CNN-based pump station fault prediction method

Abstract

The invention discloses a CNN-based pump station fault prediction method, and particularly relates to the field of industrial equipment fault prediction, and the method comprises the following steps: S1, deploying a sensor in a pump station, obtaining the operation state data of the pump station in real time, removing the noise and abnormal values in the data, and carrying out the normalization processing of the data; s2, extracting features related to pump station faults by using a multi-layer convolutional network, and independently performing convolution operation on each time sequence; s3, processing the time sequence data by using a bidirectional GRU network and introducing an Attention mechanism, so that the model can automatically learn different time steps and features; s4, training the model by using historical data, and optimizing the performance of the model by adjusting the parameters and the structure of the model; according to the invention, through the advanced deep learning algorithm and multi-source data fusion, efficient prediction and early warning of the pump station fault are realized, the operation efficiency and safety of equipment are significantly improved, and the method has important application value and popularization prospect.

Description

Technical Field

[0001] The present invention relates to the field of industrial equipment fault prediction, and more specifically, to a pump station fault prediction method based on CNN. Background Art

[0002] At present, pump station fault prediction mainly relies on fixed-cycle preventive maintenance and simple signal monitoring. These methods cannot provide accurate fault warnings. Traditional methods lack in-depth analysis of time series data and cannot effectively predict potential faults in complex systems. In addition, the limitations of a single data source lead to insufficient accuracy and reliability of predictions. In order to solve these problems, the present invention proposes a pump station fault prediction method based on multivariate time series CNN+bidirectional GRU+Attention mechanism. Through multi-source data fusion and deep learning algorithms, the comprehensiveness and reliability of fault prediction are improved, and efficient fault warning is achieved. Summary of the invention

[0003] In order to overcome the above-mentioned defects of the prior art, the present invention provides a pump station fault prediction method based on CNN, S1: deploy sensors in the pump station, obtain the operation status data of the pump station in real time and remove noise and outliers in the data, and normalize the data;

[0004] S2: Use a multi-layer convolutional network to extract features related to pump station failures, and perform convolution operations on each time series separately;

[0005] S3: Use a bidirectional GRU network to process time series data and introduce an Attention mechanism to enable the model to automatically learn different time steps and features;

[0006] S4: Use historical data to train the model and optimize the performance of the model by adjusting the parameters and structure of the model;

[0007] S5: The model predicts real-time data and determines the risk of pump station failure;

[0008] S6: Update and maintain the model regularly to improve the prediction accuracy and stability of the model by adding new data, adjusting model parameters and optimizing model structure.

[0009] In a preferred embodiment: In S1, sensors are deployed at key locations of the pump station in the industrial park, including water level sensors, flow meters, pressure sensors, temperature sensors and vibration sensors, to ensure real-time acquisition of the operating status data of the pump station, and a data collection frequency of once per second is set to ensure the timeliness and continuity of the data. Then, the data is preprocessed and noise and outliers in the data are removed. The anomaly detection algorithm of the sliding window method is adopted: outlier = {x||x-μ|>kσ}, where μ is the data mean, σ is the standard deviation, and k is the outlier determination coefficient. The data is normalized to ensure that the features are compared on the same scale: Among them, x is the original data, x ′ is the normalized data, min(x) and max(x) are the minimum and maximum values ​​of the data respectively; for missing data, interpolation method is used, and the linear interpolation formula is: Among them, x i is the value at the i-th time point, t i is the time of the ith time point, x i-1 and x i+1 are the values ​​of adjacent time points.

[0010] In a preferred embodiment: In S2, a multi-layer convolutional network is used to extract features related to pump station failures, and a convolution operation is performed on each time series separately for multivariate time series data: Among them, σ is the activation function, w is the convolution kernel weight, b is the bias, and k is the convolution kernel size; the convolutional features are pooled to reduce the feature dimension and the maximum pooling operation is used: p(x) = max(x t-k+1 ,…,x t ), where k is the pooling window size; the time series are convolved separately, and the same CNN network structure is used for different time series, which consists of two convolutional layers and two pooling layers. The convolution kernel sizes are 5 and 3 respectively, and the two pooling layers reduce the original feature vector to 1 / 2 and 1 / 4 respectively.

