Sewage treatment process abnormal working condition identification method and system based on deep neural network

Abstract

The invention discloses a sewage treatment process abnormal working condition identification method and system based on a deep neural network, and the method comprises the steps: obtaining and preprocessing multi-source heterogeneous data in a sewage treatment process, and obtaining a standardized data set; performing dimension reduction processing on the standardized data set to obtain a low-dimensional feature vector; constructing a feature extraction module based on the low-dimensional feature vectors, extracting spatial features, and obtaining depth feature representation; a time sequence analysis module is introduced, a dynamic change rule of time sequence data is captured, and time sequence feature representation is obtained; and fusing the spatial feature representation and the time sequence feature representation, and outputting probability distribution of abnormal working conditions through a classifier to obtain an anomaly recognition result. The method not only can accurately identify abnormal conditions in the sewage treatment process, but also can adapt to complex and changeable working conditions and environments, and provides powerful support for safe and stable operation of a sewage treatment plant.

Description

Technical Field

[0001] The present invention belongs to the technical field of sewage treatment, and in particular to a method and system for identifying abnormal operating conditions in a sewage treatment process based on a deep neural network. Background Art

[0002] In the sewage treatment process, the identification of abnormal conditions is a key link to ensure the stable operation of the sewage treatment system. Traditional methods mainly rely on manually set thresholds or rules for anomaly detection. These methods determine whether abnormal conditions occur by monitoring changes in key parameters, such as influent flow rate, pollutant concentration, etc. In addition, some data-driven models, such as artificial neural networks, have also been applied to the modeling and prediction of sewage treatment processes. For example, feedforward neural networks are used to perform soft measurements of parameters such as effluent chemical oxygen demand (COD).

[0003] However, existing technologies have many shortcomings in identifying abnormal conditions in sewage treatment. First, the sewage treatment process involves multiple links and parameters, and the heterogeneity and high dimensionality of the data make it difficult to extract features. Secondly, the occurrence of abnormal conditions is random and sudden, and traditional methods are difficult to quickly and accurately identify anomalies in massive data. In addition, the sewage treatment environment is complex and changeable, which places higher demands on the robustness and generalization ability of the model. At the same time, existing neural network-based methods have problems such as high model complexity and long training time, which makes it difficult to strike a balance between recognition accuracy and computational efficiency in practical applications. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method and system for identifying abnormal operating conditions in a sewage treatment process based on a deep neural network to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for identifying abnormal conditions in a sewage treatment process based on a deep neural network, comprising:

[0006] Obtain multi-source heterogeneous data from the sewage treatment process and pre-process them to obtain a standardized data set;

[0007] Perform dimensionality reduction on the standardized data set to obtain a low-dimensional feature vector;

[0008] Construct a feature extraction module based on low-dimensional feature vectors to extract spatial features and obtain deep feature representation;

[0009] Introduce the time series analysis module to capture the dynamic change rules of time series data and obtain the time series feature representation;

[0010] The spatial feature representation is fused with the temporal feature representation, and the probability distribution of abnormal working conditions is output through the classifier to obtain the abnormality recognition result.

[0011] Preferably, the process of obtaining a standardized data set includes:

[0012] Performing data cleaning on the multi-source heterogeneous data to remove noise and outliers;

[0013] The normalization method is used to map data of different dimensions to the same scale to obtain a standardized data set.

[0014] Preferably, the process of obtaining the low-dimensional feature vector includes:

[0015] Standardize the original high-dimensional data of the standardized data set;

[0016] The principal component analysis algorithm is used to reduce the dimension of the standardized high-dimensional data set;

[0017] The dimension threshold is used to determine whether the reduced-dimensionality data set meets the requirements. If not, the parameters are adjusted and the dimension is reduced again.

[0018] Preferably, the process of obtaining the deep feature representation includes:

[0019] Build the convolutional neural network architecture and set up the input layer;

[0020] Configure multiple convolution kernels in the convolution layer and connect the pooling layer to perform feature map dimensionality reduction processing;

[0021] Repeatedly stack multiple convolutional layers and pooling layers to form a feature extraction structure;

[0022] Add a fully connected layer after the last convolutional layer to convert spatial features into deep feature representation;

[0023] If the feature dimension does not meet the requirements, adjust the number of convolution kernels or the number of nodes in the fully connected layer;

[0024] The network parameters are optimized via a back-propagation algorithm to capture spatial feature patterns.

