Abnormality detection method for stationary pollution source monitoring data

Abstract

The invention discloses a stationary pollution source monitoring data-oriented anomaly detection method, which comprises the following steps of: inputting target pollution source monitoring data into a trained adversarial network model, and outputting a probability that the target pollution source monitoring data is normal data; the training process of the adversarial network model comprises the following steps: preprocessing acquired fixed pollution source monitoring data to obtain time sequence data and auxiliary features; the time sequence data and the auxiliary features are fused, and the fused features are input into a generator formed based on VAE-diffusion to be reconstructed; marking the reconstructed data as false, marking the obtained stationary pollution source monitoring data as true, inputting the reconstructed data and the obtained stationary pollution source monitoring data into a neural network discriminator, and outputting a discrimination result; and based on a discrimination result output by the neural network discriminator, back propagation is carried out on the generator by using a total loss function, and training of the generator is completed. According to the method, the accuracy and efficiency of anomaly detection in the stationary pollution source monitoring data can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of ecological environmental protection and data analysis, and more specifically, to an anomaly detection method for fixed pollution source monitoring data. Background Art

[0002] With the rapid development of new generation information technologies such as big data, cloud computing, mobile Internet, and Internet of Things, the total amount of global pollution source monitoring data has shown an explosive growth. These technologies enable the collection of a large amount of monitoring data from fixed pollution sources such as industrial emission outlets. However, due to factors such as equipment failures, operation errors, or external environmental impacts, outliers are inevitable in the data. If these outliers are not effectively identified and processed, they will seriously affect the accuracy of environmental assessment, the scientific nature of policy formulation, and the efficiency of environmental governance.

[0003] Currently, there are the following defects in the anomaly detection of fixed pollution source monitoring data:

[0004] 1. Traditional anomaly detection of fixed pollution source monitoring data mainly relies on rules summarized manually. Although this method can identify preliminary anomalies to a certain extent, the amount of abnormal data that can be marked is limited. And due to the influence of human factors, the reliability of these marks has certain limitations.

[0005] 2. Traditional manual review and machine learning methods are insufficient to process a large amount of multi-dimensional data and are difficult to efficiently achieve anomaly detection of complex data structures, thus affecting the efficiency and accuracy of anomaly detection.

[0006] 3. Most of the currently widely used anomaly detection algorithms are designed based on a single open-source dataset or for image data. For the highly heterogeneous multi-dimensional time-series data of fixed pollution sources, directly applying these algorithms has poor effects. Especially when there are no reliable labels, the existing mainstream algorithms often cannot well adapt to and solve problems.

[0007] 4. The mainstream time-series data preprocessing technology mainly uses a sliding window to extract time-series data. However, the differences between different pollution sources are very large. The traditional sliding window method is mainly designed for relatively unified datasets. If the simple methods of window extraction and normalization are used, the important differences between pollution sources may be ignored, resulting in deviations in subsequent analysis results.

[0008] Therefore, how to improve the accuracy and efficiency of anomaly detection in fixed pollution source monitoring data is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] In view of the above problems, the present invention provides an anomaly detection method for fixed pollution source monitoring data to at least solve some of the technical problems mentioned in the above background art.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] The present invention provides an anomaly detection method for fixed pollution source monitoring data, comprising the following steps:

[0012] Input the target pollution source monitoring data into the trained adversarial network model, and output the probability that the target pollution source monitoring data is normal data;

[0013] Wherein, the training process of the adversarial network model includes:

[0014] S1. Preprocess the obtained fixed pollution source monitoring data to obtain time series data and auxiliary features;

[0015] S2. Fuse the time series data and auxiliary features, and input the fused features into the generator composed of VAE-diffusion for reconstruction to obtain reconstructed data;

[0016] S3. Mark the reconstructed data as false and the obtained fixed pollution source monitoring data as true, and input them into the neural network discriminator composed of a temporal convolutional network to output a discrimination result;

[0017] S4. Based on the discrimination result output by the neural network discriminator, use the total loss function to perform backpropagation on the generator to complete the training of the generator.