[0011] In a preferred embodiment: in S3, a bidirectional GRU network is first used to process time series data, and the update formula of GRU is as follows:

[0012] z t =σ(W z ·[h t-1 ,x t ])

[0013] r t =σ(W r ·[h t-1 ,x t ])

[0014]

[0015] Among them, z t is the update gate, r t is the reset gate, h t is the hidden state, W z ,W r ,W h is the weight matrix, and then the Attention mechanism is introduced. The model learns the importance of different time steps and features. The weight calculation formula of Attention is:

[0016] e t =tanh(W e ·h t +b e )

[0017]

[0018] Among them, W e is the weight matrix, b e is the bias, α t is the Attention weight.

[0019] In a preferred embodiment: in S4, the historical data is divided into a training set, a validation set and a test set in a ratio of 70:20:10; the model is trained using the training set, and cross-validation and regularization techniques are used to prevent overfitting: Among them, k is the fold number, D i is the i-th fold data set, error is the error function, regularization: L2 regularization is used to prevent overfitting. The L2 regularization formula is: Among them, λ is the regularization parameter, θ j are model parameters; the model is optimized using the gradient descent method. Among them, α is the learning rate, is the gradient of the loss function. The Adam optimization algorithm is used in the training process to improve the training efficiency and prediction accuracy of the model. The formula of the Adam optimization algorithm is:

[0020]

[0021] Among them, m t is the momentum term, v t is the acceleration term, β1, β2 are momentum parameters, and ∈ is a small constant to prevent division by zero. The accuracy, precision, recall and F1 value are used to evaluate the model performance. The F1 value formula is:

[0022] In a preferred embodiment: In S5, the real-time data is input into the trained model for prediction, and the model outputs the predicted value according to the input data. Setting and alarm mechanism, set the fault alarm threshold T according to historical data and experience, and the threshold is set according to different parameters: Water level threshold: T water , Traffic threshold: T flow , Pressure threshold: T pressure , Temperature threshold: T temperature , Vibration threshold: T vibration ; Set the threshold judgment rule. When the predicted value of the parameter exceeds the set threshold, the alarm mechanism is triggered: if Then alarm, if Then alarm, if Then alarm, if By setting this rule, it is ensured that when any key parameter exceeds the set threshold, the system will trigger the alarm mechanism to remind the operation and maintenance personnel to take corresponding measures.

[0023] The technical effects and advantages of the present invention are as follows: In S6, new data is used regularly to update the model, adjust parameters and optimize the structure: first, data accumulation is performed, new operating data is collected, and a new training set is formed; secondly, the model is retrained using the new data set, and then the model parameters are adjusted according to the new training results to optimize the model structure; the performance of the model in practical applications is monitored to ensure its prediction effect; the performance monitoring indicators include accuracy, recall rate, and F1 value; according to the changes in data in practical applications, the model parameters are adaptively adjusted to ensure that the model always maintains the best performance, and the adaptive learning algorithm includes an online learning algorithm and an incremental learning algorithm; the model is regularly evaluated to ensure its effectiveness on new data; the evaluation indicators include accuracy, precision, recall rate and F1 value; in the process of model updating, the optimization algorithm Adam is used to improve the training efficiency and prediction accuracy of the model, and the formula of the Adam optimization algorithm is:

[0024]

[0025]

[0026] Among them, m t is the momentum term, which accumulates the exponential decay average of the gradient, v t is the acceleration term, which accumulates the exponentially decaying average of the square of the gradient. β1 and β2 are momentum parameters, which control the update speed of the momentum term and the acceleration term. ∈ is a small constant to prevent division by zero.