[0025] Preferably, the process of obtaining the time series feature representation includes:

[0026] Based on the time series data in the standardized data set, a deep feature representation model is constructed, and the deep features of the data are extracted through a deep learning algorithm;

[0027] On the basis of deep feature representation, the LSTM network module is introduced, and the memory and forget gate mechanism of LSTM is used to capture the dynamic change law of time series data;

[0028] The time series data is modeled through the LSTM network module to obtain the time series feature representation;

[0029] Adopt anomaly detection algorithm to judge whether there are abnormal emergencies in the data according to the time series characteristics;

[0030] If an abnormal emergency is detected, the early warning mechanism is triggered;

[0031] Continuously iterate and optimize the parameters of the deep feature representation model and LSTM network module, and optimize the model parameters through incremental learning;

[0032] The optimized model is applied to actual scenarios to perform real-time monitoring and anomaly detection on new data.

[0033] Preferably, the process of judging whether the data has an abnormal sudden situation according to the time series feature representation includes:

[0034] Obtaining time points in a continuous time series, dividing the time points into multiple data segments according to a preset time window, and obtaining multiple sets of data segments arranged in time order;

[0035] Using a preset sampling rate, the time series points in each data segment in the data segment set are discretized to obtain multiple time series point values ​​after discretization of each data segment. According to the time series point values ​​in each data segment, the volatility of each data segment is calculated;

[0036] According to the time point value in each data segment, the periodicity of each data segment is calculated to obtain the periodicity value of each data segment, and according to the periodicity value of each data segment, the time difference between adjacent data segments is calculated;

[0037] Obtain the value of the time series point in each data segment, calculate the change rate of the value of the time series point in the adjacent data segment, and if the change rate is greater than the preset change rate threshold, calculate the increase of the value of the time series point in the adjacent data segment;

[0038] According to the volatility and periodicity of each data segment, the outliers of each data segment are calculated. If the number of outliers is greater than the preset threshold, the data segment is judged to be bursty.

[0039] The correlation between adjacent data segments is determined based on the periodicity value of each data segment and the time difference between adjacent data segments. If the correlation between adjacent data segments is lower than a preset correlation threshold, the adjacent data segment is determined to be an abnormal data segment.

[0040] Preferably, the process of obtaining anomaly identification results includes:

[0041] The spatial feature representation vector and the temporal feature representation vector are fused by concatenation or weighted average to obtain a fused feature representation vector;

[0042] The fused feature representation vector is input into the fully connected layer to construct an abnormal operating condition classifier, and the probability distribution of abnormal operating conditions is output through the softmax activation function;

[0043] According to the probability distribution of abnormal working conditions, the threshold is set to determine whether the current equipment operating condition is abnormal and obtain the abnormal identification result.

[0044] Preferably, the process of outputting the probability distribution of abnormal operating conditions through the softmax activation function includes:

[0045] According to the fused feature vector, a fully connected layer structure of the abnormal operating condition classifier is constructed, and the appropriate number of neurons and layers are set;

[0046] The softmax activation function is used as the output layer of the fully connected layer to map the feature vector to the probability distribution space of abnormal working conditions;

[0047] By training sample data, the back propagation algorithm is used to optimize the weight parameters of the fully connected layer so that the classifier can identify abnormal working conditions;

[0048] For different types of abnormal working conditions, different loss weights are set to enhance the sensitivity of the classifier to key abnormalities;

[0049] In the inference stage, the feature vector of the data to be tested is input into the trained fully connected layer, and the probability distribution of abnormal conditions is calculated through the softmax function.

[0050] In a second aspect, the present invention provides a system for identifying abnormal conditions in a sewage treatment process based on a deep neural network, comprising:

[0051] The data preprocessing module is used to obtain and preprocess multi-source heterogeneous data in the sewage treatment process to obtain a standardized data set;

[0052] A dimension reduction processing module is used to perform dimension reduction processing on the standardized data set to obtain a low-dimensional feature vector;

[0053] A spatial feature extraction module is used to construct a feature extraction module based on a low-dimensional feature vector, extract spatial features, and obtain a deep feature representation;

[0054] The time series feature extraction module is used to introduce the time series analysis module to capture the dynamic change rules of time series data and obtain the time series feature representation;

[0055] The feature fusion and classification module is used to fuse the spatial feature representation with the temporal feature representation, and output the probability distribution of abnormal working conditions through the classifier to obtain the abnormal recognition result.

