Method, medium and system for determining the health of filter screen of plankton pump

Abstract

The present invention provides a method, medium and system for determining the health of a filter of a plankton pump, belonging to the technical field of plankton pumps, and comprising: continuously acquiring multiple high-definition images of a plankton pump filter that is positioned and set for a long time; preprocessing each high-definition image to obtain a preprocessed image; processing multiple preprocessed images using a frame difference method to obtain a change area; using a target tracking method to delete the change area in each preprocessed image to obtain an image to be identified; using a pre-set microbial identification model to identify the first image to be identified to obtain an image of microorganisms attached to the filter and the identified microbial species; establishing different types of microbial reproduction and diffusion models based on the microbial images and microbial species, and calculating the total volume proportion of microorganisms on the filter at each time point as the health of the filter; and solving the technical problem that the prior art is difficult to achieve dynamic monitoring and analysis of the filter status.

Description

Technical Field

[0001] The present invention belongs to the technical field of plankton pumps, and in particular, relates to a method, medium and system for determining the health of a filter screen of a plankton pump. Background Art

[0002] The plankton pump is a highly efficient biochemical device widely used in the field of water supply and drainage treatment. Its function is to use plankton in the water to remove pollutants, and to promote the full contact of dissolved oxygen and nutrients in the water by aerating and stirring the water body, thereby achieving efficient water purification. One of the key components of the plankton pump is the filter, which is used to filter and separate microorganisms in the water and provide a fixed carrier for these microorganisms to grow and reproduce. The working state of the filter directly affects the operating efficiency and effluent quality of the entire plankton pump.

[0003] At present, the filter status monitoring of most plankton pumps relies on manual inspections, that is, people are sent to observe and inspect the filter regularly. This manual monitoring method has many shortcomings: first, the frequency of manual inspections is limited, and it is difficult to find problems on the filter in time, which can easily cause filter blockage and affect the normal operation of the pump station; second, the results of manual inspections are easily affected by the subjective factors of the inspectors, making it difficult to make quantitative assessments; third, manual inspections are time-consuming and labor-intensive, and it is difficult to continue for a long time, making it difficult to achieve dynamic monitoring and analysis of the filter status. Summary of the invention

[0004] In view of this, the present invention provides a method, medium and system for determining the health of the filter of a plankton pump, which can solve the technical problems that the existing method of regularly sending people to observe and inspect the filter is time-consuming and labor-intensive, difficult to sustain for a long time, and difficult to achieve dynamic monitoring and analysis of the filter status.

[0005] The present invention is achieved in that:

[0006] A first aspect of the present invention provides a method for determining the health of a filter screen of a plankton pump, comprising the following steps:

[0007] S10, continuously acquiring multiple high-definition images of the plankton pump filter screen that is positioned and set for a long time;

[0008] S20, preprocessing each high-definition image to obtain a preprocessed image;

[0009] S30, using a frame difference method to process the multiple pre-processed images to obtain a changed area;

[0010] S40, using a target tracking method to delete the changed area in each pre-processed image to obtain an image to be identified;

[0011] S50, using a pre-set microorganism recognition model to recognize the first image to be recognized, and obtaining an image of microorganisms attached to the filter and the recognized types of microorganisms;

[0012] S60, establishing different types of microbial reproduction and diffusion models according to the microbial images and microbial types;

[0013] S70. According to the reproduction and diffusion model of different microorganism types in step S60, the total volume percentage of microorganisms on the filter at each time point is calculated as the health of the filter.

[0014] Wherein, the step S10 specifically includes:

[0015] Step 101, install a high-definition camera on the filter screen of the plankton pump, and position it so that it can be aimed at the filter screen in the same area for a long time and stably to shoot;

[0016] Step 102, setting the camera to automatically take a high-definition image at a certain time interval to continuously obtain the image change process of the filter over a long period of time;

[0017] Step 103: Save the captured high-definition image to a computer storage device to form a data set containing multiple images;

[0018] Step 104: During the shooting process, ensure that the focal length, aperture and exposure parameters of the camera are set appropriately to avoid image blur or distortion caused by heat, vibration or ambient light.