[0027] Through the above steps, the present invention realizes efficient prediction and early warning of pump station failures in industrial parks, and has significant beneficial effects. First, the multivariate time series data collected by various sensors is used to ensure the quality and consistency of the data through data preprocessing and normalization, thereby improving the training effect of the model. Secondly, a convolutional neural network (CNN) is used to extract the features of the time series data, and the trends, periodicity and mutation points in the data are captured through a time-series segmentation strategy, thereby enhancing the model's ability to understand the data features. In addition, the introduction of a bidirectional gated recurrent unit (GRU) network and an attention mechanism can automatically learn the important features in the time series data, dynamically weight them, and improve the prediction accuracy and stability of the model. Model training is performed through cross-validation, regularization and optimization algorithms (such as Adam), which effectively prevents overfitting and improves the generalization ability and training efficiency of the model. The real-time prediction and alarm mechanism, by setting reasonable thresholds and alarm rules, ensures timely warnings when the risk of failure occurs, and helps operation and maintenance personnel take measures quickly to avoid equipment damage and production interruptions. Update and maintain the model regularly, and use methods such as adaptive learning and online learning to ensure that the model can adapt to the ever-changing data environment and maintain efficient and stable performance. In short, the present invention uses advanced deep learning algorithms and multi-source data fusion to achieve efficient prediction and early warning of pump station failures, significantly improve the operating efficiency and safety of equipment, reduce the risk of sudden failures, and has important application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the system module flow of the present invention. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] S1: Deploy sensors in the pumping station to obtain the operating status data of the pumping station in real time, remove noise and outliers in the data, and normalize the data;

[0031] S2: Use a multi-layer convolutional network to extract features related to pump station failures, and perform convolution operations on each time series separately;

[0032] S3: Use a bidirectional GRU network to process time series data and introduce an Attention mechanism to enable the model to automatically learn different time steps and features;

[0033] S4: Use historical data to train the model and optimize the performance of the model by adjusting the parameters and structure of the model;

[0034] S5: The model predicts real-time data and determines the risk of pump station failure;

[0035] S6: Update and maintain the model regularly, and improve the prediction accuracy and stability of the model by adding new data, adjusting model parameters and optimizing model structure. In S1, sensors are deployed at key locations of the pump station in the industrial park, including water level sensors, flow meters, pressure sensors, temperature sensors and vibration sensors, to ensure real-time acquisition of the operating status data of the pump station. The data collection frequency is set once per second to ensure the timeliness and continuity of the data. Then, the data is preprocessed and noise and outliers in the data are removed. The anomaly detection algorithm of the sliding window method is adopted: outlier = {x||x-μ|>kσ}, where μ is the data mean, σ is the standard deviation, and k is the outlier determination coefficient. The data is normalized to ensure that the features are compared on the same scale: Among them, x is the original data, x ′ is the normalized data, min(x) and max(x) are the minimum and maximum values ​​of the data respectively; for missing data, interpolation method is used, and the linear interpolation formula is: Among them, x i is the value at the i-th time point, t i is the time of the ith time point, x i-1 and x i+1 are the values ​​at adjacent time points respectively; various types of sensors are deployed at key locations of the pump station in the industrial park, including water level sensors, flow meters, pressure sensors, temperature sensors, and vibration sensors, to ensure real-time acquisition of the operating status data of the pump station. Set a reasonable data collection frequency, such as once per second or once per minute, to ensure the timeliness and continuity of the data. After the data collection is completed, the data needs to be preprocessed. The first is data cleaning to remove noise and outliers in the data, such as extreme values ​​caused by sensor failure. This can be achieved using a sliding window method or a statistical anomaly detection algorithm, and outliers are judged by thresholds of the mean and standard deviation. Next is data normalization, which normalizes data of different dimensions to ensure that each feature is compared on the same scale, which helps to improve the stability and effect of model training. Normalization can be done by adjusting each data point to a range of 0 to 1. In addition, missing data needs to be filled, and commonly used methods include interpolation and filling the average value. Interpolation can estimate the value of the missing point by the values ​​of the previous and next data points, thereby maintaining the continuity of the data. Through these steps, the preprocessed data will have higher quality and consistency, providing a reliable data foundation for subsequent model training.

