A method for detecting abnormality of centrifugal fan equipment in cement plants

Abstract

The present invention discloses a method for detecting abnormalities of centrifugal fan equipment in a cement plant, which is characterized by comprising the following steps: step 1, collecting data; step 2, using a deep learning algorithm to generate an adversarial network and a variational autoencoder network to construct a detection model; step 3, training the detection model, and calculating the final abnormal threshold value during training through a threshold selection calculation module; in step 3, a generalized Pareto distribution with parameters is used to fit the tail of the probability distribution, and the abnormal score threshold step is learned; and step 4, online detection is performed through the trained detection model, and the evaluation result is judged according to the abnormal threshold and output for display. Not only can the fan status be correctly evaluated through the overall abnormal evaluation score, but also the contribution of each data dimension to the score can be shown to the user, so that the user can have a better understanding of the specific situation of the abnormality.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent detection, and in particular relates to a method for detecting abnormality of centrifugal fan equipment in a cement plant. Background Art

[0002] In the cement industry, centrifugal fans are essential key equipment in the production process, such as grate cooler fan group, kiln exhaust fan, kiln exhaust fan, high-temperature fan, powder selector, etc. The stability of the fan determines the stable operation of the entire production line. Therefore, it is necessary to regularly inspect and monitor each fan to ensure the stable operation of the fan and try to find fan operation problems and hidden dangers as soon as possible.

[0003] The prior art has the following defects and shortcomings.

[0004] 1. On-site manual inspection and patrol: The cement plant fan equipment is manually inspected and patrolled on-site. For important fan equipment, the inspection frequency is higher, which requires a large number of personnel and increases the potential safety hazards on site. At the same time, it is difficult to detect internal problems of fan equipment in real time by manual inspection. Many abnormal faults are not intuitive and difficult to detect manually.

[0005] 2. Central control room monitoring equipment signals: The central control room of the cement plant conducts linkage control and status monitoring of the entire plant's equipment, but the amount of monitoring is relatively small. It only collects the current, frequency, wind pressure and speed of the fan equipment. It is difficult to conduct a full range of monitoring of the fan equipment based on these parameters, and there is no matching alarm strategy. It only alarms for abnormalities in a single parameter, and cannot correctly and comprehensively evaluate the operating status of the fan. Summary of the invention

[0006] The purpose of the present invention is to provide a method for detecting abnormalities of centrifugal fan equipment in a cement plant to solve the technical problem that in the prior art, only a single parameter of the fan can be judged as abnormal, and the fan equipment cannot be fully monitored and a matching alarm strategy cannot be adopted, so the operating status of the fan cannot be correctly and comprehensively evaluated.

[0007] The method for detecting abnormality of centrifugal fan equipment in a cement plant comprises the following steps:

[0008] Step 1: Collect data;

[0009] Step 2: Use deep learning algorithms to generate adversarial networks and variational autoencoder networks to build a detection model;

[0010] Step 3: Train the detection model and calculate the final anomaly threshold during training through the threshold selection calculation module;

[0011] Step 4: Perform online detection using the trained detection model, determine the evaluation results based on the anomaly threshold, and output them for display.

[0012] Preferably, the step three specifically includes:

[0013] Data preprocessing: standardize and transform the data set and split it into sequences;

[0014] Model training: Send a period of multivariate time series to the model training to learn the detection model, which captures the normal pattern of the multivariate time series and outputs the anomaly assessment score for each observation;

[0015] Threshold calculation is performed during training. During offline training, a multivariate time series of N observations is used to calculate an anomaly assessment score for each observation. Then all anomaly scores form a univariate time series {S1, S2, …, SN}. Then, the above univariate time series is calculated according to the extreme value theory (EVT) principle, and the anomaly threshold is set offline.

[0016] Preferably, in step 3, a generalized Pareto distribution with parameters is used to fit the tail of the probability distribution and learn the abnormal score threshold. In this method, the GPD function is adjusted to:

[0017]

[0018] Where th is the initial anomaly score threshold, γ and β are the characteristic parameter and scale parameter of the generalized Pareto distribution, respectively, S is an arbitrary value of the univariate time series {S1, S2, …, SN}, and the part below the threshold th is recorded as th-S; s is the settable anomaly difference, and the result of th-S is greater than the set anomaly difference s, and the probability P is calculated;

[0019] Then use the maximum likelihood estimate and The value of , and then calculate the final abnormal threshold, the formula is as follows:

[0020]

[0021] Where q is the expectation of S<th, N′ is the number of observations, and N′ th is the number of Si that is less than the threshold th, where Si is an element of the univariate time series, i = 1, 2, …, N.