[0019] Step 105: Perform necessary preprocessing operations on the collected image data set, such as denoising, white balance, histogram equalization, etc., to improve the effect of subsequent image processing and analysis.

[0020] Wherein, the step S20 specifically includes:

[0021] Step 201: normalize the original high-definition image and scale it to a fixed size, such as 224x224 pixels;

[0022] Step 202: Use classic noise reduction algorithms such as median filtering and bilateral filtering to remove various noise interferences in the original image;

[0023] Step 203: performing a histogram equalization operation to increase the contrast of the image and make the microbial detail features more prominent;

[0024] Step 204: convert the original RGB color space image into a color space that is more conducive to subsequent feature extraction, such as LAB, HSV, etc.;

[0025] Step 205: If necessary, other preprocessing operations can be performed on the image, such as sharpening, edge detection, etc., to further highlight the texture features of the microorganisms.

[0026] Wherein, the step S30 specifically includes:

[0027] Step 301, selecting two consecutive pre-processed images I(t-1) and I(t);

[0028] Step 302: Calculate the frame difference D(t)=|I(t)-I(t-1)| between the two images, where |·| represents an absolute value operation;

[0029] Step 303: Apply an empirical threshold T to the frame difference image D(t) to obtain a binary changed region mask M(t);

[0030] Step 304: for multiple consecutive images, repeat steps 301-303 to obtain the changed area in the entire image sequence;

[0031] Step 305: Perform morphological filtering on the changed region mask, such as opening operation, closing operation, etc., to further extract the continuous and stable changed region.

[0032] Wherein, the step S40 specifically includes:

[0033] Step 401: Using the changed region mask M(t) obtained in step S30, locate and extract the changed regions on each pre-processed image I(t);

[0034] Step 402: for each changing region, a target tracking algorithm (such as SORT, DEEP SORT, etc.) is used to track it, and a motion trajectory of the region in the continuous image sequence is obtained;

[0035] Step 403: According to the target tracking result, a region deletion operation is performed on each pre-processed image I(t) to remove the detected changed region from the image;

[0036] Step 404: After this operation, a set of "clean" images to be identified Ic(t) can be obtained, in which there is no interference from the changing areas caused by the growth activities of microorganisms.

[0037] Wherein, the microorganism identification model adopted in step S50 includes:

[0038] Step 501: A low-level recognition network is used to obtain the types of microorganisms in the image to be recognized based on the texture features, color features and shape features of the image to be recognized Ic(1) and mark the types of microorganisms in each area;

[0039] Step 502: supplementing the recognition network, which is used to supplement the microorganisms blocked in the image to be recognized based on the output of the low-level recognition network, and obtain the actual distribution matrix of each microorganism;

[0040] Step 503, an advanced recognition network is used to generate a comprehensive feature vector describing the distribution and quantity of plankton on the filter based on the actual distribution matrix of each microorganism, reflecting the overall distribution and quantity information of the microorganisms on the filter.

[0041] Wherein, the step S60 specifically includes:

[0042] Step 601: Analyze the dynamic change rules of each type of microorganism on the filter screen by using the actual distribution matrix of each type of microorganism obtained in step S50 and combining it with the corresponding timestamp information;

[0043] Step 602: For different types of microorganisms, mathematical models of their reproduction and diffusion are established, such as Logistic model, Ricker model, etc., and parameter fitting is performed in combination with actual observation data;

[0044] Step 603: Combine the established microbial reproduction models to form a comprehensive microbial community evolution model to describe the dynamic changes of the entire filter microbial community over time.

