Intelligent fire pump fault diagnosis method based on BP neural network

Abstract

The invention provides an intelligent fire pump fault diagnosis method based on a BP neural network, and the method comprises the steps: obtaining original basic data of an intelligent fire pump in a normal state and a fault state, and building a fault diagnosis database; performing data enhancement on the obtained original fault data by constructing a conditional expression Wasserstein generative adversarial network model with gradient penalty optimization so as to obtain enhanced fault data; performing data fusion on the enhanced fault data and the original basic data, and performing data dimension reduction and data feature extraction on the fused data through a principal component analysis method; and according to an output result of the pre-constructed intelligent fire pump fault diagnosis model, judging whether the intelligent fire pump has a fault and judging an abnormal point location. The fault diagnosis accuracy of the intelligent fire pump is effectively improved, potential faults can be found and repaired in time, and it is ensured that a fire fighting system can be rapidly started and effectively run under the emergency situation.

Description

Technical Field

[0001] The present application belongs to the technical field of fire pump fault diagnosis, and in particular to an intelligent fire pump fault diagnosis method based on BP neural network. Background Art

[0002] With the acceleration of urbanization and the increase in building density, the risk of fire is increasing. The reliability and efficiency of the fire protection system have become key factors in protecting people's lives and property. As the core equipment of the fire protection system, the fire pump is responsible for providing the necessary water pressure and water volume when a fire occurs. Its operating status directly affects the effectiveness of fire fighting. However, in actual applications, fire pumps may fail due to various reasons such as long-term use, improper maintenance or environmental factors, such as motor overheating, bearing wear, seal failure, etc. These problems will not only reduce the performance of the fire pump, but may also cause system failure in an emergency, thereby delaying the time of fire fighting and increasing the risk of casualties and property losses.

[0003] Although the existing disclosed technologies can also realize the fault diagnosis of fire pumps, most of them have the following technical disadvantages: 1. The data collected is of a single dimension, and mostly single acoustic emission signals, vibration signals or motor data are obtained, and then such signals are analyzed, without fully considering the changing relationship between environmental parameters, the current state of the fire pump and the operating parameters of the fire pump.

[0004] 2. Existing technologies have mostly studied water pump fault diagnosis, but less research has been done on fire pump fault diagnosis. Unlike other water pumps, which operate continuously, fire pumps are mostly in standby or full power / low speed operation when there is no fire. Since other water pumps are continuously in operation, the sample size of fault data and normal data is large, which makes fault diagnosis more difficult; however, the sample size of fire pump fault data is small, so it is necessary to make full use of data enhancement methods to make up for it during fault diagnosis.

[0005] 3. Some technologies use machine learning methods such as clustering and Bayesian networks to predict the fault type of the pump, but they are mostly classified according to a single historical fault type. They do not consider the coupling of fault types under multiple fault conditions and the fault types that have not been detected in the past, nor do they determine the fault point. Summary of the invention

[0006] In view of this, the present application aims to propose a fire pump fault diagnosis method, system, electronic device and storage medium to solve the problem that in actual working conditions, most intelligent fire pumps are in normal state, resulting in scarcity of fault samples, differences between fault data, and low accuracy of fault category identification.

[0007] To achieve the above purpose, the technical solution of this application is implemented as follows: In a first aspect, the present application provides an intelligent fire pump fault diagnosis method based on BP neural network, comprising: Acquire original basic data of the intelligent fire pump in normal state and fault state, and establish a fault diagnosis database, wherein the basic data at least includes environmental parameters, power supply parameters and fire pump sensor data; The acquired original fault data is enhanced by constructing a conditional Wasserstein generative adversarial network model with gradient penalty optimization to obtain enhanced fault data. The enhanced fault data is fused with the original basic data, and the fused data is subjected to data dimension reduction and data feature extraction through principal component analysis. According to the output result of the pre-constructed intelligent fire pump fault diagnosis model, it is judged whether the intelligent fire pump has a fault and the abnormal point is determined, wherein the intelligent fire pump fault diagnosis model is trained by the extracted data features.

[0008] Furthermore, the intelligent fire pump fault samples are collected and the fault data set is defined; It is known that the actual distribution obeyed by the fault samples, and there is a set of noise vectors obeying the Gaussian distribution. By establishing a mapping relationship model between the actual distribution and the Gaussian distribution, artificial samples that satisfy the real data distribution relationship are generated based on random noise.