[0022] Preferably, the detection model calculates the observation result xt, where xt includes the reconstruction probabilities calculated for various data and the overall abnormality assessment score of the wind turbine formed by the reconstruction probabilities of various data.

[0023] Preferably, the detection model adopts adversarial learning anomaly detection, and the detection model is built on a type of generative adversarial network GAN; the generative adversarial network consists of two types of competing networks, a generator and a discriminator. The generator maps a random variable z sampled from a latent distribution to an input data space, and the discriminator attempts to distinguish between real data samples x(i) and samples G(z) generated by the generator. The two types of networks compete with each other. The generator attempts to generate samples similar to real data, while the discriminator attempts to distinguish between samples generated by the generator and real data samples.

[0024] Preferably, the generative adversarial network GAN simultaneously learns an encoder network E, which is a variational autoencoder network that maps data samples x to a latent space z during training; calculating the latent representation of a data point x in such a model is simply by passing x to the encoder network E.

[0025] Preferably, in step 4, the online detection module stores the trained model, inputs the preprocessed sequence, and obtains a new observation result xt. If the anomaly assessment score of xt is lower than the anomaly threshold, the judgment result of xt is output as abnormal, otherwise it is normal; if the judgment result is abnormal, the contribution of each data dimension of xt is estimated and sorted to explain the abnormal situation, and the contribution of the data dimension is the reconstruction probability of the data; when displaying, the reconstruction probabilities corresponding to each data are sorted in ascending order to form a list A.

[0026] Preferably, in step one, data is collected through sensors and data collectors to form a data set, and the data include the fan set speed, three stator temperatures of the fan motor A, B, and C, the fan inlet temperature, the fan outlet temperature, the fan outlet pressure, the fan inlet pressure, the fan free end vibration, the fan positioning end vibration, the fan inverter speed feedback, the fan inverter current, voltage and power feedback, and the fan motor free end and output end vibration.

[0027] The present invention has the following advantages: the present invention realizes the training and calculation of the abnormal threshold through the above-mentioned method, so that the abnormal threshold is most suitable for the fan system corresponding to the current training data, and the abnormal evaluation score is used to determine the fan state. The most reliable judgment basis is provided, and the automatic training and determination of the abnormal threshold is realized through this method, thereby improving the training efficiency.

[0028] Monitoring the fan through the present invention can conduct all-round monitoring of the fan's operating status in real time, and realize intelligent evaluation of the fan status in combination with the detection model to form a better alarm strategy. It can not only correctly evaluate the fan status through the overall abnormal evaluation score, but also show the user the contribution of each data dimension to the score, so that the user can have a better understanding of the specific situation of the abnormality, thereby providing sufficient clues and basis for the staff to understand the abnormal status and subsequent processing solutions.

[0029] The detection model provided by the present invention enhances the model's ability to reconstruct and identify the normal state of the wind turbine through adversarial learning, increases the difference between the abnormal state and the normal state, and combines a variational autoencoder network to ensure that only the main part of the information is reconstructed and identified by the generator, which facilitates the calculation of latent representations and improves the training efficiency and the detection ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The present invention is a flow chart of a method for detecting abnormality of centrifugal fan equipment in a cement plant.

[0031] Figure 2 It is a schematic diagram of the detection model in the present invention. DETAILED DESCRIPTION

[0032] The specific implementation modes of the present invention will be further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0033] like Figure 1-2 As shown, the present invention provides a method for detecting abnormality of centrifugal fan equipment in a cement plant, comprising the following steps.

[0034] Step 1: Collect data.

[0035] Data is collected through sensors and data collectors to form a data set, which includes the fan set speed, three stator temperatures of the fan motor A, B, and C, the fan inlet temperature, the fan outlet temperature, the fan outlet pressure, the fan inlet pressure, the fan free end vibration, the fan positioning end vibration, the fan inverter speed feedback, the fan inverter current, voltage and power feedback, and the fan motor free end and output end vibration.

[0036] Step 2: Use deep learning algorithms to generate adversarial networks and variational autoencoder networks to build detection models.

[0037] Generative adversarial networks (GANs) consist of two competing networks, a generator G and a discriminator D. The generator G maps random variables z sampled from a latent distribution (Gaussian distribution or uniform distribution) to the input data space. The discriminator D tries to distinguish between real data samples x(i) and samples G(z) generated by the generator G. The two networks compete with each other, G tries to generate samples similar to real data, while D tries to distinguish between samples generated by the generator and real data samples.