[0045] Wherein, the step S70 specifically includes:

[0046] Step 701: using the microbial propagation and diffusion model established in step S60, predicting the distribution matrix of each microorganism at a certain time point t+1 in the future;

[0047] Step 702: Add up the area or volume occupied by each type of microorganism on the filter to obtain the total microorganism proportion H(t+1);

[0048] Step 703, continuously calculating the total proportion of microorganisms H(t+1) at different time points t+1 to obtain the dynamic change trend of the microorganisms on the filter during the entire monitoring period;

[0049] Step 704: Based on this change trend, the health index of the filter can be defined, such as the total proportion of microorganisms at a certain moment as the instantaneous health index, and the average proportion of microorganisms in the life cycle of the filter as the comprehensive health index;

[0050] Step 705: A health threshold may be further set. When the total proportion of microorganisms is lower than a certain threshold, the filter is considered to be in a healthy state; when it exceeds another higher threshold, the filter is considered to be in an unhealthy state.

[0051] Optionally, the size of the high-definition image is uniformly 224x224 pixels; the shooting time interval of the camera is 5 minutes; the value range of the empirical threshold T is 1% to 5% of the image grayscale value; the target tracking algorithm includes the SORT algorithm and the DEEP SORT algorithm; the microbial reproduction and diffusion model includes the Logistic model and the Ricker model; the lower value of the health threshold is 20% of the total proportion of microorganisms, and the higher value is 60%; the health index also includes the balance of microbial species and the changing trend of dominant strains.

[0052] Furthermore, the microorganism recognition model includes a low-level recognition network, a supplementary recognition network, and a high-level recognition network.

[0053] The steps of training the microbial recognition model are briefly described as follows:

[0054] 1. First, train the low-level recognition network so that it can accurately identify the species characteristics of various microorganisms.

[0055] 2. Then train the supplementary recognition network so that it can supplement the occluded microorganisms and obtain the actual distribution matrix of each microorganism.

[0056] 3. Finally, train the advanced recognition network so that it can comprehensively utilize microscopic and macroscopic features to generate a feature vector that describes the microbial status of the filter.

[0057] Furthermore, the low-level recognition network is used to obtain the types of microorganisms in the image to be identified based on the texture features, color features and shape features of the image to be identified and to mark the types of microorganisms in each area of ​​the image to be identified.

[0058] Furthermore, the supplementary recognition network is used to supplement the microorganisms that are blocked in the image to be recognized based on the output of the low-level recognition network, so as to obtain the actual distribution matrix of each microorganism.

[0059] Furthermore, the advanced recognition network is used to generate a comprehensive feature vector describing the distribution and quantity of plankton on the filter based on the actual distribution matrix of each organism, reflecting the overall distribution and quantity information of microorganisms on the filter.

[0060] Furthermore, the low-level recognition network specifically includes:

[0061] Input layer: receives the image to be recognized

[0062] Multi-scale convolution layer: Use convolution kernels of different scales to process the input image in parallel to capture subtle features of different sizes;

[0063] Spatial pyramid pooling layer, or SPP layer: divides the feature map into regions of different scales, performs maximum pooling in each region, and extracts local features at different scales;

[0064] Attention mechanism module: including channel attention and spatial attention branches, which enable the model to adaptively focus on the most important subtle regional features;

[0065] Fully connected layer: classify microbial species based on the extracted features;

[0066] Output layer: Output the microbial species label of each area;

[0067] The supplementary recognition network specifically includes:

[0068] Input layer: receives the output of the low-level recognition network;

[0069] Attention mechanism module: including channel attention and spatial attention branches, further enhancing the ability to capture occluded microorganisms;

[0070] Feature pyramid module: integrates feature information of different scales to obtain a richer microbial distribution matrix;

[0071] Deconvolution layer: Generates the distribution matrix of occluded microorganisms;

[0072] Output layer: Output the actual distribution matrix of each microorganism.

[0073] Furthermore, the advanced recognition network specifically includes:

[0074] Input layer: receives the actual distribution matrix of each microorganism;

[0075] Deep convolutional network: Use ResNet or VGG deep convolutional neural network as the backbone network to extract powerful semantic features;

[0076] Dilated convolution: Dilated convolution is used in the last few layers of the backbone network to expand the receptive field and capture broader contextual information;

[0077] Feature pyramid module: fuses feature maps of different depth layers to obtain fine-grained and coarse-grained features;

[0078] Feature fusion layer: splice and fuse micro features and macro features;

[0079] Output layer: Generates a comprehensive feature vector describing the microbial status of the filter.