[0009] Furthermore, the conditional Wasserstein generative adversarial network model is composed of a generator network and a discriminator network, wherein the generator network forms a new implicit representation by splicing and combining the noise vector and the category condition; in the discriminator network, the actual fault samples or artificial samples are input together with the category condition for discrimination; The difference between the actual distribution and the generated distribution is measured by the Wasserstein distance, the Kantorovich-Rubinstein dual form is adopted, and the category condition is introduced; By adding a gradient penalty term to the discriminator's loss function, the gradient norm of the discriminator is constrained.

[0010] Furthermore, in the conditional Wasserstein generative adversarial network model based on gradient penalty optimization, the loss functions of the generator and the discriminator are: ; ; In the formula, z represents noise samples, x represents fault samples, c represents category conditions, G(z) represents artificial samples, represents the actual distribution, represents the generated distribution, λ represents the weight coefficient, represents the interpolation between the actual sample and the generated sample, Indicates that the discriminator is The gradient norm on .

[0011] Furthermore, the data dimensionality reduction and data feature extraction of the fused data by principal component analysis method include: The fused data is normalized and SelectKBest is used for feature selection to select the features most relevant to the target variable. The data after feature selection is reduced in dimension by principal component analysis, and the principal component is used to detect anomalies to determine the fault location.

[0012] Furthermore, the interquartile range of each principal component is calculated, and the threshold of the outlier is determined according to the interquartile range. If the principal component value of a data point exceeds the threshold range, the data point is marked as an outlier.

[0013] Furthermore, an intelligent fire pump fault diagnosis model based on a Transformer structure is constructed, wherein the Transformer structure includes: The multi-head attention mechanism is configured to use a scaled inner product attention module to calculate the attention weights of feature embeddings; The forward network consists of two linear transformation layers and one nonlinear activation layer, which is used to map the features extracted by multi-head attention into more abstract features and adopts the Gaussian error linear unit activation function; Residual connections and layer regularization are set at the output of the multi-head attention mechanism and the forward network.

[0014] Compared with the prior art, the intelligent fire pump fault diagnosis method based on BP neural network described in this application has the following beneficial effects: The intelligent fire pump fault diagnosis method based on BP neural network described in the present application has a higher accuracy rate in identifying the fault category of the intelligent fire pump, effectively improving the accuracy of the intelligent fire pump fault diagnosis, being able to timely discover and repair potential faults, ensuring that the fire protection system can be quickly started and effectively operated in an emergency, and also improving the maintenance efficiency and reliability of the fire pump, extending the service life of the equipment, and reducing the cost of maintenance and replacement, thereby improving the level of public safety and protecting people’s lives and property. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings: Figure 1This is a flow chart of a method for fault diagnosis of an intelligent fire pump based on a BP neural network according to an embodiment of the present application; Figure 2 A schematic diagram of the basic structure of a conditional generative adversarial network described in an embodiment of the present application; Figure 3 This is a schematic diagram of the Transformer neural network intelligent fire pump fault diagnosis model framework described in the embodiment of the present application; Figure 4 Schematic diagram of the multi-head attention mechanism described in an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0017] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present application belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] Explanation of the name: Conditional Wasserstein Generative Adversarial Network Model, CWGAN-GP.

[0019] Principal Component Analysis, PCA.

[0020] Wasserstein distance, also known as the Earth Mover distance.

[0021] 1-Lipschitz condition, that is, Lipschitz continuity condition (Lipschitz continuity).

[0022] Interquartile range (IQR), also known as four-point difference.

[0023] In view of the problem that in actual working conditions, most intelligent fire pumps are in normal state, resulting in scarcity of fault samples, differences between fault data, and low accuracy of fault category identification, this embodiment provides an intelligent fire pump fault diagnosis method based on BP neural network, which aims to solve the fault diagnosis problem of fire water pumps running at low speed under non-firefighting state.

[0024] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0025] See also Figure 1 As shown, this embodiment provides an intelligent fire pump fault diagnosis method based on BP neural network, which specifically includes the following steps: Step S101, obtaining original basic data of the intelligent fire pump in normal state and fault state, and establishing a fault diagnosis database, wherein the basic data at least includes environmental parameters, power supply parameters and fire pump sensor data; Specifically, in this embodiment, by experimentally simulating normal working conditions and fault conditions, the environmental parameters (temperature, humidity, atmospheric pressure, etc.), power supply parameters (input voltage, input current, input power, etc.), and fire pump sensor parameters (inlet gate valve status, inlet pressure, inlet flange vibration value, anchor vibration value, operating noise, rotation speed, shaft temperature, outlet flange vibration value, outlet pressure, outlet gate valve status, etc.) of the intelligent fire pump are obtained to establish a fault diagnosis database.

