CN115019095A

CN115019095A - Active sludge state online monitoring method based on convolutional neural network

Info

Publication number: CN115019095A
Application number: CN202210628913.2A
Authority: CN
Inventors: 王爱杰; 赵媛; 陶彧; 许铁夫
Original assignee: Shenzhen Graduate School Harbin Institute of Technology
Current assignee: Heilongjiang University; Shenzhen Graduate School Harbin Institute of Technology
Priority date: 2022-06-06
Filing date: 2022-06-06
Publication date: 2022-09-06

Abstract

The invention discloses an active sludge state online monitoring method based on a convolutional neural network. Periodically acquiring activated sludge microscopic images by using a real-time online microscopic image capturing device; and (3) constructing a sludge form health number model by taking the convolutional neural network as a main network and training, and evaluating the sludge state in the periodically input activated sludge microscopic image in real time by using the trained sludge form health number model. The method takes the microscopic image of the activated sludge as a monitoring object, establishes the sludge form health number model based on the convolutional neural network, and utilizes the sludge form health number model to monitor the activated sludge in the biological treatment system of the sewage treatment plant, so that the health state of the activated sludge can be evaluated, and early warning technical support can be provided for the water inlet risk control of the sewage treatment plant to a certain extent through classification of the output types of the model and training of the model.

Description

Active sludge state online monitoring method based on convolutional neural network

Technical Field

The invention relates to the technical field of sewage treatment, in particular to an active sludge health state online monitoring method based on a convolutional neural network.

Background

In traditional scientific research, evaluation of the state of activated sludge is often analyzed from traditional water quality parameters (removal of pollutants and the like) and sludge property indexes and the like; under the background of the rapid development of a molecular biology technology based on nucleic acid sequencing, deeper and multidimensional information mining and analysis can be carried out on the activated sludge by utilizing methods such as a metagenomic technology, a macrotranscriptomic and a unicellular genomics. Although the above indexes can perform comprehensive and deep state evaluation on the sludge, the detection is time-consuming and labor-consuming, has great time delay (counting for short hours and more months), and is difficult to apply to daily monitoring of actual engineering.

Compared with the prior art, the monitoring operation of the microscope on the sludge is very simple and efficient, and the health state of the activated sludge can be initially judged only by sampling and placing under the microscope. However, the microscopic examination project has the problems of low operation frequency, small sample volume, high requirements on technical personnel and the like in the daily operation of a sewage treatment plant, and often cannot capture the change of the sludge form in time and acutely, and the visual advantage of microscopic examination is lost. In addition, a scientific, systematic, objective and standard sludge evaluation system for the overall health state of the activated sludge in the aeration tank in the actual operation of a sewage plant is lacked, and the sludge state is naturally difficult to monitor in real time.

The research on the sludge morphology finds that the microscopic morphological characteristics of the activated sludge have certain correlation with the physicochemical properties and the water quality condition of the activated sludge. It is considered that there is also a correlation between the image of activated sludge and the health state of activated sludge.

The toxic effect of various types of abnormal sewage on the activated sludge is manifold, and the toxic effect can not only directly or indirectly influence the metabolic activity of the sludge, but also cause deep and irreversible damage to the sludge by influencing the physiological structure of the sludge. Extremely abnormal water inflow can often cause harm to an activated sludge system, so that the treatment effect of sewage is reduced, and the normal operation of a sewage treatment plant is seriously influenced, so that huge economic loss is caused. Therefore, monitoring the water inflow condition in the bioreactor system is of great importance, but in the actual water plant operation process, the monitoring cannot sense the content of each parameter of various types of inflow water in time to respond, the sludge system can be found to be broken down after the water outflow abnormality occurs, especially under the condition that the destabilization process of activated sludge is long. Therefore, directly paying attention to the state health of the activated sludge is the most direct method for monitoring whether the sewage treatment plant is normally operated, and is also a problem to be solved currently.

Disclosure of Invention

Aiming at the problems, the invention provides an online monitoring method for the health state of activated sludge based on a convolutional neural network.