[0080] The low-level recognition network focuses on the preliminary classification of microbial species based on underlying visual features such as texture, color and shape; the supplementary recognition network focuses on supplementing the occluded microorganisms and obtaining the actual distribution matrix of each microorganism in the image; the high-level recognition network comprehensively utilizes these microscopic and macroscopic features to generate a feature vector that describes the microbial status of the entire filter.

[0081] The microbial reproduction and diffusion model is generated based on population dynamics models commonly used in biological research, such as the Logistic model and the Ricker model.

[0082] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for determining the health of a filter of a plankton pump.

[0083] A third aspect of the present invention provides a system for determining the health of a filter screen of a plankton pump, which includes the above-mentioned computer-readable storage medium.

[0084] Compared with the prior art, the method, medium and system for determining the health of a filter screen of a plankton pump provided by the present invention have the following beneficial effects:

[0085] 1. The accuracy of microbial identification is greatly improved. The present invention adopts a hierarchical deep learning network, including a low-level recognition network, a supplementary recognition network and a high-level recognition network, which can effectively capture the multi-scale visual features of microorganisms and greatly improve the accuracy of microbial identification in complex environments. The low-level recognition network focuses on preliminary classification based on underlying features such as texture, color and shape; the supplementary recognition network focuses on supplementing the obscured microorganisms to obtain a more accurate distribution matrix; the high-level recognition network comprehensively utilizes these microscopic and macroscopic features to generate a comprehensive feature vector that describes the microbial status of the entire filter. This hierarchical network design makes the entire recognition process more robust and reliable.

[0086] 2. The dynamic evolution of microorganisms is accurately modeled. Based on the information of the distribution of microorganisms on the filter, the present invention further establishes mathematical models for the reproduction and diffusion of different types of microorganisms, such as the Logistic model and the Ricker model. These models can accurately describe the dynamic changes of various types of microorganisms on the filter, and provide a reliable basis for predicting the distribution trend of microorganisms in the future. These microbial reproduction models are combined to form a comprehensive microbial community evolution model, which better reflects the dynamic changes of the entire filter microbial community over time.

[0087] 3. The health assessment index is more complete. On the basis of the above steps, the present invention further calculates the proportion of the total volume of microorganisms in the filter at each time point according to the dynamic distribution of microorganisms, as a comprehensive health index of the filter. This indicator can not only intuitively reflect the current operating status of the filter, but also divide the health status by setting reasonable thresholds. In addition, other indicators such as the balance of microbial species, the changing trend of dominant strains, etc. can also be considered to further enrich the dimensions of health assessment. These comprehensive and accurate health assessment indicators provide a reliable decision-making basis for subsequent maintenance.

[0088] In general, the proposed method for determining the health of the plankton pump filter screen achieves accurate identification of microorganisms through a hierarchical deep learning network, predicts future evolution trends based on microbial dynamic modeling, and calculates the overall health of the filter screen by integrating multiple indicators. Compared with the existing technology, this method not only improves the accuracy and reliability of microbial identification, but also realizes dynamic monitoring and comprehensive evaluation of the filter screen status, providing strong support for the intelligent operation and maintenance management of the plankton pump.