[0026] In this step, by comprehensively considering the coupling relationship between environmental parameters, power supply parameters, vibration data, fire pump operating status and other factors, the corresponding fault diagnosis model input is established, breaking through the data limitations of a single signal and reducing the impact on the fault diagnosis model.

[0027] Step S102: Perform data enhancement on the acquired original fault data by constructing a conditional Wasserstein generative adversarial network model with gradient penalty optimization to obtain enhanced fault data; Specifically, in this embodiment, the imbalance problem between normal data and fault data samples is a key factor restricting the application effect of machine learning in the field of intelligent fire pump fault diagnosis. In order to overcome the defect that the traditional sampling method does not consider the overall distribution information of the data, a fault data enhancement method based on deep learning is proposed, that is, a conditional Wasserstein generative adversarial network model (CWGAN-GP) optimized by gradient penalty is established to achieve the category balancing goal of the sample library.

[0028] Assuming that n sets of intelligent fire pump fault samples are currently collected, the fault data set can be expressed as ,in, Represents the feature vector of the nth group of fault samples. It is known that the actual distribution of the fault sample x is Pr, and there is a set of noise vectors z that obey the Gaussian distribution Pz. By establishing a mapping relationship model between Pz and Pr, it is possible to generate artificial samples based on random noise that meet the distribution relationship of real data.

[0029] The enhancement model consists of a generator network G(z; θ (G) ) and a discriminator network D(x; θ (D) ), where θ(G) represents the weight parameter of the generator network and θ(D) represents the weight parameter of the discriminator network. The basic structure of the CGAN model that combines category information is as follows: Figure 2 shown.

[0030] Compared with the traditional GAN ​​model, the generator in this embodiment forms a new implicit representation by concatenating the noise vector z with the category condition c; and in the discriminator, the actual fault sample x or the artificial sample G(z) is input together with the category condition c for discrimination.

[0031] Therefore, the training objective function of CGAN is as follows:

[0032] (1) In order to solve the training stability problem of the CGAN model, this embodiment uses the Wasserstein distance to measure the difference between the actual distribution Pr and the generated distribution Pg. As shown in formula (2), the smaller the Wasserstein distance, the higher the similarity between the two distributions.

[0033] (2) in, Indicates and The joint probability distribution of the marginal distribution A collection of; express The expected lower bound.

[0034] Since the infimum of the Wasserstein distance cannot be solved directly, its Kantorovich-Rubinstein dual form is used to perform an equivalent transformation. On this basis, the category condition c is further introduced to ensure that the variation range of the network parameters is constrained within a certain range during training and that the gradient optimization is normal.

[0035] At the same time, in order to correct the loss function of the discriminator network so that it satisfies the 1-Lipschitz condition, a gradient penalty is added to the original loss function. Therefore, in the conditional Wasserstein generative adversarial network model based on gradient penalty optimization, the loss functions of the generator and the discriminator are shown in equations (3) and (4), respectively.

[0036] (3) (4) In the formula, z represents noise samples, x represents fault samples, c represents category conditions, G(z) represents artificial samples, represents the actual distribution, represents the generated distribution, λ represents the weight coefficient, represents the interpolation between the actual sample and the generated sample, Indicates that the discriminator is The gradient norm on .

[0037] In this step, when achieving the goal of class balancing of the sample library, a conditional Wasserstein generative adversarial network model (CWGAN-GP) with gradient penalty optimization is established. On the one hand, this method adds class constraints to the model to better guide the generation of data of different categories; on the other hand, on the basis of using Wasserstein distance instead of JS divergence as the optimization target, the gradient penalty term is introduced, which can effectively solve the problems of gradient vanishing and mode collapse in the original GAN ​​training, and greatly improve the stability of training.

[0038] In order to reduce the impact of fewer fault samples on the fault diagnosis of intelligent fire pumps, this embodiment uses a data enhancement model of the conditional Wasserstein generative adversarial network model (CWGAN-GP) with gradient penalty optimization in the process of data processing to generate more fault data based on the original small amount of fault data features, so that the positive and negative samples are evenly distributed, thereby improving the accuracy of the fault diagnosis model.