[0018] Further, the step S1 specifically includes:

[0019] Perform grouped normalization processing on the obtained fixed pollution source monitoring data;

[0020] Adopt the sliding window technique to extract continuous time series data from the grouped and normalized fixed pollution source monitoring data according to a preset time scale;

[0021] For the fixed pollution source monitoring data corresponding to different pollution sources, identify the categorical variables therein; and for each categorical variable, create a binary vector with the same data length as the categorical variable data as the auxiliary feature after one-hot encoding;

[0022] Store the time series data and auxiliary features as CSV files respectively.

[0023] Further, the performing grouped normalization processing on the obtained fixed pollution source monitoring data specifically includes:

[0024] Obtain the fixed pollution source monitoring data corresponding to different monitoring points;

[0025] The monitoring data of stationary pollution sources are initially classified according to pollutant types to obtain multiple large groups;

[0026] The monitoring data of stationary pollution sources in each large group are normalized.

[0027] Further, in the step S1, the MinMax normalization method is selected.

[0028] Further, the fusion of the time series data and the auxiliary features specifically includes:

[0029] The time series data are subjected to feature compression processing through a one-dimensional convolutional layer to obtain convolutional features;

[0030] The auxiliary features are successively subjected to convolution and pooling processing to obtain pooling features;

[0031] The convolutional features and the pooling features are fused through a fully connected layer to obtain fused features.

[0032] Further, in step S2, in the generator composed of VAE-diffusion, the following processing is performed on the fused features:

[0033] In the encoder, the latent space of the fused features is modeled by introducing a probability distribution, the distribution law of the fused features in the latent space is output, and dimensionality reduction sampling is performed to obtain latent variables;

[0034] In the diffusion model, the latent variables are respectively subjected to forward diffusion processing and reverse diffusion processing to generate initial reconstructed data;

[0035] In the decoder, the feature structure of the initial reconstructed data is restored to obtain the final reconstructed data.

[0036] Further, in step S4, the steps for obtaining the total loss function include:

[0037] (1) Obtain a first loss function generated from the reconstruction error of the generator, expressed as:

[0038]

[0039] where x i is the i-th element in the original input time series data; n represents that there are n elements in the original input data; is the i-th element after reconstruction.

[0040] (2) Obtain a second loss function generated from the generator, expressed as:

[0041]

[0042] Meanwhile, obtain the third loss function for negative feedback training directly acting on the neural network discriminator, expressed as:

[0043]

[0044] where x ∼ p d (x) represents sampling a sample x from the original data; x ∼ p g (x) represents sampling a sample x from the generated data; denotes the expectation; D(x) represents the discrimination result;

[0045] (3) Input the reconstructed data into the artificial rule discriminator aggregated by artificial anomaly detection rules, output the violation situation of the reconstructed data, and obtain the corresponding fourth loss function based on the violation situation;

[0046] Perform weighted summation on the first loss function, the second loss function, and the fourth loss function to obtain the total loss function.

[0047] Furthermore, the fourth loss function is expressed as:

[0048]

[0049] where r j represents the number of violations of the j-th artificial anomaly detection rule; w j represents the current weight of the j-th artificial anomaly detection rule.

[0050] Furthermore, the current weight w j of the j-th artificial anomaly detection rule will be updated according to the total number of violations V in each round of training. Let the total weight W = {w 1 , w 2 ,..., w j}; the update formula is expressed as:

[0051]

[0052] where W new represents the total weight after updating the artificial anomaly detection rules; η represents the learning rate.

[0053] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an anomaly detection method for fixed pollution source monitoring data, having the following beneficial effects:

[0054] The present invention combines the variational diffusion model of convolution and LSTM, innovatively applies the diffusion model to time-series multi-dimensional data, effectively captures complex patterns and abnormal behaviors in highly heterogeneous time-series data, and improves the accuracy of anomaly detection.

[0055] The new loss function designed by the present invention combines the anomaly monitoring rules summarized manually, the reconstruction error, and the discriminant results of the discriminator, enabling the model to effectively train and detect anomalies on unlabeled data, greatly reducing the dependence on labeled data.