In a first aspect of the present invention, a method for online monitoring of health status of activated sludge based on a convolutional neural network is provided, the method comprising the following steps:

periodically acquiring activated sludge microscopic images by using a real-time online microscopic image capturing device;

and (3) constructing a sludge form health number model by taking the convolutional neural network as a main network:

the sludge form health number model input layer randomly clips input microscopic image data to 224 multiplied by 224, and a built-in normaize function is called to carry out standardization processing on the image data;

the output layer of the sludge form health number model outputs one-dimensional vectors of target quantity categories, which respectively represent the prediction result and the corresponding probability value of the input sample, and the category corresponding to the maximum probability is taken as a classification result;

training a sludge form health number model, wherein pictures in a training set are randomly selected by an input layer of the model in the training process to be overturned in the horizontal direction and then are subjected to standardization treatment;

and the trained sludge form health model evaluates the sludge state in the periodically input activated sludge microscopic image in real time.

Further, the method also comprises the step of displaying the monitored activated sludge state on an online interface in real time, wherein the online interface comprises: an image display window of the online activated sludge, an operation and result display window of sludge state evaluation, and a dissolved oxygen value and pH value display window of the online monitoring sludge suspension.

Further, the target quantity category is a 10-grade category, increasing from the least healthy sludge grade at grade 1 to the most healthy sludge grade at grade 10.

Further, the convolutional neural network is an attention-integrating EfficientNet B0.

Further, the method also comprises the step of carrying out periodic statistical analysis on the sludge grade health number output by the sludge form health number model.

Further, the method further comprises screening abnormal data by using a box type graph before the statistical analysis.

Further, the method further comprises the step of optimizing the prediction result output by the sludge form health number model and the corresponding probability value, and the method comprises the following steps:

utilizing a maximum value normalization method to adjust the probability value corresponding to the prediction result output by the sludge form health number model to be between 0.8 and 1.2, and obtaining an adjusted amplification or reduction coefficient;

and performing product operation on the amplification or reduction coefficient and the prediction result output by the sludge form health number model to obtain an optimized prediction result.

The invention provides an active sludge health state on-line monitoring method based on a convolutional neural network, which is characterized in that a micro-image of active sludge is taken as a monitoring object, a sludge form health number model based on the convolutional neural network is established, the sludge form health number model is utilized to monitor the active sludge in a biological treatment system of a sewage treatment plant, the health state of the active sludge can be evaluated, and early warning technical support can be provided for the water inlet risk control of the sewage treatment plant to a certain extent through classification of output classes of the model and training of the model.

Drawings

FIG. 1 is a schematic flow chart of an online monitoring method for health status of activated sludge based on a convolutional neural network in the embodiment of the invention;

FIG. 2 is a schematic diagram of the structure of MB conv in the embodiment of the present invention;

FIG. 3is a diagram of the SENET structure in the embodiment of the present invention;

FIG. 4 is a schematic drawing of a box chart according to an embodiment of the present invention;

FIG. 5 is a schematic illustration of the use of a boxplot to screen out anomalous data in an embodiment of the present invention;

FIG. 6 is a schematic diagram of a real-time online interface display for monitoring the status of activated sludge in an embodiment of the present invention;

FIG. 7 is a confusion matrix diagram based on the EfficientNet B0 model in an embodiment of the invention;

FIG. 8 is a statistical chart of the results of the online monitoring of the sludge health number in the example of the present invention;

FIG. 9 is a schematic view of a representative sludge image corresponding to the sludge morphology health grade in the definition of the sludge morphology health number in the example of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.

Before discussing exemplary embodiments in greater detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although a flowchart may describe the steps as a sequential process, many of the steps can be performed in parallel, concurrently or simultaneously. In addition, the order of the steps may be rearranged. The process may be terminated when its operations are completed, but may have additional steps not included in the figure. The processes may correspond to methods, functions, procedures, subroutines, subprograms, and the like.

The embodiment of the invention provides an online monitoring method for the health state of activated sludge based on a convolutional neural network, which comprises the following steps:

example 1 based on the invention

The embodiment is used for explaining the online monitoring method of the activated sludge health state based on the convolutional neural network, and the method is used for establishing a sludge form health number model based on the convolutional neural network and evaluating the sludge health state in real time by utilizing the sludge form health number model.

The active sludge health state online monitoring method based on the convolutional neural network in the embodiment is shown in fig. 1, and comprises the following steps:

s100, periodically acquiring an activated sludge microscopic image by using a real-time online microscopic image capturing device;

in a specific implementation process, the real-time online microscopic image capturing device of this embodiment records the activated sludge in each period of the aeration phase in a video (frame rate: 60fps, image resolution: 1920 × 1080) form by using a rolling tape sheet, and the recorded sludge image can represent the whole sludge condition of the period. The camera matched with the on-line microscopic image capturing device is an E3ISPM series camera, a USB3.0 interface is adopted to ensure high transmission rate, and camera picture capturing software MIIIMageViewSetup matched with the camera picture capturing device downloaded from an official website can be used with a computer to display pictures in a visual field.