[0089] The invention solves the technical problems that the existing method of regularly sending people to observe and inspect the filter is time-consuming and labor-intensive, difficult to sustain for a long time, and difficult to realize dynamic monitoring and analysis of the filter status. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0091] Figure 1 A flow chart of the method provided by the present invention;

[0092] Figure 2 This is the health statistics chart in Example 2. DETAILED DESCRIPTION

[0093] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0094] like Figure 1 FIG. 1 is a flow chart of a method for determining the health of a filter screen of a plankton pump provided by the present invention. The method comprises the following steps:

[0095] S10, continuously acquiring multiple high-definition images of the plankton pump filter screen that is positioned and set for a long time;

[0096] S20, preprocessing each high-definition image to obtain a preprocessed image;

[0097] S30, using a frame difference method to process the multiple pre-processed images to obtain a changed area;

[0098] S40, using a target tracking method to delete the changed area in each pre-processed image to obtain an image to be identified;

[0099] S50, using a pre-set microorganism recognition model to recognize the first image to be recognized, and obtaining an image of microorganisms attached to the filter and the recognized types of microorganisms;

[0100] S60, establishing different types of microbial reproduction and diffusion models based on microbial images and microbial species;

[0101] S70. According to the reproduction and diffusion model of different microorganism types in step S60, the total volume percentage of microorganisms on the filter at each time point is calculated as the health of the filter.

[0102] The specific implementation methods of the above steps are described in detail below:

[0103] The specific implementation of step S10 is: continuously acquiring multiple high-definition images of the plankton pump filter screen that is positioned for a long time. Specifically, the following sub-steps can be used to implement it:

[0104] 1. Install a high-definition camera on the plankton pump filter and position it so that it can be aimed at the filter in the same area for a long time and stably.

[0105] 2. Set the camera to automatically take a high-definition image at a certain time interval (such as 5 minutes). This can continuously obtain the image changes of the filter over a long period of time.

[0106] 3. Save the captured high-definition images to a computer storage device to form a data set containing multiple images. This data set will be used for subsequent image processing and microbial identification analysis.

[0107] 4. To ensure image quality and accuracy of subsequent analysis, please note the following points during the shooting process: Ensure that the focal length, aperture, and exposure parameters of the camera are set appropriately; avoid image blur or distortion caused by heat, vibration, or ambient light; and maintain the stability of the shooting angle and distance.

[0108] 5. For the collected image data set, necessary preprocessing operations need to be performed, such as denoising, white balance, histogram equalization, etc., to improve the effect of subsequent image processing and analysis.

[0109] The main purpose of this step is to obtain a high-definition image sequence of the plankton pump filter over a long period of time, providing the required raw data support for subsequent microbial identification and health assessment. By taking pictures automatically at a fixed time, the dynamic changes in the filter state can be continuously recorded, providing a basis for analyzing the proliferation and spread of microorganisms over time. At the same time, corresponding measures need to be taken to ensure image quality and lay a good foundation for subsequent computer vision analysis.

[0110] The specific implementation of step S20 is: preprocessing each high-definition image to obtain a preprocessed image. Specifically, it includes the following sub-steps:

[0111] 1. Normalize the original HD image. Since the image sizes captured by the camera may vary, they need to be scaled to a fixed size, such as 224x224 pixels. This ensures consistency in subsequent processing.

[0112] 2. De-noise the image. Due to the complexity of the shooting environment, the original image may contain various noise interferences, such as thermal noise, sensor noise, etc. Using classic noise reduction algorithms such as median filtering and bilateral filtering can effectively remove these noises and enhance image details.

[0113] 3. Perform histogram equalization. By adjusting the grayscale histogram of the image, the contrast of the image can be improved, making the microbial details more prominent and easier to identify.

[0114] 4. Perform color space conversion. Convert the original RGB color space image to a color space that is more conducive to subsequent feature extraction, such as LAB, HSV, etc. This color space conversion can make the color features play a better role in subsequent processing.

[0115] 5. If necessary, other preprocessing operations can be performed on the image, such as sharpening, edge detection, etc., to further highlight the texture features of the microorganisms and prepare for subsequent image analysis.

[0116] Through the above preprocessing steps, the quality and analyzability of the image can be effectively enhanced, providing high-quality input data for subsequent image analysis and microbial identification tasks. These preprocessing operations help improve the robustness and generalization ability of the algorithm, thereby obtaining more accurate microbial identification results.