[0038] This method uses an adversarial learning anomaly detection model, which is built on a class of generative adversarial networks (GANs) that simultaneously learn an encoder network E, which is a variational autoencoder network that maps data samples x to a latent space z during training. The latent representation of a data point x in such a model is simply calculated by passing x to the encoder network.

[0039] Figure 2 This is a schematic diagram of the final GAN ​​model, where Dzz, Dxz and Dxx represent discriminators, G represents generators, and E represents encoders. These networks learn simultaneously during the training process to train the optimal discriminator Dxz. Where z is the latent distribution of the input data, and x is the real data space.

[0040] After the detection model is trained, it can determine in real time whether the device is in an abnormal state, and identify abnormal situations, including items, events, or observations that do not conform to the expected pattern or other item standards in the data set. The detection model uses abnormal indicators as evaluation indicators, and the system automatically determines whether the device is in an abnormal state, and users can review it retrospectively.

[0041] Step 3: Train the detection model and calculate the final anomaly threshold during training through the threshold selection calculation module.

[0042] This step specifically includes:

[0043] Data preprocessing: The data set is normalized and then divided into sequences using a sliding window of length t+1.

[0044] Model training: After preprocessing, a period of multivariate time series is sent to model training to learn a detection model that captures the normal pattern of the multivariate time series and outputs an anomaly assessment score for each observation.

[0045] Threshold calculation during training: During offline training, using a multivariate time series of N observations, we calculate an anomaly assessment score for each observation. Then all anomaly scores form a univariate time series {S1, S2, …, SN}. Then, according to the extreme value theory EVT principle, the above univariate time series is calculated and the anomaly threshold is set offline.

[0046] EVT is a statistical theory that aims to find the law of extreme values, which are usually located at the tail of the probability distribution. It uses a generalized Pareto distribution (GPD) with parameters to fit the tail of the probability distribution and learn the abnormal score threshold. This method adjusts the GPD function to:

[0047]

[0048] Where th is the initial anomaly score threshold, γ and β are the characteristic parameter and scale parameter of GPD (generalized Pareto distribution) respectively, S is an arbitrary value of the univariate time series {S1, S2, …, SN}, and the part below the threshold th is recorded as th-S; s is the settable anomaly difference, and the result of th-S is greater than the set anomaly difference s, and the probability P is calculated.

[0049] Then use the maximum likelihood estimate and The value of is then used to calculate the final abnormal threshold, and the formula is as follows.

[0050]

[0051] Where q is the expectation of S<th, N′ is the number of observations, and N′ th is the number of Si that is less than the threshold th, where Si is an element of the univariate time series, i = 1, 2, …, N.

[0052] By implementing the training and calculation of the abnormal threshold as described above, the abnormal threshold is most suitable for the fan system corresponding to the current training data, providing the most reliable judgment basis for determining the fan status by the abnormal evaluation score, and the automatic training and determination of the abnormal threshold is implemented by this method, thereby improving the training efficiency.

[0053] The detection model calculates the observation result xt, which includes the reconstruction probability calculated for each data, and the overall abnormal evaluation score of the wind turbine formed by the reconstruction probability of each data. The abnormal evaluation score is used to compare with the final abnormal threshold th F Compare and determine the judgment result of xt.

[0054] Step 4: Perform online detection using the trained detection model.

[0055] The online detection module stores the trained model, inputs the preprocessed sequence, and obtains a new observation result xt. If the anomaly assessment score of xt is lower than the anomaly threshold, the judgment result of xt is output as an anomaly, otherwise it is normal. If the judgment result is abnormal, the contribution of each data dimension of xt (i.e., reconstruction probability) is estimated and sorted to explain the abnormal situation. When displayed, the data is sorted in ascending order according to the reconstruction probability corresponding to each data, and a list A is formed. The higher the ranking of each type of data in list A, the greater its contribution to the judgment result of xt. The ordered list A is presented to the operator as an explanation of the anomaly. The top few dimensions can provide sufficient clues for the operator to understand and handle the detected entity anomalies.

[0056] Monitoring the fan through the present invention can conduct all-round monitoring of the fan's operating status in real time, and realize intelligent evaluation of the fan status in combination with the detection model to form a better alarm strategy. It can not only correctly evaluate the fan status through the overall abnormal evaluation score, but also show the user the contribution of each data dimension to the score, so that the user can have a better understanding of the specific situation of the abnormality, thereby providing sufficient clues and basis for the staff to understand the abnormal status and subsequent processing solutions.