[0056] Compared with the prior art, the present invention has the following advantages and technical effects:

[0057] The present invention provides a method for identifying abnormal operating conditions in a sewage treatment process based on a deep neural network. First, the multi-source heterogeneous data in the sewage treatment process is processed and analyzed, and a deep learning model integrating a feature extraction module and a time series analysis module is constructed through steps such as data standardization, dimensionality reduction and feature extraction. The model can simultaneously capture the spatial characteristics and temporal characteristics of the data, effectively improving the ability to identify abnormal operating conditions. The present invention can not only accurately identify abnormal conditions in the sewage treatment process, but also adapt to complex and changeable working conditions, providing strong support for the safe and stable operation of sewage treatment plants. The technical effect of the present invention is significant, and can greatly improve the efficiency and accuracy of abnormal detection in sewage treatment systems, which is of great significance to ensuring water quality safety and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] 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:

[0059] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0060] Figure 2 Schematic diagram of a system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0062] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0063] Embodiment 1

[0064] like Figure 1 As shown, this embodiment provides a method for identifying abnormal conditions in a sewage treatment process based on a deep neural network, including:

[0065] S101. Obtain multi-source heterogeneous data in a sewage treatment process, and obtain a standardized data set through data cleaning and normalization, wherein the standardized data set eliminates data dimension differences.

[0066] Obtain multi-source heterogeneous data in the sewage treatment process, and use corresponding data cleaning methods to pre-process data from different sources and in different formats to remove noise and outliers and ensure data quality. According to the needs of sewage treatment business, determine the target range of data normalization, and use methods such as minimum-maximum normalization or z-score normalization to map data of different dimensions to the same scale to eliminate the impact of dimensional differences. After data cleaning and normalization, a standardized sewage treatment data set is obtained with a unified data format and consistent dimensions, which is convenient for subsequent data analysis and modeling applications.

[0067] Specifically, multi-source heterogeneous data in the sewage treatment process include water quality parameters, equipment operating status, environmental factors, etc. Data cleaning is a key step to ensure data quality, such as eliminating or correcting data with abnormal pH values, and marking COD values ​​that exceed the normal range as abnormal. Normalization can eliminate the dimensional differences between different indicators, such as mapping the BOD5 concentration from 0-500 mg / L to the range of 0-1, which facilitates comparison and analysis between different indicators. Cluster analysis can identify different types of sewage, such as classifying sewage with high ammonia nitrogen and low COD into one category, and sewage with high organic matter and low nitrogen and phosphorus into another category, which helps to optimize the treatment process in a targeted manner.

[0068] S102. For the standardized data set, a principal component analysis method is used to perform dimensionality reduction processing on high-dimensional data to obtain a low-dimensional feature vector, and the low-dimensional feature vector reduces the complexity of subsequent model calculations.

[0069] According to the preset data standard, the acquired original high-dimensional data is standardized to obtain a standardized high-dimensional data set; for the standardized high-dimensional data set, the principal component analysis algorithm is used to perform feature dimensionality reduction, and the main feature information of the data is extracted to obtain the reduced-dimensional data set; according to the preset dimensionality threshold, it is judged whether the dimension of the reduced-dimensional data set meets the requirements, and if it meets the requirements, the reduced-dimensional data set is output as a low-dimensional feature vector; if it does not meet the requirements, the parameters of the principal component analysis algorithm are adjusted, and the feature dimensionality reduction process is performed again until the dimension meets the requirements.

[0070] Specifically, data standardization is a key step to ensure data quality and consistency. Taking the water quality monitoring data of a sewage treatment plant as an example, the original data may include multiple indicators such as pH, dissolved oxygen, and chemical oxygen demand, and the dimensions and ranges of these indicators vary greatly. Through standardization, all indicators can be mapped to a unified scale, such as between 0 and 1, to eliminate the impact of dimensions and facilitate subsequent analysis. Principal component analysis (PCA) is a commonly used feature dimensionality reduction method. In sewage treatment data, there may be multiple water quality indicators with high correlation. PCA can extract the main change information of these indicators and compress the data that may originally have dozens of dimensions into a few principal components. For example, it may be found that the first three principal components can explain more than 80% of the variance of the original data, greatly reducing the data dimension. The setting of the dimension threshold requires a trade-off between information retention and computational efficiency. In the analysis of sewage treatment data, a cumulative variance contribution rate of 95% may be set as the threshold. If the dimension after dimensionality reduction is still too high, the feature space can be further compressed by adjusting the PCA parameters, such as increasing the regularization strength.