[0117] The specific implementation of step S30 is: using the frame difference method to process multiple pre-processed images to obtain the changed area. Specifically, it includes the following sub-steps:

[0118] 1. Select two consecutive preprocessed images, denoted as I(t-1) and I(t).

[0119] 2. Calculate the frame difference between the two images: D(t) = |I(t)-I(t-1)|

[0120] Here, |·| represents the absolute value operation. The frame difference D(t) reflects the degree of change between two consecutive images.

[0121] 3. Apply an empirical threshold T to the frame difference image D(t) to obtain the binary change area mask M(t):

[0122]

[0123] By setting a reasonable threshold T, the area with obvious changes in the image (such as microbial proliferation) can be effectively segmented from the background. The value of the threshold T can be adjusted according to the actual situation, usually 1% to 5% of the image gray value.

[0124] 4. Repeat steps 2-3 for multiple consecutive images to obtain the changed area mask M(t) of each frame. By superimposing these mask images, we can obtain the changed area in the entire image sequence.

[0125] 5. In order to eliminate the influence of noise, morphological filtering can be performed on the change area mask, such as opening operation, closing operation, etc., to further extract the continuous and stable change area.

[0126] The frame difference method is a simple and effective image change detection algorithm. By calculating the grayscale difference between adjacent frames, it can quickly locate the changed areas in the image. These changed areas are likely to correspond to the growth and reproduction process of microorganisms on the filter, providing important clues for subsequent microbial identification tasks.

[0127] The specific implementation of step S40 is: using the target tracking method, deleting the changed area in each pre-processed image, and obtaining the image to be identified. Specifically, it includes the following sub-steps:

[0128] 1. Using the changed region mask M(t) obtained in step S30, locate and extract these changed regions on each pre-processed image I(t).

[0129] 2. For each changing area, use the target tracking algorithm (such as SORT, DEEP SORT, etc.) to track it and obtain the motion trajectory of the area in the continuous image sequence.

[0130] 3. Based on the results of target tracking, a region deletion operation can be performed on each pre-processed image I(t) to remove the detected changed area from the image.

[0131] 4. After this step, a set of "clean" images to be identified Ic(t) can be obtained, in which there is no interference from the changing areas caused by microbial growth activities.

[0132] In the target tracking algorithm, the visual features of the target, such as position, size, and shape, can be used in combination with Kalman filtering or association matching to track the same target in continuous images. By tracking the motion trajectory of the changing area, it is possible to distinguish which areas are due to continuous changes caused by microbial growth and which areas are just temporary interference or noise.

[0133] The key role of this step is to eliminate the dynamic interference caused by microbial growth and obtain a set of relatively "static" images to be identified, creating more favorable conditions for subsequent microbial species identification tasks.

[0134] The specific implementation of step S50 is: use a pre-set microorganism recognition model to recognize the first image to be recognized, and obtain the image of microorganisms attached to the filter and the recognized microorganism types. This microorganism recognition model includes the following three sub-networks:

[0135] 1. Low-level recognition network:

[0136] Input layer: receives the image to be recognized;

[0137] Multi-scale convolution layer: Use convolution kernels of different scales (such as 3x3, 5x5, 7x7) to process the input image in parallel to capture subtle features of different sizes;

[0138] Spatial Pyramid Pooling (SPP) layer: divides the feature map into regions of different scales (such as 1x1, 2x2, 4x4), performs maximum pooling in each region, and extracts local features at different scales;

[0139] Attention mechanism module: including channel attention and spatial attention branches, which enable the model to adaptively focus on the most important subtle regional features;

[0140] Fully connected layer: classify microbial species based on the extracted features;

[0141] Output layer: Output the microbial species label of each area;

[0142] 2. Supplementary recognition network:

[0143] Input layer: receives the output of the low-level recognition network;

[0144] Attention mechanism module: including channel attention and spatial attention branches, further enhancing the ability to capture occluded microorganisms;

[0145] Feature pyramid module: integrates feature information of different scales to obtain a richer microbial distribution matrix;