[0056] The group normalization and sliding window extraction techniques adopted by the present invention effectively process highly heterogeneous and multi-dimensional time-series data, improving the adaptability and generalization ability of the model to different data sources.

[0057] Through adversarial network training, the model of the present invention can effectively evaluate the training effect during the training process, further reducing the false alarm rate. The adversarial network training improves the model's ability to identify abnormal data, while the multi-loss value mechanism enables the model to effectively evaluate the training effect during the training process of unlabeled data, thereby reducing false alarms.

[0058] The technical solution of the present invention will be further described in detail below through the accompanying drawings and embodiments. Description of the Drawings

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0060] Figure 1 It is a schematic flow chart of the anomaly detection method for fixed pollution source monitoring data provided by the embodiment of the present invention.

[0061] Figure 2 It is a schematic diagram of the adversarial network model framework provided by the embodiment of the present invention.

[0062] Figure 3 It is a schematic diagram of the preprocessing provided by the embodiment of the present invention.

[0063] Figure 4 It is a schematic diagram of the acquisition of reconstructed data provided by the embodiment of the present invention.

[0064] Figure 5 It is a schematic diagram of the VAE-diffusion structure provided by the embodiment of the present invention.

[0065] Figure 6 It is a schematic diagram of the dilated convolutional layer structure provided by the embodiment of the present invention. Detailed Embodiments

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0067] An embodiment of the present invention discloses an anomaly detection method for fixed pollution source monitoring data. As shown in Figure 1 the following, it includes the following steps:

[0068] Input the target pollution source monitoring data into the trained adversarial network model, and output the probability that the target pollution source monitoring data is normal data;

[0069] Among them, the training process of the adversarial network model includes:

[0070] S1. Preprocess the obtained fixed pollution source monitoring data to obtain time series data and auxiliary features;

[0071] S2. Fuse the time series data and auxiliary features, and input the fused features into the generator composed of VAE-diffusion for reconstruction to obtain reconstructed data;

[0072] S3. Mark the reconstructed data as false and the obtained fixed pollution source monitoring data as true, and input them into the neural network discriminator composed of a temporal convolutional network to output a discrimination result;

[0073] S4. Based on the discrimination result output by the neural network discriminator, use the total loss function to perform backpropagation on the generator to complete the training of the generator.

[0074] In the embodiment of the present invention, the generator and the discriminator form an adversarial network (GAN) model structure, as shown in Figure 2 the following. The reconstructed data generated by the generator (i.e., the autoencoder) is used by the discriminator to distinguish normal data from abnormal data. The reconstruction error of the generator, the output of the discriminator, and the artificial rule discrimination result are combined to guide the training of the generator.

[0075] Next, each training step of the above adversarial network model will be described separately.

[0076] In the above step S1, the obtained fixed pollution source monitoring data is preprocessed to obtain time series data and auxiliary features; specifically, it includes:

[0077] (1) Perform grouped normalization processing on the obtained fixed pollution source monitoring data; as shown in Figure 3 the following:

[0078] ① Obtain the fixed pollution source monitoring data corresponding to different pollution sources. For example, in Figure 3 , obtain the fixed pollution source monitoring data of each monitoring point corresponding to Company A under Industry A, and obtain the fixed pollution source monitoring data of each monitoring point corresponding to Company B under Industry B; among them, each monitoring point of Company A and Company B can be regarded as different pollution sources;

[0079] ② Initially divide the fixed pollution source monitoring data according to the pollutant type to obtain multiple large groups. For example, in Figure 3 , divide the fixed pollution source monitoring data of each monitoring point into Pollutant A and Pollutant B; each pollutant type corresponds to a large group;

[0080] ③ Perform normalization processing on the fixed pollution source monitoring data in each large group; specifically, MinMax, Z-score or Robust normalization processing can be used to eliminate the dimensional influence and heterogeneity between different monitoring points and improve the consistency of the data. After experimental verification, MinMax normalization has a slight advantage over the other two methods at the model training level, and its principle is as follows:

[0081]

[0082] Among them, x is a value in the original data; x max is the minimum value in the original data; x min is the maximum value in the original data; x ′ is the normalized data value.