[0057] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for detecting abnormality of centrifugal fan equipment in a cement plant, characterized in that: The following steps are involved: Step 1: Collect data; Step 2: Use deep learning algorithms to generate adversarial networks and variational autoencoder networks to build a detection model; Step 3: Train the detection model and calculate the final anomaly threshold during training through the threshold selection calculation module; Step 4: Perform online detection using the trained detection model, determine the evaluation results based on the anomaly threshold, and output them for display; The detection model adopts adversarial learning anomaly detection, and the detection model is built on a type of generative adversarial network GAN; the generative adversarial network consists of two types of competing networks, the generator and the discriminator. The generator maps the random variable z sampled from the latent distribution to the input data space, and the discriminator attempts to distinguish between the real data sample x(i) and the sample G(z) generated by the generator. The two types of networks are competitive with each other. The generator attempts to generate samples similar to real data, while the discriminator attempts to distinguish between the samples generated by the generator and the real data samples. The generative adversarial network GAN also learns an encoder network E, which is a variational autoencoder network that maps real data samples x to latent space z during training; calculating the latent representation of data point x in such a model is simply by passing x to the encoder network E; The detection model calculates the observation result xt, which includes the reconstruction probability calculated for each data and the overall abnormality assessment score of the wind turbine formed by the reconstruction probability of each data. The abnormality assessment score is used to compare with the final abnormality threshold to determine the judgment result; The discriminator includes three discriminators Dzz, Dxz and Dxx, the input of the discriminator Dzz includes the latent representation in the latent space z and the latent representation processed by the generator and the encoder network in sequence to obtain E(G(z)); the input of the discriminator Dxx includes the real data sample x and the real data sample processed by the encoder network and the generator in sequence to obtain G(E(x)); the input of the discriminator Dxz includes the latent representation in the latent space z and the real data sample processed by the encoder network to obtain E(x); In the step 4, the online detection module stores the trained model, inputs the preprocessed sequence, and obtains a new observation result xt. If the anomaly assessment score of xt is lower than the anomaly threshold, the judgment result of xt is output as abnormal, otherwise it is normal; if the judgment result is abnormal, the contribution of each data dimension of xt is estimated and sorted to explain the abnormal situation, and the contribution of the data dimension is the reconstruction probability of the data; when displaying, the reconstruction probabilities corresponding to each data are sorted in ascending order to form a list A.

2. A method for detecting abnormality of centrifugal fan equipment in a cement plant according to claim 1, characterized in that: The step three specifically includes: Data preprocessing: standardize and transform the data set and split it into sequences; Model training: Send a period of multivariate time series to the model training to learn the detection model, which captures the normal pattern of the multivariate time series and outputs the anomaly assessment score for each observation; Threshold calculation is performed during training. During offline training, a multivariate time series of N observations is used to calculate an anomaly assessment score for each observation. Then all anomaly scores form a univariate time series {S1, S2, …, SN}. Then, the above univariate time series is calculated according to the extreme value theory (EVT) principle, and the anomaly threshold is set offline.

3. A method for detecting abnormality of centrifugal fan equipment in a cement plant according to claim 2, characterized in that: In step 3, the generalized Pareto distribution with parameters is used to fit the tail of the probability distribution and learn the abnormal score threshold. This method adjusts the GPD function to: Where th is the initial anomaly score threshold, γ and β are the characteristic parameter and scale parameter of the generalized Pareto distribution, respectively, S is an arbitrary value of the univariate time series {S1, S2, …, SN}, and the part below the threshold th is recorded as th-S; s is the settable anomaly difference, and the result of th-S is greater than the set anomaly difference s, and the probability P is calculated; Then use the maximum likelihood estimate and The value of , and then calculate the final abnormal threshold, the formula is as follows: where q is the expectation of S < th, N' is the number of observations, and N′ th is the number of Si that are less than the threshold th, Si is an element of a univariate time series, and i = 1, 2, …, N.

4. A method for detecting abnormality of centrifugal fan equipment in a cement plant according to claim 1, characterized in that: In the step 1, data is collected by sensors and data collectors to form a data set, and the data includes a given speed of the fan, three stator temperatures of the fan motor A, B, and C, a fan inlet temperature, a fan outlet temperature, a fan outlet pressure, a fan inlet pressure, a fan free end vibration, a fan positioning end vibration, a fan inverter speed feedback, a fan inverter current, voltage and power feedback, and a fan motor free end and output end vibration.

Images