[0071] S103. Construct a feature extraction module based on a convolutional neural network according to the low-dimensional feature vector, extract the spatial features in the standardized data set through multi-layer convolution and pooling operations, and obtain a deep feature representation.

[0072] Construct a convolutional neural network architecture and set the input layer to receive a standardized data set. Configure multiple convolution kernels in the convolution layer, each with a preset size and step size parameter. Connect a pooling layer after the convolution layer and use the maximum pooling method to reduce the dimension of the feature map. Repeatedly stack the convolution layer and the pooling layer to form a multi-layer feature extraction structure. Add a fully connected layer after the last convolution layer to convert the extracted spatial features into a fixed-dimensional deep feature representation. If the feature dimension does not meet the preset requirements, adjust the number of convolution kernels or the number of nodes in the fully connected layer. Optimize the network parameters through the back-propagation algorithm so that the feature extraction module can accurately capture the spatial feature patterns in the data.

[0073] Specifically, convolutional neural networks are a powerful deep learning architecture that is particularly suitable for processing data with spatial structure, such as images and videos. When building a network, the input layer is first set to receive a standardized dataset, which helps speed up the training process and improve model performance. For example, for an image classification task, the input may be a 224x224x3 RGB image. The convolution layer is the core of the network, which extracts features by configuring multiple convolution kernels. Each convolution kernel has a specific size and stride to capture features at different scales. For example, the first layer of convolution may use 32 3x3 convolution kernels with a stride of 1, which can detect low-level features such as edges and textures. As the network goes deeper, the number of convolution kernels may be increased and the size may be reduced, such as 64 3x3 convolution kernels, to capture more complex features. The pooling layer follows the convolution layer and is mainly used to reduce the spatial dimension of the feature map. Max pooling is a common method that retains the most significant features of each region. For example, a 2x2 max pooling operation will halve the height and width of the feature map, effectively reducing the number of parameters while maintaining the main feature information. By repeatedly stacking convolutional layers and pooling layers, the network forms a multi-layer feature extraction structure. This hierarchical structure enables the network to gradually learn feature representations from simple to complex. For example, a typical VGG style network may contain 5 convolutional blocks, each containing 2-3 convolutional layers and a pooling layer. After the last convolutional layer, a fully connected layer is added to convert the spatial features into a deep feature representation of fixed dimensions. This step is crucial for subsequent classification or regression tasks. For example, two fully connected layers can be used, with 1024 nodes in the first layer and 512 nodes in the second layer, and finally output a fixed-dimensional feature vector. If the extracted feature dimension does not meet the preset requirements, it can be modified by adjusting the number of convolution kernels or the number of nodes in the fully connected layer. For example, if the goal is to obtain a 256-dimensional feature vector, but the current output is 512-dimensional, the number of nodes can be reduced from 512 to 256 in the last fully connected layer. The optimization of network parameters is achieved through the back-propagation algorithm. This process involves calculating the gradient of the loss function for each parameter and updating the parameters using an optimizer such as Adam or SGD. Through multiple rounds of training, the network gradually adjusts its weights to more accurately capture the spatial feature patterns in the data. For example, in an image classification task, it may take tens or hundreds of epochs for the network to converge to a good feature extractor. The design and optimization process of this convolutional neural network architecture can not only effectively extract the spatial features of the data, but also greatly reduce the computational complexity through dimensionality reduction and parameter sharing. This enables the model to handle high-dimensional input data while maintaining high computational efficiency and generalization capabilities. Through carefully designed network structures and training strategies, excellent performance can be achieved in various computer vision tasks.

[0074] S104. Based on the deep feature representation, a long short-term memory network module is introduced to capture the dynamic change law of the time series data in the standardized data set to obtain a time series feature representation. The time series feature representation enhances the model's ability to recognize abnormal bursts.