[0036] At this time, in S2, a multi-layer convolutional network is used to extract features related to pump station failures, and convolution operations are performed on each time series separately for multivariate time series data: Among them, σ is the activation function, w is the convolution kernel weight, b is the bias, and k is the convolution kernel size; the convolutional features are pooled to reduce the feature dimension and the maximum pooling operation is used: p(x) = max(x t-k+1 ,…,x t ), where k is the pooling window size; convolution is performed on the time series separately, and the same CNN network structure is used for different time series, which consists of two convolutional layers and two pooling layers. The convolution kernel sizes are 5 and 3 respectively, and the two pooling layers reduce the original feature vector to 1 / 2 and 1 / 4 respectively; in addition, in the preprocessed data, a multi-layer convolutional network is used to extract features related to pump station failures. These features may include trends, periodicity, mutation points, etc. of time series data. Convolutional neural networks (CNNs) can effectively extract local features, especially for the processing of images or one-dimensional sequence data. In this step, for multivariate time series data, each time series is convolved separately. The convolutional network consists of multiple convolutional layers and pooling layers. The convolutional layer is used to extract features, and the pooling layer is used to reduce dimensionality and reduce the number of parameters. For each time series, the same CNN network structure is used to extract features through a combination of multiple convolutional layers and pooling layers. The convolution kernel size and the configuration of the pooling layer are set according to the specific situation of the data. Different time series use the same CNN network structure, which consists of two convolutional layers and two pooling layers. The convolution kernel sizes are 5 and 3 respectively. The two pooling layers reduce the original feature vectors to 1 / 2 and 1 / 4 respectively. Through the strategy of time series segmentation, each time series is convoluted separately, so as to extract more representative features from the time series data and provide richer information for subsequent model training.

[0037] Next, in S3, we first use a bidirectional GRU network to process time series data. The update formula of GRU is as follows:

[0038] z t =σ(W z ·[h t-1 ,x t ])

[0039] r t =σ(W r ·[h t-1 ,x t ])

[0040]

[0041] Among them, zt is the update gate, r t is the reset gate, h t is the hidden state, W z ,W r ,W h is the weight matrix, and then the Attention mechanism is introduced. The model learns the importance of different time steps and features. The weight calculation formula of Attention is:

[0042] e t =tanh(W e ·h t +b e )

[0043]

[0044] Among them, W e is the weight matrix, b e is the bias, α t is the attention weight; after feature extraction, in order to further improve the accuracy and reliability of pump station fault prediction, a bidirectional gated recurrent unit (GRU) network is used to process time series data, and an attention mechanism is introduced. The bidirectional GRU network is an improved recurrent neural network (RNN) that can simultaneously consider the past and future information of time series data. Traditional RNN models are prone to gradient vanishing or explosion problems when dealing with long-term dependencies, while GRU effectively alleviates this problem through a gating mechanism. The GRU network consists of an update gate and a reset gate, which can control the flow of information, allowing the model to selectively remember and forget information, thereby better capturing long-term dependencies in time series.

[0045] In the present invention, a bidirectional GRU network is used to process feature sequences extracted from a convolutional neural network (CNN). By inputting the feature sequence into the bidirectional GRU network, the model can capture the time dependency in the feature sequence. This processing method enables the model to understand the information in the time series data more comprehensively and improve the accuracy of fault prediction. The forward processing of the bidirectional GRU network can capture the influence of past information, and the reverse processing can capture the influence of future information, thereby taking into account both past and future information at the same time step, enhancing the prediction ability of the model.

[0046] In the encoder part, a bidirectional GRU is used to encode the feature sequence extracted from CNN. The task of the encoder is to convert the input feature sequence into a fixed-length context vector, which contains important information in the input sequence. Through the processing of the bidirectional GRU, the encoder can generate a more comprehensive and accurate context vector, providing effective information support for the subsequent decoder part. In the decoder part, the attention mechanism is introduced. The attention mechanism can dynamically adjust the importance of each feature by calculating the attention weight of the input feature, thereby improving the prediction ability of the model.

[0047] The Attention mechanism can automatically learn and focus on the most critical time steps and features for fault prediction by calculating the attention weights for each time step. Specifically, the Attention mechanism first weights the context vector output by the encoder through a weight matrix, and then inputs the weighted context vector into the decoder. Through this dynamic weighting method, the Attention mechanism enables the model to pay more attention to input elements related to the fault and ignore irrelevant information. The introduction of the Attention mechanism avoids the information loss problem that may occur in traditional models. Especially when processing long time series data, the Attention mechanism can significantly improve the prediction accuracy and stability of the model.

[0048] In the decoder part, the Attention mechanism is used to perform weighted averaging on the context vectors generated from the bidirectional GRU encoder. By calculating the attention weights, the model can dynamically adjust the importance of each feature, thereby improving the accuracy of fault prediction. The Attention mechanism can dynamically adjust the weights of features by learning the importance of different features, so that the model can more accurately capture the key features related to pump station failures. This processing method can significantly improve the model's predictive ability, allowing it to maintain a high level of prediction accuracy when facing complex and changing time series data.