S200, constructing a sludge form health number model by taking the convolutional neural network as a main network:

in the specific implementation process, the convolutional neural network has certain representative network models including, but not limited to, ResNet 34, MobileNetV2, EfficientNet B0, and the like.

Preferably, the convolutional neural network is an attention-integrating EfficientNet B0, and a specific network structure is shown in table 1:

TABLE 1 EfficientNet-B0 model Structure

The most core structure of the EffectientNet B0 model is a Mobile Inverted Bottleneck convolution module (MB Conv), the structure uses convolution kernels of 3 × 3 and 5 × 5 to extract features in a crossed manner, compared with a residual structure, an attention mechanism is integrated into an Inverted residual, attention and feature extraction of a Network to important parts of a target are enhanced, and the specific structure of the MB Conv is shown in FIG. 2, wherein "SE" is a compression and Excitation Network module (SE Net), and is an attention mechanism module commonly used in the visual field, and the module can help the Network focus on channel features with larger information content, inhibit unimportant channel features, and further improve the representation effect of the model. The specific structure of SE Net is shown in fig. 3.

In the implementation process, although the adopted real-time online microscopic image capturing device performs normalized, standardized and streamlined data collection, all the existing environmental interference factors, such as: and the acquisition working conditions such as illumination, temperature and the like add more interference information such as water drops, bubbles and the like to the image. Variables other than these targets will cause the degree of interest of the target classification object in the visual field to be weakened, so that the network model is more required to transfer attention to the target object, pay more attention to the important features of the target object and suppress unnecessary features. This is exactly what attention mechanism is concerned, and therefore the efficiency of EfficientNet B0 incorporated into attention mechanism is also better.

In the specific implementation process, comprehensive analysis is carried out according to specific parameters such as sewage treatment effect and sludge state of the sludge, and the actual state of the sludge is evaluated, and a new index called sludge form health number is defined by the inventor and used for describing the health state degree of the sludge. Thus, the defined sludge health classification is mainly based on the following: the method has the advantages that the method is supported by objective numerical analysis and qualitative evaluation of the sludge based on manual experience under the conditions of impact time, pollutant removal effect, sludge activity, sludge form in a visual field and the like, and comprehensively considers various indexes such as sludge and sludge function effect, so that the method meets the ultimate goal of exploring a sludge image and sludge health state. The health indexes of the sludge are graded to define 10 grades, and the grade from the 1 st grade of the most unstable sludge grade to the tenth grade of the most healthy sludge is gradually increased, specifically, the 1 st grade, the 2 nd grade, the 3 rd grade, the 4 th grade, the 5 th grade, the 6 th grade, the 7 th grade, the 8 th grade, the 9 th grade and the 10 th grade are respectively, wherein the 10 th grade is healthy sludge which is not impacted, and the 1 st grade is seriously instable sludge which is impacted by high salinity for a long time, and the denitrification and dephosphorization effects are completely lost. The sludge properties of the specific stages are shown in table 2.

TABLE 2 definition of sludge morphology health number

S300, training a sludge form health number model, wherein pictures in a training set are randomly selected by an input layer of the model in the training process, turned over in the horizontal direction and then standardized;

in the specific implementation process, input picture data is randomly cut to 224 multiplied by 224; randomly selecting pictures in a training set to turn over in the horizontal direction, so that the interference of subjective factors possibly brought by manual adjustment on the model effect is fully avoided; and calling a built-in normaize function to carry out standardization processing on the picture data, and converting the data into standard Gaussian distribution so as to accelerate the convergence of the model.

Selecting a training set and a test set with the quantity ratio of 8:2, and adding 625 single-label picture data sheets and 6250 calling pictures in total. And finally, storing the screened picture data in each label folder respectively: 500 training sets (20% of verification sets) and 125 test sets, wherein the data of the verification sets are randomly extracted in proportion in the training sets by calling Pythrch built-in functions.

S400, evaluating the sludge state in the periodically input activated sludge microscopic image in real time by the trained sludge form health digital model.

Example 2 based on the invention

In this embodiment, on the basis of embodiment 1, the method for online monitoring of the health state of activated sludge based on the convolutional neural network further includes periodically performing statistical analysis on the health number of the sludge grade output by the sludge morphology health number model, and screening out abnormal data by using a box diagram before the statistical analysis.