[0039] Step S103: Fusing the enhanced fault data with the original basic data, and performing data dimension reduction and data feature extraction on the fused data by principal component analysis; Specifically, in this embodiment, the enhanced fault data is fused with the original basic data to form an enhanced feature vector, but the data volume and data dimension are large, which has a certain impact on the training speed and computing efficiency of the model. Feature selection is a data dimensionality reduction method, which is often used to process high-dimensional and complex data. By selecting meaningful and helpful features from all features, it avoids the situation where all features must be imported into the model for training, thereby improving the training speed and efficiency of the model and improving the accuracy. At the same time, the principal component analysis (PCA) is used to remove abnormal data, noise data and extract key features.

[0040] Furthermore, the fused data is normalized, and then SelectKBest is used to select the features with the highest correlation with the target variable. This method can identify the features that have the greatest impact on the target variable, ensuring that only the most relevant information is used during model training. This feature selection method not only improves the performance of the model, but also avoids the negative impact of redundant features on the model.

[0041] After feature selection is completed, the most important features and their corresponding scores are listed, which provides a more concise and efficient feature set for the subsequent classification model. Through the comprehensive application of these methods, the accuracy and efficiency of the classification task can be effectively improved. The data after PCA dimension reduction can be used to detect anomalies using principal components. The specific method is to calculate the interquartile range (IQR) of each principal component, and then determine the threshold of the outlier (1.5 times the IQR) based on the IQR. If the principal component value of a data point exceeds this threshold range, the data point is marked as an outlier. In this way, possible anomalies in the data set can be identified, which is very important for equipment fault detection and preventive maintenance.

[0042] In this step, principal component analysis (PCA) is used to remove abnormal data, noise data and extract key features. This feature selection method not only improves the performance of the model, but also avoids the negative impact of redundant features on the model. At the same time, PCA is used to combine the deviation points with the final prediction results to determine the fault points.

[0043] Step S104: judging whether the intelligent fire pump has a fault and determining the abnormal point according to the output result of the pre-constructed intelligent fire pump fault diagnosis model, wherein the intelligent fire pump fault diagnosis model is trained by the extracted data features.

[0044] Specifically, in this embodiment, an intelligent fire pump fault diagnosis model based on the Transformer structure is constructed, such as Figure 3As shown in the figure, the model is implemented by stacking multiple independent basic learning modules. Each basic learning module contains a multi-head attention mechanism and a forward network. Residual connections and layer regularization are used to improve model accuracy and generalization ability, as follows: Multi-head attention mechanism The multi-head attention mechanism can assign different learned weights to different time series segments, so that each tag in the tag embedding sequence Tseq extracts useful information from other tags, such as Figure 4 As shown in Figure 1, the mechanism uses a scaled inner product attention module to calculate the attention weight of feature embedding, which can be expressed as: (5) Where Q represents the query matrix, with dimension (n, ), where n is the sequence length, represents the dimension of the key, K represents the key matrix, and the dimension is (n, ), V represents the value matrix, with dimension (n, ),in represents the dimension of the value, Represents the scaling factor, which is used to stabilize the gradient and prevent the inner product value from being too large, causing the softmax function to saturate.

[0045] Feedforward Network The forward network contains two linear transformation layers and one nonlinear activation layer, which are used to map the features extracted by multi-head attention into more abstract features. In order to improve the convergence ability of the network, the Gaussian error linear unit (GeLU) activation function is used to replace the traditional ReLU activation function.

[0046] (6) Here, x represents the input and erf(·) represents the Gaussian error function.

[0047] Residual connections and layer regularization In the Transformer structure, in order to improve the model accuracy and generalization ability, residual connections and layer regularization are used in the output part of the multi-head attention mechanism and the forward network. The output can be expressed as: (7) Based on the constructed fault diagnosis model, the acquired basic data is input into the model, and whether the intelligent fire pump is faulty is determined according to the output results of the model, and the fault point is determined through the abnormal data points of the sensor.

[0048] In this step, in the intelligent fire pump fault diagnosis model based on the Transformer structure, a multi-head attention mechanism is used to perform weighted operations on features at different positions in multiple different subspaces, so that the feature expression ability is stronger; the GeLU activation function is used in the forward network to replace the traditional ReLU activation function, so that the activation function will be smoother when there are a large number of derivative discontinuities. In addition, GeLU has a gradient in the negative part, which avoids the dead zone of ReLU. In addition, the Dropout mechanism is also introduced into the forward network to avoid overfitting of the model; residual connection and layer regularization are used in the output part to improve the model accuracy and generalization ability. According to the output of the model results, this embodiment can not only determine whether the intelligent fire pump is faulty, but also determine the fault point through the abnormal data point of the sensor.