[0146] Deconvolution layer: Generates the distribution matrix of occluded microorganisms;

[0147] Output layer: output the actual distribution matrix of each microorganism;

[0148] 3. Advanced recognition network:

[0149] Input layer: receives the actual distribution matrix of each microorganism;

[0150] Deep convolutional network: Use classic deep convolutional neural networks such as ResNet or VGG as the backbone network to extract powerful semantic features;

[0151] Dilated convolution: Dilated convolution is used in the last few layers of the backbone network to expand the receptive field and capture broader contextual information;

[0152] Feature pyramid module: fuses feature maps of different depth layers to obtain fine-grained and coarse-grained features;

[0153] Feature fusion layer: splice and fuse micro features and macro features;

[0154] Output layer: Generates a comprehensive feature vector describing the microbial status of the filter.

[0155] The steps of microbial recognition model training are briefly described as follows:

[0156] 1. First, train the low-level recognition network so that it can accurately identify the species characteristics of various microorganisms.

[0157] 2. Then train the supplementary recognition network so that it can supplement the occluded microorganisms and obtain the actual distribution matrix of each microorganism.

[0158] 3. Finally, train the advanced recognition network so that it can comprehensively utilize microscopic and macroscopic features to generate a feature vector that describes the microbial status of the filter.

[0159] The steps of microbial recognition model training are described as follows:

[0160] 1. Training of low-level recognition network:

[0161] (1) Data preprocessing: Collect a large number of high-definition image samples of the plankton pump filter and annotate them, marking the types of microorganisms in each area. Perform data enhancement on the original images, such as random cropping, flipping, and rotation, to expand the training data. Convert the image input to a format acceptable to the network model, such as an RGB image with a uniform size of 224x224x3.

[0162] (2) Network training: Initialize the parameters of the low-level recognition network. You can use pre-trained models such as VGG or ResNet as a starting point. Set the loss function to cross entropy loss, and the goal is to minimize the classification error on the training set. Use mini-batch stochastic gradient descent to optimize network parameters. Momentum terms and L2 regularization can also be used to improve training stability and generalization. Set appropriate hyperparameters, such as learning rate, batch size, regularization strength, etc., and adjust through cross-validation to obtain the best performance.

[0163] When both the training error and the verification error converge and tend to be stable, the training of the low-level recognition network can be considered completed.

[0164] (3) Model evaluation: Use the test set to evaluate the trained low-level recognition network and calculate indicators such as classification accuracy, recall, and precision. Analyze the misclassified samples to find out the possible reasons and further optimize the network structure and training strategy.

[0165] 2. Supplement the training of recognition network:

[0166] (1) Data preparation: Using the trained low-level recognition network, batch predictions are performed on a large number of plankton pump filter images to obtain the microbial species labels for each area. From these prediction results, the areas with obscured microorganisms are manually marked. The original images and the corresponding microbial distribution labels are used as training samples for the supplementary recognition network.

[0167] (2) Network training: Initialize the parameters of the supplementary recognition network. The weights of the low-level recognition network can be used for transfer learning. Set the loss function to mean square error loss, and the goal is to minimize the difference between the predicted microbial distribution matrix and the true distribution matrix. Use the mini-batch stochastic gradient descent method to optimize the network parameters, and use momentum terms and L2 regularization. Adjust hyperparameters such as learning rate, batch size, and regularization strength, and find the best configuration through cross-validation. When both the training error and the validation error converge and tend to be stable, the training of the supplementary recognition network can be considered complete.

[0168] (3) Model evaluation: Use the test set to evaluate the supplementary recognition network and calculate the mean absolute error, mean relative error and other indicators of the microbial distribution matrix. Analyze samples with poor prediction results and explore possible improvement directions, such as optimizing the attention mechanism module and enhancing the feature pyramid.