[0075] Based on the time series data in the standardized data set, a deep feature representation model is constructed, and the deep features of the data are extracted through a deep learning algorithm. On the basis of deep feature representation, the LSTM network module is introduced, and the memory and forget gate mechanisms of LSTM are used to capture the dynamic changes of time series data. The time series data is modeled through the LSTM network module to obtain the time series feature representation, which enhances the model's ability to represent time series data. The anomaly detection algorithm is used to determine whether the data has abnormal emergencies based on the time series feature representation, thereby improving the model's anomaly recognition ability. If an abnormal emergencies are detected, the early warning mechanism is triggered, and relevant personnel are notified to handle them through visualization or alarms. The parameters of the deep feature representation model and the LSTM network module are continuously iterated and optimized, and the feature extraction and anomaly recognition performance of the model are improved through incremental learning. The optimized model is applied to actual business scenarios, and real-time monitoring and anomaly detection of new time series data are performed to timely discover and handle potential anomalies and ensure the stable operation of the business.

[0076] Specifically, the deep feature representation model is an effective way to extract complex features from standardized data sets through deep learning algorithms. The introduction of LSTM network modules can effectively handle long-term dependencies in time series data. Based on the time series feature representation, the anomaly detection algorithm can identify data points that deviate from the normal pattern. Through this comprehensive deep learning and anomaly detection framework, enterprises can better understand and predict changes in time series data, detect potential problems in advance, and take corresponding measures, thereby improving the efficiency and stability of business operations.

[0077] Anomaly detection algorithms are used to determine whether the data has abnormal emergencies based on time series characteristics, thereby improving the model's anomaly recognition capabilities.

[0078] Obtain the time points in the continuous time series, divide the time points into multiple data segments according to the preset time window, and obtain multiple sets of data segments arranged in time order. Use the preset sampling rate to discretize the time points in each data segment in the data segment set, obtain multiple time point values ​​after discretization of each data segment, and calculate the volatility of each data segment according to the time point values ​​in each data segment. Calculate the periodicity of each data segment according to the time point values ​​in each data segment, obtain the periodicity value of each data segment, and calculate the time difference between adjacent data segments according to the periodicity value of each data segment. Obtain the time point values ​​in each data segment, calculate the rate of change of the time point values ​​of adjacent data segments, and if the rate of change is greater than the preset rate of change threshold, calculate the growth of the time point values ​​of the adjacent data segments. Calculate the outliers of each data segment according to the volatility and periodicity of each data segment, and if the number of outliers is greater than the preset number threshold, it is determined that the data segment has burstiness. According to the periodicity value of each data segment and the time difference between adjacent data segments, the correlation of adjacent data segments is determined. If the correlation of adjacent data segments is lower than the preset correlation threshold, the adjacent data segment is determined to be an abnormal data segment. According to the burstiness of each data segment and the correlation of adjacent data segments, the abnormal data segment set is determined, and the final abnormal burst recognition result is generated according to the abnormal data segment set and the preset threshold line.

[0079] Specifically, time series data analysis is an important data processing method that can be used to identify anomalies and emergencies. First, continuous time series data needs to be divided into multiple data segments. Next, each data segment is discretized. This step can help reduce the complexity of the data and highlight key features. This discretization makes subsequent analysis more convenient and efficient. Calculating the volatility and periodicity of data segments is the key to understanding the characteristics of time series. Volatility reflects the magnitude of data changes, while periodicity reveals the repetitive pattern of data. By calculating the time difference between adjacent data segments, it can be further confirmed whether this periodicity is stable. The calculation of the rate of change helps to discover emergencies. By setting a suitable rate of change threshold, these important business events can be captured in time. Outlier analysis is an important means to discover anomalies. Through the preset quantity threshold, it can be determined whether this anomaly is statistically significant. Correlation analysis of adjacent data segments can help identify sudden changes in data patterns. Finally, by comprehensively considering the suddenness of data segments and the correlation of adjacent segments, the set of abnormal data segments can be determined. Through this systematic approach, valuable information can be extracted from massive time series data, potential risks and opportunities can be discovered in time, and strong support can be provided for the operation and decision-making of enterprises.

[0080] S105, fusing the spatial feature representation and the temporal feature representation, building a classifier through a fully connected layer, outputting the probability distribution of abnormal working conditions, and obtaining a preliminary abnormality recognition result.

[0081] The original data during the operation of the equipment, including spatial data and time series data, is obtained, and the data is preprocessed to remove noise and outliers. For the preprocessed spatial data, a convolutional neural network is used to extract spatial features to obtain a spatial feature representation vector. For the preprocessed time series data, a long short-term memory network is used to extract time series features to obtain a time series feature representation vector. The spatial feature representation vector and the time series feature representation vector are fused, and the fused feature representation vector is obtained by splicing or weighted averaging. The fused feature representation vector is input into the fully connected layer, an abnormal condition classifier is constructed, and the probability distribution of the abnormal condition is output through the softmax activation function. According to the probability distribution of the abnormal condition, a threshold is set to determine whether the current equipment operating condition is abnormal, and the abnormal identification result is obtained. If the preliminary abnormal identification result is abnormal, the alarm mechanism is triggered to notify relevant personnel to perform equipment inspection and maintenance; if the preliminary abnormal identification result is normal, the equipment operation status monitoring continues.

[0082] Specifically, feature fusion is a key step to integrate information from different sources. The spatial feature vector extracted by CNN (such as [0.8, 0.2, 0.5]) can be simply concatenated with the temporal feature vector extracted by LSTM (such as [0.6, 0.7, 0.4]) to obtain the fused feature vector [0.8, 0.2, 0.5, 0.6, 0.7, 0.4]. Alternatively, different weights can be assigned to spatial and temporal features based on experience, such as a spatial feature weight of 0.6 and a temporal feature weight of 0.4, and the fused features can be obtained by weighted average. The fully connected layer, as the core of the classifier, maps the fused features to different abnormal conditions. For example, three categories can be defined: "normal", "slightly abnormal", and "severely abnormal". The softmax function converts the output of the fully connected layer into a probability distribution, such as [0.85, 0.12, 0.03], which means that the probability of normal is 85%, the probability of slight abnormality is 12%, and the probability of severe abnormality is 3%. Setting an appropriate threshold is crucial to balancing detection sensitivity and false alarm rate. For example, it can be stipulated that an alarm is triggered when the probability of "serious anomaly" exceeds 10%. Such a setting can not only detect potential problems in a timely manner, but also avoid too frequent false alarms that disturb maintenance personnel. Through long-term operation and feedback, this threshold can be continuously adjusted to make the system more intelligent and reliable. This deep learning-based anomaly detection method has significant advantages over traditional methods. It can automatically learn complex features, has strong adaptability, and can detect abnormal patterns that are difficult to define manually. At the same time, by fusing spatial and temporal information, it provides a more comprehensive equipment status assessment, which helps to improve detection accuracy and interpretability.

[0083] The fused feature representation vector is input into the fully connected layer to build an abnormal operating condition classifier, and the probability distribution of abnormal operating conditions is output through the softmax activation function.

[0084] According to the fused feature vector, the fully connected layer structure of the abnormal working condition classifier is constructed, and the appropriate number of neurons and layers are set. The softmax activation function is used as the output layer of the fully connected layer to map the feature vector to the probability distribution space of the abnormal working condition. Through the training sample data, the weight parameters of the fully connected layer are optimized using the back propagation algorithm so that the classifier can accurately identify abnormal working conditions. In the training process of the fully connected layer, regularization techniques such as L1 and L2 regularization are introduced to prevent model overfitting and improve the generalization ability of the classifier. Different loss weights can be set for different types of abnormal working conditions to enhance the sensitivity of the classifier to key abnormalities. In the inference stage, the feature vector of the data to be tested is input into the trained fully connected layer, and the probability distribution of the abnormal working condition is calculated by the softmax function. According to the prior probability and loss cost of the abnormal working condition, an appropriate decision threshold is set, and abnormal working conditions above the threshold are judged as faults, triggering corresponding alarms and handling measures.