[0049] In the process of intelligent fire pump fault diagnosis, the PCA method is not only used to effectively enhance the model's ability to eliminate abnormal data and extract features, but also combined with the fault diagnosis model results of the Transformer structure to further determine whether the intelligent fire pump has a fault and the fault point.

[0050] The intelligent fire pump fault diagnosis method based on BP neural network described in this embodiment has a higher accuracy rate in identifying the fault category of the intelligent fire pump, effectively improving the accuracy of the intelligent fire pump fault diagnosis, being able to timely discover and repair potential faults, ensuring that the fire protection system can be quickly started and effectively operated in an emergency, and also improving the maintenance efficiency and reliability of the fire pump, extending the service life of the equipment, and reducing the cost of maintenance and replacement; thereby improving the level of public safety and protecting people’s lives and property.

[0051] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0052] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. An intelligent fire pump fault diagnosis method based on BP neural network, characterized in that: include: Acquire original basic data of the intelligent fire pump in normal state and fault state, and establish a fault diagnosis database, wherein the basic data at least includes environmental parameters, power supply parameters and fire pump sensor data; The acquired original fault data is enhanced by constructing a conditional Wasserstein generative adversarial network model with gradient penalty optimization to obtain enhanced fault data. The enhanced fault data is fused with the original basic data, and the fused data is subjected to data dimension reduction and data feature extraction through principal component analysis. According to the output result of the pre-constructed intelligent fire pump fault diagnosis model, it is judged whether the intelligent fire pump has a fault and the abnormal point is determined, wherein the intelligent fire pump fault diagnosis model is trained by the extracted data features.

2. The method according to claim 1, characterized in that: Collect intelligent fire pump fault samples and define the fault data set; It is known that the actual distribution obeyed by the fault samples, and there is a set of noise vectors obeying the Gaussian distribution. By establishing a mapping relationship model between the actual distribution and the Gaussian distribution, artificial samples that satisfy the real data distribution relationship are generated based on random noise.

3. The method according to claim 2, characterized in that: The conditional Wasserstein generative adversarial network model is composed of a generator network and a discriminator network, wherein the generator network forms a new implicit representation by splicing and combining the noise vector and the category condition; in the discriminator network, the actual fault samples or artificial samples are input together with the category condition for discrimination; The difference between the actual distribution and the generated distribution is measured by the Wasserstein distance, using its Kantorovich-Rubinstein dual form and introducing the category condition; By adding a gradient penalty term to the discriminator's loss function, the gradient norm of the discriminator is constrained.

4. The method according to claim 3, characterized in that: In the conditional Wasserstein generative adversarial network model based on gradient penalty optimization, the loss functions of the generator and the discriminator are: ； ； In the formula, z represents noise samples, x represents fault samples, c represents category conditions, G(z) represents artificial samples, represents the actual distribution, represents the generated distribution, λ represents the weight coefficient, represents the interpolation between the actual sample and the generated sample, Indicates that the discriminator is The gradient norm on .

5. The method according to claim 1, characterized in that The method of performing data dimension reduction and data feature extraction on the fused data by principal component analysis includes: The fused data is normalized and SelectKBest is used for feature selection to select the features most relevant to the target variable. The data after feature selection is reduced in dimension by principal component analysis, and the principal component is used to detect anomalies to determine the fault location.

6. The method according to claim 5, characterized in that: By calculating the interquartile range of each principal component, the threshold of the outlier is determined according to the interquartile range. If the principal component value of a data point exceeds the threshold range, the data point is marked as an outlier.

7. The method according to claim 1, characterized in that An intelligent fire pump fault diagnosis model based on a Transformer structure is constructed, wherein the Transformer structure includes: The multi-head attention mechanism is configured to use a scaled inner product attention module to calculate the attention weights of feature embeddings; The forward network consists of two linear transformation layers and one nonlinear activation layer, which is used to map the features extracted by multi-head attention into more abstract features and adopts the Gaussian error linear unit activation function; Residual connections and layer regularization are set at the output of the multi-head attention mechanism and the forward network.

Images