[0083] (2) Adopt the sliding window technique to extract continuous time series data from the grouped and normalized fixed pollution source monitoring data according to a preset time scale (such as 5 hours or 8 hours) to capture abnormal patterns in the time series and prepare input data for subsequent anomaly detection;

[0084] (3) Identify the categorical variables in the fixed pollution source monitoring data corresponding to different pollution sources; and for each categorical variable (such as pollutant type, industry category, production suspension mark, etc.), create a binary vector (usually represented by 0 and 1) with the same length as the data of this categorical variable as the auxiliary feature after one-hot encoding;

[0085] In the embodiments of the present invention, the auxiliary feature is one-hot encoded through one-hot encoding of the auxiliary feature, which enhances the adaptability of the model to the unlabeled fixed pollution source data set and improves the generalization ability of the model;

[0086] (4) Store the time series data and auxiliary features as CSV files respectively for subsequent model reading.

[0087] In the above step S2, a variational diffusion model (VAE-diffusion) that combines convolution and long short-term memory network (LSTM): The present invention constructs a variational diffusion model that combines a convolutional neural network and a long short-term memory network, making full use of data features to improve the accuracy of anomaly detection. The data features after dimensionality reduction by the variational encoder can effectively reduce the complexity of the diffusion model, and the diffusion model can not only learn the effective representation of the data, but also simulate the uncertainty of the latent representation, which is particularly important for anomaly detection. The goal of the variational diffusion model is to minimize the reconstruction loss and the KL divergence loss in the latent space to ensure that the generated reconstructed data can fit the pattern of the original data as much as possible. Specifically:

[0088] Fuse the time series data and auxiliary features, and input the fused features into the generator based on VAE-diffusion for reconstruction to obtain reconstructed data; see Figure 4 as shown:

[0089] (1) Fuse the time series data and auxiliary features, specifically including:

[0090] Perform feature compression processing on the time series data through a one-dimensional convolutional layer to obtain convolutional features;

[0091] Perform convolution and pooling processing on the auxiliary features in sequence to obtain pooled features;

[0092] Fuse the convolutional features and the pooled features through a fully connected layer to obtain fused features.

[0093] (2) Input the fused features into the generator based on the variational autoencoder-diffusion model (VAE-diffusion) for reconstruction to obtain reconstructed data; both the encoder and the decoder in this variational autoencoder (VAE) are composed of LSTM to better extract the core time series features;

[0094] The structure of VAE-diffusion is as Figure 5 shown; specifically, in the generator based on VAE-diffusion, the fused features are processed as follows:

[0095] 1) In the encoder:

[0096] Model the latent space of the fused features by introducing a probability distribution (such as a Gaussian distribution), so as to introduce uncertainty in the reconstruction process, enabling the model to learn the distribution law of the fused features in the latent space, denoted as the latent distribution;

[0097] In this process, two related parameters need to be calculated, namely z mean (the mean of the latent space) and z log_var(Log variance of the latent space), and their respective calculation formulas are as follows:

[0098] z mean = f μ (E(x))

[0099] z log_var = f σ (E(x))

[0100] Among them, E(x) is the output of the encoder; f μ is a fully connected layer that maps the output of the encoder to the mean of the latent space; f σ is another fully connected layer that maps the output of the encoder to the log variance of the latent space.

[0101] After that, the reparameterization trick is used to perform dimensionality reduction sampling on the latent distribution to obtain the latent variable z; this step allows the model to learn the distribution of the data latent space during training;

[0102] The above reparameterization trick allows sampling from the latent space without destroying the gradient. The sampling formula is as follows:

[0103]

[0104] Among them, ∈ is a random noise sampled from the standard normal distribution N(0,1);

[0105] 2) In the diffusion model:

[0106] The latent variable z is input into the diffusion model. Without changing the dimension, it first undergoes several rounds of forward diffusion processing to gradually add noise; then it undergoes reverse diffusion processing through several layers of denoising models composed of LSTM, and the denoising models are gradually trained during this process to achieve more precise noise removal and generate higher-quality reconstructed data; The noise addition for each step of training is shown as follows:

[0107]

[0108] Among them, t represents the number of rounds of adding noise; β represents the mixing ratio of noise in each step; ∈t represents the normal distribution random noise added in each step;

[0109] 3) In the decoder:

[0110] Restore the feature structure of the initial reconstructed data (i.e., restore it to the original temporal feature structure) to obtain the final reconstructed data. Specifically, after removing the noise, the restored z is obtained, and then it is upsampled to the dimension of the original data through the decoder to obtain the reconstructed data of the generator. The mean squared error (MSE) between the reconstructed data and the original data is used as a loss value Loss of the modelreconstruction 。

[0111] In the above step S3, the reconstructed data is marked as false (0), and the obtained fixed pollution source monitoring data is marked as true (1), and both are input into a neural network discriminator composed of a temporal convolutional network (TCN) to output a discrimination result D(x) (that is, the probability that the input data is real data); the core structure of the TCN is a dilated convolutional layer, and its schematic diagram is as Figure 6 shown.

[0112] In the above step S4, based on the discrimination result output by the neural network discriminator, the generator is backpropagated using the total loss function to complete the training of the generator;

[0113] Among them, the steps for obtaining the total loss function include:

[0114] (1) Obtain a first loss function generated from the reconstruction error of the generator, expressed as:

[0115]

[0116] Among them, x i is the i-th element in the original input time series data; n represents that there are n elements in the original input data; is the i-th element reconstructed by the autoencoder.

[0117] (2) Obtain a second loss function acting on the generator, expressed as:

[0118]

[0119] At the same time, a third loss function for negative feedback training directly acting on the neural network discriminator can be obtained, expressed as:

[0120]

[0121] Among them, x ∼ p d (x) represents sampling a sample x from the original data, x ∼ p g (x) represents sampling a sample x from the generated data, Here represents the expectation, and in actual calculation, it is approximated by cumulative averaging; this third loss function is only used for the independent training of the discriminator and does not participate in the subsequent total loss function and generator training;

[0122] (3) Input the reconstructed data into an artificial rule discriminator composed of a summary of artificial anomaly detection rules, output the violation situation of the reconstructed data, and obtain a corresponding fourth loss function based on the violation situation, expressed as:

[0123] LOSS ruler = ∑rj w j

[0124] where r j represents the number of violations of the j-th manual anomaly detection rule; w j represents the current weight of the j-th manual anomaly detection rule; this w j will be updated after each round of training according to the total number of violations V in this round. Let the total weight W = {w 1 , w 2 ,..., w j}; the update formula is expressed as:

[0125]

[0126] where W new represents the total weight after updating the manual anomaly detection rules; η represents the learning rate, usually set to 0.01; e represents the base of the natural logarithm.

[0127] (4) Perform weighted summation on the first loss function, the second loss function, and the fourth loss function to obtain the total loss function, which is used to train the generator.

[0128] The present invention designs a new loss function by combining the manually summarized anomaly monitoring rules, the reconstruction error, and the discrimination result of the temporal convolutional network (TCN) discriminator, enabling the model to effectively evaluate the training effect during the training process with unlabeled data.

[0129] In summary, the embodiments of the present invention provide an anomaly detection method for fixed pollution source monitoring data, which can achieve special preprocessing of fixed pollution source monitoring data and output the anomaly score of the monitoring data by the model. Specifically:

[0130] The embodiments of the present invention use a neural network to process temporal and non-temporal data. To effectively solve the problem of high heterogeneity of discharge data of different enterprises, a grouping extraction method is introduced in combination with the actual situation, improving the processing efficiency of a large amount of data.

[0131] The embodiments of the present invention perform anomaly detection on temporal data based on an adversarial network using a reconstruction method, and combine the currently manually summarized anomaly monitoring rules, the reconstruction error, and the discrimination result of the discriminator to design a new loss function, making full use of manual experience and facilitating the effective evaluation of the training effect during the training process with unlabeled data.