[0085] Specifically, constructing a fully connected layer structure for abnormal condition classifiers is a key step in identifying abnormal conditions of equipment. The design of the fully connected layer needs to consider the number of neurons and the number of layers to balance the complexity and performance of the model. The application of the Softmax activation function in the output layer enables the classifier to output the probability distribution of each abnormal condition. Assuming that the monitored industrial equipment has five possible abnormal conditions, the softmax function converts the output of the fully connected layer into five probability values, the sum of which is 1. This probabilistic output facilitates subsequent decision-making and threshold setting. During the training process, the back propagation algorithm minimizes the difference between the predicted results and the true labels by continuously adjusting the network weights. In order to improve the generalization ability of the model, it is crucial to introduce regularization technology. L1 regularization can produce sparse solutions, which is helpful for feature selection; L2 regularization can prevent the weight value from being too large and suppress overfitting. In practice, the appropriate regularization method can be selected according to the characteristics of the data, or two regularization techniques can be used at the same time. For different types of abnormal conditions, setting different loss weights can increase the sensitivity of the model to key abnormalities. Doing so can enable the model to pay more attention to important abnormal types during training and improve the ability to identify key faults. In the inference stage, the real-time collected equipment operation data is extracted and fused, and then input into the trained fully connected layer. The probability distribution calculated by the softmax function reflects the possibility that the current equipment state belongs to each abnormal condition. When setting the decision threshold, it is necessary to consider the prior probability of abnormal conditions and the cost of misjudgment. The abnormal condition classifier constructed by this method can effectively identify the abnormal state of the equipment, providing an important basis for timely maintenance and fault prevention. At the same time, by continuously accumulating data and optimizing the model, the performance of the classifier can be continuously improved, providing a strong guarantee for the safe and stable operation of industrial equipment.

[0086] Embodiment 2

[0087] Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a system for identifying abnormal operating conditions in a sewage treatment process based on a deep neural network, including:

[0088] The data preprocessing module is used to obtain and preprocess multi-source heterogeneous data in the sewage treatment process to obtain a standardized data set;

[0089] A dimension reduction processing module is used to perform dimension reduction processing on the standardized data set to obtain a low-dimensional feature vector;

[0090] A spatial feature extraction module is used to construct a feature extraction module based on a low-dimensional feature vector, extract spatial features, and obtain a deep feature representation;

[0091] The time series feature extraction module is used to introduce the time series analysis module to capture the dynamic change rules of time series data and obtain the time series feature representation;

[0092] The feature fusion and classification module is used to fuse the spatial feature representation with the temporal feature representation, and output the probability distribution of abnormal working conditions through the classifier to obtain the abnormal recognition result.

[0093] The abnormal operating condition identification system for sewage treatment process based on deep neural network provided in this embodiment has all the advantages of the abnormal operating condition identification method for sewage treatment process based on deep neural network provided in the first embodiment.

[0094] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for identifying abnormal conditions in a sewage treatment process based on a deep neural network, characterized in that: The following steps are involved: Obtain multi-source heterogeneous data from the sewage treatment process and pre-process them to obtain a standardized data set; Perform dimensionality reduction on the standardized data set to obtain a low-dimensional feature vector; Construct a feature extraction module based on low-dimensional feature vectors to extract spatial features and obtain deep feature representation; Introduce the time series analysis module to capture the dynamic change rules of time series data and obtain the time series feature representation; The spatial feature representation is fused with the temporal feature representation, and the probability distribution of abnormal working conditions is output through the classifier to obtain the abnormality recognition result.

2. The method according to claim 1, characterized in that The process of obtaining a standardized data set includes: Performing data cleaning on the multi-source heterogeneous data to remove noise and outliers; The normalization method is used to map data of different dimensions to the same scale to obtain a standardized data set.

3. The method according to claim 1, characterized in that The process of obtaining a low-dimensional feature vector includes: Standardize the original high-dimensional data of the standardized data set; The principal component analysis algorithm is used to reduce the dimension of the standardized high-dimensional data set; The dimension threshold is used to determine whether the reduced-dimensionality data set meets the requirements. If not, the parameters are adjusted and the dimension is reduced again.

4. The method according to claim 1, characterized in that: The process of obtaining deep feature representation includes: Build the convolutional neural network architecture and set up the input layer; Configure multiple convolution kernels in the convolution layer and connect the pooling layer to perform feature map dimensionality reduction processing; Repeatedly stack multiple convolutional layers and pooling layers to form a feature extraction structure; Add a fully connected layer after the last convolutional layer to convert spatial features into deep feature representation; If the feature dimension does not meet the requirements, adjust the number of convolution kernels or the number of nodes in the fully connected layer; The network parameters are optimized via a back-propagation algorithm to capture the spatial feature patterns.

5. The method according to claim 1, characterized in that: The process of obtaining the time series feature representation includes: Based on the time series data in the standardized data set, a deep feature representation model is constructed, and the deep features of the data are extracted through a deep learning algorithm; On the basis of deep feature representation, the LSTM network module is introduced, and the memory and forget gate mechanism of LSTM is used to capture the dynamic change law of time series data; The time series data is modeled through the LSTM network module to obtain the time series feature representation; Adopt anomaly detection algorithm to judge whether there are abnormal emergencies in the data according to the time series characteristics; If an abnormal emergency is detected, the early warning mechanism is triggered; Continuously iterate and optimize the parameters of the deep feature representation model and LSTM network module, and optimize the model parameters through incremental learning; The optimized model is applied to actual scenarios to perform real-time monitoring and anomaly detection on new data.

6. The method according to claim 5, characterized in that The process of judging whether the data has abnormal sudden conditions based on the time series characteristics includes: Obtaining time points in a continuous time series, dividing the time points into multiple data segments according to a preset time window, and obtaining multiple sets of data segments arranged in time order; Using a preset sampling rate, the time series points in each data segment in the data segment set are discretized to obtain multiple time series point values ​​after discretization of each data segment. According to the time series point values ​​in each data segment, the volatility of each data segment is calculated; According to the time point value in each data segment, the periodicity of each data segment is calculated to obtain the periodicity value of each data segment, and according to the periodicity value of each data segment, the time difference between adjacent data segments is calculated; Obtain the value of the time series point in each data segment, calculate the change rate of the value of the time series point in the adjacent data segment, and if the change rate is greater than the preset change rate threshold, calculate the increase of the value of the time series point in the adjacent data segment; According to the volatility and periodicity of each data segment, the outliers of each data segment are calculated. If the number of outliers is greater than the preset threshold, the data segment is judged to be bursty. The correlation between adjacent data segments is determined based on the periodicity value of each data segment and the time difference between adjacent data segments. If the correlation between adjacent data segments is lower than a preset correlation threshold, the adjacent data segment is determined to be an abnormal data segment.

7. The method according to claim 1, characterized in that The process of obtaining anomaly identification results includes: The spatial feature representation vector and the temporal feature representation vector are fused by concatenation or weighted average to obtain a fused feature representation vector; The fused feature representation vector is input into the fully connected layer to construct an abnormal operating condition classifier, and the probability distribution of abnormal operating conditions is output through the softmax activation function; According to the probability distribution of abnormal working conditions, the threshold is set to determine whether the current equipment operating condition is abnormal and obtain the abnormal identification result.

8. The method according to claim 1, characterized in that The process of outputting the probability distribution of abnormal conditions through the softmax activation function includes: According to the fused feature vector, a fully connected layer structure of the abnormal operating condition classifier is constructed, and the number of neurons and layers are set; The softmax activation function is used as the output layer of the fully connected layer to map the feature vector to the probability distribution space of abnormal working conditions; By training sample data, the back propagation algorithm is used to optimize the weight parameters of the fully connected layer so that the classifier can identify abnormal working conditions; For different types of abnormal working conditions, different loss weights are set to enhance the sensitivity of the classifier to key abnormalities; In the inference stage, the feature vector of the data to be tested is input into the trained fully connected layer, and the probability distribution of abnormal conditions is calculated through the softmax function.

9. A sewage treatment process abnormal condition identification system based on deep neural network, characterized in that: include: The data preprocessing module is used to obtain and preprocess multi-source heterogeneous data in the sewage treatment process to obtain a standardized data set; A dimension reduction processing module is used to perform dimension reduction processing on the standardized data set to obtain a low-dimensional feature vector; A spatial feature extraction module is used to construct a feature extraction module based on a low-dimensional feature vector, extract spatial features, and obtain a deep feature representation; The time series feature extraction module is used to introduce the time series analysis module to capture the dynamic change rules of time series data and obtain the time series feature representation; The feature fusion and classification module is used to fuse the spatial feature representation with the temporal feature representation, and output the probability distribution of abnormal working conditions through the classifier to obtain the abnormal recognition result.

Images