[0049] In addition, the Attention mechanism is also interpretable to a certain extent. By observing the distribution of Attention weights, we can understand which time steps and features the model pays most attention to when making predictions. This is of great significance for model debugging and optimization. By analyzing the changes in Attention weights, operation and maintenance personnel can better understand the causes and processes of failures, and thus take more effective prevention and maintenance measures.

[0050] In summary, by introducing the bidirectional GRU network and the Attention mechanism after feature extraction, the pump station fault prediction model of the present invention can more comprehensively capture the information in the time series data, thereby improving the accuracy and reliability of fault prediction. The bidirectional GRU network captures the forward and backward dependencies of the data through forward and reverse processing, so that the model has better performance when processing long-term dependencies. The Attention mechanism improves the model's attention to key features through dynamic weighting, avoids the problem of information loss, and significantly improves the model's prediction ability. Through this combination, the present invention can more accurately predict pump station failures and provide strong support for the operation and maintenance of pump stations in industrial parks.

[0051] Then, in S4, the historical data is divided into training set, validation set and test set with a ratio of 70:20:10; the model is trained using the training set, and cross-validation and regularization techniques are used to prevent overfitting: Among them, k is the fold number, D i is the i-th fold data set, error is the error function, regularization: L2 regularization is used to prevent overfitting. The L2 regularization formula is: Among them, λ is the regularization parameter, θ j are model parameters; the model is optimized using the gradient descent method. Among them, α is the learning rate, is the gradient of the loss function. The Adam optimization algorithm is used in the training process to improve the training efficiency and prediction accuracy of the model. The formula of the Adam optimization algorithm is:

[0052]

[0053] Among them, m t is the momentum term, v t is the acceleration term, β1, β2 are momentum parameters, and ∈ is a small constant to prevent division by zero. The accuracy, precision, recall and F1 value are used to evaluate the model performance. The F1 value formula is: Use historical data to train the model, and optimize the model's performance by adjusting the model's parameters and structure. During the training process, cross-validation and regularization techniques are used to prevent overfitting and improve the generalization ability of the model. Cross-validation evaluates the performance of the model on different data sets by dividing the data set into multiple subsets for training and validation, respectively, to ensure the robustness and stability of the model. Regularization prevents the model from being overly dependent on the training data by adding a penalty term to the loss function, thereby improving the performance of the model on new data. During the model training process, the gradient descent method is used to optimize the model and update the model's weights. In order to improve training efficiency and prediction accuracy, optimization algorithms such as Adam and RMSprop are used. These optimization algorithms can converge to the optimal solution faster by adaptively adjusting the learning rate and momentum, thereby improving the training efficiency of the model. The Adam optimization algorithm combines the advantages of the momentum method and RMSprop, and can quickly find the optimal solution in a complex high-dimensional parameter space. During the model training process, it is also necessary to set a variety of evaluation indicators such as accuracy, precision, recall, and F1 value to evaluate the model performance. The F1 value is particularly suitable for evaluating data sets with imbalanced classifications. It provides a balanced evaluation standard by comprehensively considering precision and recall. Through these technical means, the model can better adapt to data in actual applications and improve the accuracy and reliability of fault prediction.

[0054] It should be noted that in S6, new data is used regularly to update the model, adjust parameters and optimize the structure: first, data accumulation is carried out, new operating data is collected to form a new training set; second, the model is retrained using the new data set, and then the model parameters are adjusted according to the new training results to optimize the model structure; the performance of the model in actual applications is monitored to ensure its prediction effect; performance monitoring indicators include accuracy, recall rate, and F1 value; according to the changes in data in actual applications, the model parameters are adaptively adjusted to ensure that the model always maintains the best performance. The adaptive learning algorithm includes online learning algorithm and incremental learning algorithm; the model is evaluated regularly to ensure its effectiveness on new data; the evaluation indicators include accuracy, precision, recall rate and F1 value; in the process of model updating, the optimization algorithm Adam is used to improve the training efficiency and prediction accuracy of the model. The formula of Adam optimization algorithm is:

[0055]