In the specific implementation process, the periodic statistical analysis of the sludge grade health number output by the sludge form health number model mainly takes a single period as a statistical unit to perform statistics on the sludge index monitoring result of the activated sludge. As shown in fig. 8, the state of activated sludge was very healthy in the 3 rd cycle, the sludge grade health number was 10 as a whole, the health state of sludge was gradually deteriorated due to the poisoning of high salinity influent water to sludge, the state of sludge was completely deteriorated by the 11 th and 12 th cycles, and the sludge health index was 1.

In another specific embodiment, based on the prediction result of the EfficientNet B0 model on the target picture, a confusion matrix is drawn by combining the real situation of each picture as shown in FIG. 7, and it can be seen that the model has a good effect on identifying the impact type of the sludge. The calculation of the sludge image recognition accuracy for each level of health state according to the prediction effect in the confusion matrix is shown in table 2.

TABLE 2 identification accuracy of sludge in various stages

As can be seen from the table, the recognition accuracy of the sludge morphology health number model constructed based on EfficientNet B0 on each impact type reaches a higher level.

In the specific implementation process, in the periodic monitoring of the sludge form health number, a certain amount of abnormal data is often generated in the long-term monitoring process of the sludge form to interfere with the statistical result, and the evaluation conclusion of the sludge image is further influenced. Therefore, when the monitoring result is periodically analyzed statistically, the abnormal value is usually first screened, and the abnormal value of the boxed graph is screened by the method adopted in the embodiment.

The box type graph is also called a box whisker graph, is a type of statistical graph which is used for sequentially arranging a group of data and calculating some special indexes to be displayed in the graph, can be used for displaying a group of data dispersion conditions and is mainly used for reflecting the characteristics of original data distribution. Besides carrying out statistical analysis on single group data in the target, helping to objectively identify abnormal values in the data and judge the skewness and the tail weight of the data, the method can also carry out comparison on distribution characteristics of the multiple groups of data. Therefore, the monitoring results were analyzed on a weekly basis using a boxed graph.

The box type graph is used for screening abnormal values, and the application is very wide because the box type graph can carry out targeted data analysis on various distribution types of data. When abnormal values are screened out compared with the normal distribution diagramThe original data is assumed to be normalized, and the distribution of the data is not set when the box type graph is subjected to data processing, so that the effects on different groups of data are kept consistent and cannot be influenced by the special distribution characteristics of each group, and the obtained final result is more reliable. The box plot is depicted in fig. 4. Wherein, the distance between the upper Quartile and the lower Quartile is an interquartile Range (IQR), as shown in formula one; the position calculation of the upper and lower edges is shown as equation two and equation three, where Q ₃ Is upper quartile, Q ₁ The number outside the upper and lower edges is the abnormal value, which is screened out when the data is cleaned.

The method comprises the following steps: iqr ═ Q ₃ -Q ₁

The second formula: upper edge ═ Q ₃ +1.5×IQR

And (3) formula III: lower edge ═ Q ₁ -1.5×IQR

The abnormal values are removed by using a box graph method, namely, corresponding values marked as the abnormal values in the box graph are removed, and the box graphs before and after the abnormal data in a group of data are removed are shown in FIG. 5.

Example 3 based on the invention

In this embodiment, based on embodiment 1, the method for online monitoring of the health status of activated sludge based on a convolutional neural network further includes optimizing a prediction result and a corresponding probability value output by a sludge morphological health number model, including:

In the specific implementation process, in order to make more practical judgment and prediction on the output result and improve the engineering guidance of monitoring data, a method is designed, the output result of a target model is adjusted and optimized, and the sludge form health number with higher precision and better descriptiveness is obtained. The optimization operation mainly comprises two steps of normalization of the sludge form health coefficient and calculation of the sludge form health coefficient.