[0132] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0133] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting anomalies in monitoring data of fixed pollution sources, characterized in that: The steps include: Input the target pollution source monitoring data into the trained adversarial network model, and output the probability that the target pollution source monitoring data is normal data; The training process of the adversarial network model includes: S1. Preprocess the acquired fixed pollution source monitoring data to obtain time series data and auxiliary features; S2, fuse the time series data and auxiliary features, and input the fused features into the generator based on VAE-diffusion for reconstruction to obtain reconstructed data; S3, marking the reconstructed data as false, marking the acquired fixed pollution source monitoring data as true, and inputting them into the neural network discriminator composed of the time series convolutional network, and outputting the discrimination result; S4. Based on the discrimination result output by the neural network discriminator, the total loss function is used to back-propagate the generator to complete the training of the generator.

2. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 1, characterized in that: The step S1 specifically includes: The acquired fixed pollution source monitoring data is grouped and normalized; Using sliding window technology, continuous time series data are extracted from the fixed pollution source monitoring data after group normalization according to the preset time scale; For the fixed pollution source monitoring data corresponding to different pollution sources, identify the categorical variables; and for each categorical variable, create a binary vector with the same length as the categorical variable data as an auxiliary feature after one-hot encoding; Store time series data and auxiliary features separately as CSV files.

3. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 2, characterized in that: The grouping and normalization processing of the acquired fixed pollution source monitoring data specifically includes: Obtain monitoring data of fixed pollution sources corresponding to different monitoring points; Initially divide the monitoring data of stationary pollution sources according to pollutant types to obtain multiple large groups; The monitoring data of stationary pollution sources in each large group are normalized.

4. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 2, characterized in that: In the step S1, the MinMax normalization method is selected.

5. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 1, characterized in that: The fusion of time series data and auxiliary features specifically includes: Performing feature compression processing on the time series data through a one-dimensional convolution layer to obtain convolution features; The auxiliary features are sequentially subjected to convolution and pooling processing to obtain pooling features; The convolutional features and pooling features are fused through a fully connected layer to obtain fused features.

6. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 1, characterized in that: In step S2, in the generator based on VAE-diffusion, the fused features are processed as follows: In the encoder, the latent space of the fused features is modeled by introducing probability distribution, the distribution law of the fused features in the latent space is output, and dimensionality reduction sampling is performed to obtain latent variables; In the diffusion model, forward diffusion processing and backward diffusion processing are respectively performed on the latent variables to generate initial reconstructed data; In the decoder, the characteristic structure of the initial reconstructed data is restored to obtain final reconstructed data.

7. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 2, characterized in that: In step S4, the step of obtaining the total loss function includes: (1) Obtain the first loss function generated by the generator reconstruction error, expressed as: Among them, x i is the i-th element in the original input time series data; n means there are n elements in the original input data; is the i-th element after reconstruction; (2) Obtain the second loss function generated by the generator, expressed as: At the same time, the third loss function of negative feedback training directly acting on the neural network discriminator is obtained, which is expressed as: Among them, x~p d (x) represents sampling a sample x from the original data; x~p g (x) represents sampling a sample x from the generated data; represents the expectation; D(x) represents the judgment result; (3) inputting the reconstructed data into an artificial rule discriminator formed by summarizing artificial anomaly detection rules, outputting the violation of the reconstructed data, and obtaining a corresponding fourth loss function based on the violation; The first loss function, the second loss function and the fourth loss function are weightedly summed to obtain the total loss function.

8. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 7, characterized in that: The fourth loss function is expressed as: LOSS ruler =Σr j w j Among them, r j It is represented by the number of violations of the j-th artificial anomaly detection rule; w j Represented as the current weight of the j-th manual anomaly detection rule.

9. The method for detecting anomalies in monitoring data of fixed pollution sources according to claim 8, characterized in that: The current weight w of the jth artificial anomaly detection rule j , will be updated after each round of training according to the total number of violations V in this round, assuming that the total weight W = {w1,w2,...,w j }; The update formula is expressed as: Among them, W new represents the total weight after the artificial anomaly detection rule is updated; η represents the learning rate.

Images