[0056] Among them, m t is the momentum term, which accumulates the exponential decay average of the gradient, v tis the acceleration term, which accumulates the exponential decay average of the square of the gradient. β1 and β2 are momentum parameters, which control the update speed of the momentum term and the acceleration term. ∈ is a small constant to prevent division by zero. The trained model is applied to real-time data to predict whether the pump station has a risk of failure. First, the real-time data is input into the model for prediction, and the predicted value is calculated according to the model output. In order to ensure the effectiveness of the prediction results, it is necessary to set the threshold of the fault alarm. The threshold can be set according to historical data and experience, such as the threshold of parameters such as water level, flow, pressure, temperature, and vibration. When the predicted value of one or more parameters exceeds the set threshold, the alarm mechanism is triggered. For example, if the predicted value of the water level exceeds the set water level threshold, an alarm signal is issued; if the predicted value of the flow exceeds the set flow threshold, an alarm signal is issued; if the predicted value of the pressure exceeds the set pressure threshold, an alarm signal is issued; if the predicted value of the temperature exceeds the set temperature threshold, an alarm signal is issued; if the predicted value of the vibration exceeds the set vibration threshold, an alarm signal is issued. By setting these if...then... rules, it can be ensured that when any key parameter exceeds the set threshold, the system will trigger the alarm mechanism to remind the operation and maintenance personnel to take corresponding measures. The alarm mechanism can include various forms such as sound alarms, visual alarms and software prompts to ensure that the operation and maintenance personnel can detect and deal with potential faults in a timely manner. In addition, the prediction results and fault trends are displayed through visualization tools to help operation and maintenance personnel understand and respond to pump station failures. Visualization tools can use time series diagrams, scatter plots and heat maps to intuitively display the change trend of each parameter over time, the distribution of the overall distribution and hot spots of the fault prediction results, thereby improving the efficiency and accuracy of fault management.

[0057] The above embodiments can be implemented in whole or in part by software, hardware, firmware and any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions and computer programs. When the computer instructions and computer programs are loaded and executed on a computer, the processes and functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server and data center to another website, computer, server and data center by wired means (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessed by a computer or a data storage device such as a server or data center that includes one or more available media sets. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0058] It should be understood that in various embodiments of the present application, the order of the sequence numbers of the above processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0059] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here;

[0060] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment;

[0061] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit;

[0062] If the functions are implemented in the form of software functional units and sold and used as independent products, they can be stored in a computer-readable storage medium; based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application; and the aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk;

[0063] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes and substitutions within the technical scope disclosed in the present application, which should be covered by the protection scope of the present application.

Claims

1. A pump station fault prediction method based on CNN, characterized in that: The fault prediction method steps are as follows: S1: Deploy sensors in the pumping station to obtain real-time operating status data of the pumping station, remove noise and outliers from the data, and normalize the data; S2: Use a multi-layer convolutional network to extract features related to pump station failures, and perform convolution operations on each time series separately; S3: Use a bidirectional GRU network to process time series data and introduce an attention mechanism to enable the model to automatically learn different time steps and features; S4: Use historical data to train the model and optimize the model's performance by adjusting the model's parameters and structure; S5: The model predicts real-time data and determines the risk of pump station failure; S6: Update and maintain the model regularly to improve the prediction accuracy and stability of the model by adding new data, adjusting model parameters and optimizing model structure.

2. A pump station fault prediction method based on CNN according to claim 1, characterized in that: In S1, sensors, including water level sensors, flow meters, pressure sensors, temperature sensors, and vibration sensors, are deployed at key locations in the industrial park pump station to ensure real-time acquisition of the pump station's operating status data. A data collection frequency of once per second is set to ensure the timeliness and continuity of the data. The data is then preprocessed to remove noise and outliers. A sliding window anomaly detection algorithm is used: outlier = *x||x-μ|>kσ+, where μ is the data mean, σ is the standard deviation, and k is the outlier determination coefficient. The data is normalized to ensure that features are compared on the same scale: Among them, x is the original data, x ′ is the normalized data, min(x) and max(x) are the minimum and maximum values ​​of the data respectively; for missing data, interpolation method is used, and the linear interpolation formula is: Among them, x i is the value at the i-th time point, t i is the time of the i-th time point, x i-1 and x i+1 are the values ​​of adjacent time points respectively.

3. The pump station fault prediction method based on CNN according to claim 1, characterized in that: In S2, a multi-layer convolutional network is used to extract features related to pump station failures. For multivariate time series data, each time series is convolved separately: Among them, σ is the activation function, w is the convolution kernel weight, b is the bias, and k is the convolution kernel size; the convolution features are pooled to reduce the feature dimension and the maximum pooling operation is used: p(x) = max(x t-k+1 ,…,x t ), where k is the pooling window size; convolution is performed on the time series separately, and the same CNN network structure is used for different time series, which consists of two convolutional layers and two pooling layers. The convolution kernel sizes are 5 and 3 respectively. The two pooling layers reduce the original feature vector to 1 / 2 and 1 / 4 respectively.

4. The pump station fault prediction method based on CNN according to claim 1, characterized in that: In S3, we first use a bidirectional GRU network to process time series data. The update formula of GRU is as follows: z t =σ(W z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) Among them, z t is the update gate, r t is the reset gate, h t is the hidden state, W z ,W r ,W h is the weight matrix, and then the Attention mechanism is introduced. The model learns the importance of different time steps and features. The weight calculation formula of Attention is: e t =tanh(W e ·h t +b e ) Among them, W e is the weight matrix, b e is the bias, α t is the Attention weight.

5. The pump station fault prediction method based on CNN according to claim 1, characterized in that: In S4, the historical data is divided into training set, validation set and test set with a ratio of 70:20:10; the model is trained using the training set, and cross-validation and regularization techniques are used to prevent overfitting: Among them, k is the fold, D i is the i-th fold data set, error is the error function, regularization: L2 regularization is used to prevent overfitting. The L2 regularization formula is: Among them, λ is the regularization parameter, θ j are model parameters; the model is optimized using the gradient descent method. Among them, α is the learning rate, is the gradient of the loss function. The Adam optimization algorithm is used in the training process to improve the training efficiency and prediction accuracy of the model. The formula of the Adam optimization algorithm is: Among them, m t is the momentum term, v t is the acceleration term, β1 and β2 are momentum parameters, and ∈ is a small constant to prevent division by zero. The accuracy, precision, recall and F1 value are used to evaluate the model performance. The F1 value formula is:

6. The pump station fault prediction method based on CNN according to claim 1, characterized in that: In S5, real-time data is input into the trained model for prediction, and the model outputs the predicted value based on the input data. Setting and alarm mechanism, set the fault alarm threshold T according to historical data and experience, and the threshold is set according to different parameters: Water level threshold: T water , Traffic threshold: T flow , Pressure threshold: T pressure , Temperature threshold: T temperature , vibration threshold: T vibration ; Set the threshold judgment rule. When the predicted value of the parameter exceeds the set threshold, the alarm mechanism is triggered: if Then alarm, if Then alarm, if Then alarm, if By setting this rule, when any key parameter exceeds the set threshold, the system will trigger the alarm mechanism to remind the operation and maintenance personnel to take corresponding measures.

7. The pump station fault prediction method based on CNN according to claim 1, characterized in that: In S6, new data is used to update the model regularly, adjust parameters, and optimize the structure: first, data accumulation is carried out, new operating data is collected, and a new training set is formed; second, the model is retrained using the new data set, and then the model parameters are adjusted and the model structure is optimized based on the new training results; the performance of the model in actual applications is monitored to ensure its prediction effect; performance monitoring indicators include accuracy, recall rate, and F1 value; according to the changes in data in actual applications, the model parameters are adaptively adjusted to ensure that the model always maintains the best performance. The adaptive learning algorithm includes online learning algorithm and incremental learning algorithm; the model is regularly evaluated to ensure its effectiveness on new data; the evaluation indicators include accuracy, precision, recall rate, and F1 value; during the model update process, the optimization algorithm Adam is used to improve the training efficiency and prediction accuracy of the model. The Adam optimization algorithm formula is: Among them, m t is the momentum term, which accumulates the exponential decay average of the gradient, v t is the acceleration term, which accumulates the exponentially decaying average of the square of the gradient. β1 and β2 are momentum parameters, which control the update speed of the momentum term and the acceleration term. ∈ is a small constant to prevent division by zero.

Images