The standardized operation of the sludge form health coefficient is mainly to carry out numerical calculation on the probability corresponding to the output value and adjust the probability to be between 0.8 and 1.2, and the specific operation process is as follows:

adjusting the original probability to be between 0.8 and 1 according to a maximum value normalization method, wherein normalization calculation is shown as a formula IV; and then, the probability is adjusted and distributed around 1 according to a symmetrical distribution operation, wherein the probability of 1 is kept unchanged, the probability of less than 1 is kept at an absolute distance from 1 to randomly expand or not change, namely, the probability is randomly changed into an amplification coefficient or a reduction coefficient, and the sludge form health coefficient is positioned between 0.8 and 1.2. Among them, it is to be noted that the sludge health index value proposed by the inventor is strictly defined between 1 and 10, and therefore, when the sludge form health number is adjusted, it should be noted that when the original sludge form health number is 1, only the amplification factor is provided and no reduction factor is provided for adjustment; when the original morphological health number is 10, only the reduction coefficient but not the amplification coefficient is used for adjustment. Therefore, the formula of the sludge form health coefficient is shown as the formula IV.

And IV, formula IV:

wherein x is the probability to be processed, and k0 is the output probability after normalization processing; min is the minimum value in the array to be changed, and max is the maximum value in the array to be changed; a is the minimum value of the target interval, and b is the maximum value of the target interval.

And a fifth formula: k ═ rand (0,2) -k ₀ I (the original sludge form health number is 2-9);

the formula is six: k is 2-k ₀ I (the original sludge form health number is 1);

the formula is seven: k is 0-k ₀ L (original sludge form health number is 10);

where k0 is the normalized output probability, k is the adjusted amplification or reduction coefficient, and rand (0,2) refers to the random selection of 0 (reduction coefficient) or 2 (amplification coefficient) to carry over into the operation.

And finally, performing product operation on the normalized sludge form health coefficient and the original sludge form health number (namely the prediction result output by the sludge form health number model), wherein the final optimization result is obtained as shown in the formula eight.

The formula eight: and (4) optimizing the sludge form health number k x the original sludge form health number.

Example 4 based on the invention

In this embodiment, based on embodiment 1, the method for online monitoring of the health status of activated sludge based on a convolutional neural network further includes performing real-time online interface display on the monitored status of activated sludge, where the online interface includes: the method comprises an image display window of the online activated sludge, an operation and result window of sludge state evaluation, and a dissolved oxygen value and pH value window of the online monitoring sludge suspension.

In the specific implementation process, the health state of the activated sludge is monitored on line based on the EfficientNet B0 as an identification model, and the morphological health number of the sludge is synchronously calculated and displayed. The interface in the actual operation process is shown in fig. 6, and mainly comprises a display window 601 of activated sludge in a visual field, an operation window 602 and a result display window 6021 for sludge state evaluation, and a display window 603 for online monitoring of dissolved oxygen value and pH value of sludge suspension.

Aiming at the above method for online monitoring of the health state of the activated sludge based on the convolutional neural network, the embodiment of the invention establishes a sludge form health number model based on the convolutional neural network by taking a microscopic image of the activated sludge as a monitoring object, and monitors the activated sludge in a biological treatment system of a sewage treatment plant by using the sludge form health number model, so that the health state of the activated sludge can be evaluated, and early warning technical support can be provided for the water inlet risk control of the sewage treatment plant to a certain extent by classifying the output types of the model and training the model.

It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. Those skilled in the art will appreciate that the present invention is not limited to the particular embodiments described herein, and that various obvious changes, rearrangements and substitutions will now be apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. An active sludge health state online monitoring method based on a convolutional neural network is characterized by comprising the following steps:

2. The convolutional neural network-based online activated sludge health monitoring method of claim 1, further comprising performing real-time online interface display on the monitored activated sludge state, wherein the online interface comprises: the system comprises an image display window of online activated sludge, an operation and result display window of sludge state evaluation, and a dissolved oxygen value and pH value display window of online monitoring sludge suspension.

3. The convolutional neural network-based online activated sludge health monitoring method as claimed in claim 1, wherein the target number category is 10 class categories, and the healthy number of sludge class grades at level 1 which is the least healthy is increased to the healthy number of sludge class grades at level 10 which is the most healthy.

4. The active sludge health state online monitoring method based on the convolutional neural network as claimed in claim 1, wherein the convolutional neural network is EfficientNet B0 integrated with attention mechanism.

5. The convolutional neural network-based online activated sludge health monitoring method as claimed in claim 3, further comprising performing periodic statistical analysis on the sludge grade health number output by the sludge morphology health number model.

6. The convolutional neural network-based online activated sludge health monitoring method of claim 5, further comprising screening abnormal data with a boxplot before statistical analysis.

7. The convolutional neural network-based online activated sludge health monitoring method as claimed in claim 1, further comprising optimizing the prediction results and corresponding probability values output by the sludge morphology health number model, including: