coefficients obtained after processing with a threshold function, where T represents the improved threshold.

[0110] S204, using the wavelet coefficients estimated in step three. (Write and reuse) Perform wavelet reconstruction to obtain the denoised signal f′(t). The signal f′(t) describes the tilt angle and displacement output by the sensor at a certain moment under the influence of only the inlet pressure of the electrostatic precipitator.

[0111] S3 constructs a BP neural network model as the basis for identifying the risk of collapse of the electrostatic precipitator body.

[0112] like Figure 3 As shown, building a BP neural network model mainly involves two stages:

[0113] In the first stage, forward propagation is activated. Input samples are passed from the input layer, processed layer by layer by the hidden layers, and then passed to the output layer to obtain the error value between the output of the output layer and the actual output. Then, the backpropagation of the error begins.

[0114] In the second stage, backpropagation is activated, and the error signal is used as the basis for correcting the weights of each unit. Based on the error, various parameters of the BP neural network model are generated.

[0115] The specific implementation process is illustrated using a single-layer neural network as an example. First, random weights are set to activate the forward propagation. Then, the weight values ​​are corrected in reverse based on the error. Finally, the weight values ​​are used to calculate the curve within the range of P.

[0116] S301, Initialize the parameters in the neural network;

[0117] S302, obtain the correspondence between the downward shift of the center of gravity and the ash accumulation mass, and input the sample;

[0118] S303, activate forward propagation, initialize the BP neural network parameters, let the input layer data be P, the parameters from the input layer to the hidden layer be w, b1, the parameters from the hidden layer to the output layer be v, b2, the activation function be g1 g2, each weight is a random real number in the range [-1, 1], each bias is a random real number in the range [0, 1], and then the forward propagation begins, the commonly used Sigmoid function is selected as the activation function.

[0119]

[0120] Therefore, from the input layer to the hidden layer: net1 = w T x+b1,h=g1(net1)

[0121] Hidden layer to output layer: net2 = v T x+b2,

[0122] In summary, the entire model has the following:

[0123]

[0124] Output The predicted value of the risk coefficient R(y):

[0125] Loss function:

[0126] S304 activates backpropagation, and calculates the error terms of the output unit and the hidden unit based on the loss function.

[0127] 1) The error term of the output unit, i.e., calculating the partial derivative of the loss function with respect to the output unit, can be obtained according to the chain rule:

[0128]

[0129]

[0130] 2) The error term of the hidden unit, i.e., calculating the partial derivative of the loss function with respect to the hidden unit, according to the chain rule, we have:

[0131]

[0132]

[0133] 3) Update the weights and biases in the neural network.

[0134] Output unit parameter update:

[0135]

[0136]

[0137] Hidden unit parameter update:

[0138]

[0139]

[0140] η represents the learning efficiency, and k = 1, 2, ..., n represents the number of iterations.

[0141] S305, repeat S303, S304, S305 until the loss function is less than the given threshold or the number of iterations is exhausted, output the optimal parameters of the BP neural network at this time, and complete the model construction.

[0142] S4 determines whether a displacement exceeding the limit has occurred and adjusts the threshold for triggering the alarm in the BP model based on the result. If an exceedance is detected, an alarm is triggered. The displacement sensor in the system measures the lateral displacement of the electrostatic precipitator body. The relationship between the horizontal displacement S collected by the displacement sensor and the length d of the supporting column and the tilt angle α of the supporting column is as follows:

[0143] S>d*tanα

[0144] If the above expression is true, it indicates that a displacement exceeding the limit has occurred, signifying that the load-bearing column has deformed. In the subsequently constructed neural network model, the threshold for the alarm trigger condition will automatically decrease to 80% of the original condition. It should be noted that the threshold for the alarm condition can be manually adjusted. In this embodiment, after actual testing, a relatively stable fixed threshold of 80% was selected as the scheme for actual operation to ensure the effectiveness of collapse monitoring.

[0145] S5 uses a constructed BP neural network model to predict the collapse risk of electrostatic precipitators. The specific steps are as follows:

[0146] S501, check if the displacement has been judged to be excessive. If it has been judged to be excessive, reduce the critical alarm value of Δm to 0.8*7‰d.

[0147] S502, based on the curve of the downward shift distance ΔL of the center of gravity as a function of Δm, outputs the corresponding relationship curve of the downward shift distance ΔL of the center of gravity in a wider range (0-2 Δm);

[0148] S503, based on the corresponding relationship curve and the current data collected by the sensor, determines the time length between the Δm value corresponding to ΔL under the current working condition and the critical alarm value, that is, predicts the time length until the current working condition continues to operate until the collapse danger state.

[0149] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the collapse risk of an electrostatic precipitator, characterized in that, Includes the following steps: S1 uses sensors to collect data and monitor the risk indicators of the electrostatic precipitator; S2, The collected data is filtered to remove the impact of electrostatic precipitator inlet pressure fluctuations on equipment amplitude, including the following steps: S201, selecting the Haar function as the wavelet basis, decomposing it into J layers, and processing all noisy signals acquired by the accelerometer. Performing discrete wavelet transform yields a series of wavelet decomposition coefficients. : in yes orthonormal basis This represents the inner product operation. Let be the wavelet coefficient at position k on scale j; S202, Select a suitable initial threshold and process it to obtain the improved threshold: The general threshold sqtwolog is selected as the initial threshold, and the initial threshold corresponding to the wavelet coefficients of the J-th layer is... for in , These are the orthogonal wavelet coefficients of the noise in the high-frequency subband, and median indicates taking the median value; The initial threshold is improved using the correlation coefficient. : in, This represents the weighting coefficient, which is between 0 and 1. S203, Select a suitable threshold function and apply it to the series of wavelet coefficients obtained from S201. The wavelet coefficients are then processed to obtain the estimated wavelet coefficients. ; A hard threshold function is selected as the threshold function, and the estimated wavelet coefficients are obtained after processing with the threshold function. : Where T represents the improved threshold; S204, using the estimated wavelet coefficients from S203 Wavelet reconstruction is performed to obtain the denoised signal. ,Signal This represents the tilt angle and displacement output by the sensor at a given moment, under the influence of only the inlet pressure of the electrostatic precipitator. S3. Construct a BP neural network model as the basis for identifying the risk of collapse of the electrostatic precipitator body; S4, determine whether a displacement exceeding the limit has occurred, and adjust the threshold for triggering the alarm condition in the BP model based on the determination result, specifically: Calculate the horizontal displacement S collected by the displacement sensor in relation to the length d of the load-bearing column and the tilt angle of the load-bearing column. The relationship between them is expressed as follows: If the above expression is true, then the displacement is determined to be excessive; S5 uses a constructed BP neural network model to predict the collapse risk of electrostatic precipitators.

2. The method for predicting the collapse risk of an electrostatic precipitator according to claim 1, characterized in that, The sensor collects data, including the following steps: S101 uses an accelerometer to measure the deflection angle of the load-bearing column; S102 uses a settlement sensor to measure electrotropic displacement; S103 uses a displacement sensor to measure the displacement of the load-bearing structure.

3. The method for predicting the collapse risk of an electrostatic precipitator according to claim 1, characterized in that, Constructing a BP neural network model includes the following steps: S301, Initialize the parameters in the neural network; S302, obtain the correspondence between the downward shift of the center of gravity and the ash accumulation mass, and input the sample; S303, activate forward propagation. The sample is introduced from the input layer, processed layer by layer by the hidden layers, and then passed to the output layer to obtain the error value between the output of the output layer and the actual output. Then, the back propagation of the error begins. S304, activate backpropagation, based on the loss function. It calculates the error terms of the output units and the hidden units, and updates the parameters such as weights and biases in the neural network. S305, output samples, determine whether training is successful. If training is successful, store the current weights and predict the center of gravity shift distance ΔL within the range of 0 to 2Δm based on the weights. Output the curve of the correspondence between ΔL and Δm, and output the time required for Δm to reach the critical alarm value. If training fails, determine whether the termination condition is met. If not, repeat steps S303, S304, and S305 until the termination condition is met. Here, Δm represents the ash accumulation quality.

4. The method for predicting the collapse risk of an electrostatic precipitator according to claim 1, characterized in that, Predicting the collapse risk of electrostatic precipitators using a BP neural network model includes the following steps: S501, check if the displacement has been judged to be excessive. If it has been judged to be excessive, reduce the critical alarm value of Δm to 0.8*7‰d. S502, based on the curve of the change of the center of gravity downward displacement ΔL with the ash accumulation mass Δm, outputs the corresponding relationship curve of the center of gravity downward displacement ΔL in a wider range (0-2Δm); S503, based on the corresponding relationship curve and the current data collected by the sensor, determines the time length from the Δm value corresponding to ΔL under the current working condition to the critical alarm value, that is, predicts the time length for the current working condition to continue operating until the collapse danger state.

5. A system for predicting the collapse risk of an electrostatic precipitator, used to implement the method for predicting the collapse risk of an electrostatic precipitator as described in any one of claims 1 to 4, characterized in that, include: Sensors, including accelerometers, displacement sensors, and settlement sensors, are used to collect risk data of the electrostatic precipitator. This risk data includes the deflection angle of the load-bearing column, vertical displacement, and displacement offset of the load-bearing structure. The noise reduction module is used to filter out the amplitude fluctuations caused by the pressure fluctuations at the inlet of the electrostatic precipitator. The neural network module, including a BP neural network, is used to generate curves corresponding to risk factors and predict collapse risk. The intelligent controller is used to calculate and transmit the results to the DCS system via communication protocols and upload them to the plant-level monitoring information system (SIS).

6. The electrostatic precipitator collapse risk prediction system according to claim 5, characterized in that, The intelligent controller is an external device, ensuring the security and independence of the DCS system, and can be disconnected at any time.

7. The electrostatic precipitator collapse risk prediction system according to claim 5, characterized in that, Accelerometers and displacement sensors are installed at the top of the load-bearing column, while settlement sensors are installed at the bottom of the ash hopper.

Images