[0169] 3. Training of advanced recognition network:

[0170] (1) Data preparation: Use the trained low-level recognition network and supplementary recognition network to perform batch prediction on a large number of plankton pump filter images to obtain the actual distribution matrix of each microorganism. At the same time, collect the corresponding filter health labels, such as the total volume percentage of microorganisms on the filter at each time point. Use the microbial distribution matrix and health labels as training samples for the high-level recognition network.

[0171] (2) Network training: Initialize the parameters of the high-level recognition network. You can use the pre-trained ResNet or VGG network as a starting point.

[0172] The loss function is set to mean square error loss, and the goal is to minimize the difference between the predicted health feature vector and the true health. The network parameters are optimized using mini-batch stochastic gradient descent, and momentum terms and L2 regularization are used. Hyperparameters such as learning rate, batch size, and regularization strength are adjusted, and the best configuration is found through cross-validation. When both the training error and the validation error converge and tend to be stable, the high-level recognition network training is considered complete.

[0173] (3) Model evaluation: Use the test set to evaluate the advanced recognition network and calculate the average absolute error, average relative error and other indicators between the health feature vector and the true health. Analyze the samples with poor prediction results and explore possible improvement directions, such as optimizing the deep convolutional network structure and adjusting the dilated convolution parameters.

[0174] During the entire training process, a reasonable data set partitioning strategy needs to be adopted, such as randomly partitioning the training set, validation set, and test set. At the same time, cross-validation can also be used to further improve the generalization performance of the model. In addition, during the training process, it is necessary to pay close attention to the convergence of the model and adjust the hyperparameters in time to avoid problems such as overfitting or gradient disappearance / explosion. After the training of each network is completed, they can be integrated into a complete microbial recognition model for actual filter health assessment.

[0175] The entire training process requires a lot of computing resources and time, and requires reasonable planning and optimization of training strategies. Through this step-by-step training method, a highly accurate and robust microbial recognition model can be obtained in the end.

[0176] The specific implementation method of the training of the microorganism recognition model is as follows:

[0177] 1. Dataset preparation: Collect a large number of datasets containing images of plankton pump filters. These images should cover a variety of microbial species, varying degrees of occlusion and contamination. Annotate the dataset, annotate the specific location and species of microorganisms in each image. This can be done with the help of professionals or by using semi-automatic annotation tools of computer vision. The original image data can be obtained from relevant research papers, public datasets provided by the government / enterprise, or by field sampling. For the annotation work, a semi-automatic annotation method based on multi-person annotation and voting mechanism can be used to improve the accuracy and consistency of the annotation. In order to increase the generalization ability of the model, the dataset can be enhanced, such as flipping, rotating, cropping, color jittering, etc.

[0178] 2. Training of low-level recognition network: Divide the labeled data set into training set, validation set and test set. Construct the model architecture of the low-level recognition network, including input layer, multi-scale convolution layer, SPP layer, attention mechanism module and fully connected layer. Use the training set to perform end-to-end supervised training on the low-level recognition network. The training goal is to minimize the classification loss of microbial species in each area. Evaluate the performance of the trained low-level recognition network on the validation set, and tune the network structure and hyperparameters until the model achieves ideal performance on the validation set. Use the test set to evaluate the final trained low-level recognition network to ensure that it can maintain good generalization performance on unseen data. In the multi-scale convolution layer, convolution kernels of different sizes (such as 3x3, 5x5, 7x7) can be used to process the input image in parallel to capture subtle features of different scales. The SPP layer can divide the feature map into regions of different scales such as 1x1, 2x2, 4x4, perform maximum pooling in each region, and extract local features at different scales. The attention mechanism module includes channel attention and spatial attention branches, which can enable the model to adaptively focus on the most important subtle regional features. During the training process, the cross entropy loss function can be used, and optimizers such as SGD and Adam can be used to update parameters. In order to prevent overfitting, techniques such as dropout and L2 regularization can be used, and an early stopping strategy can be adopted.

[0179] 3. Training of the supplementary recognition network: Use the trained low-level recognition network to infer the training set and obtain the microbial species label of each area in each image. According to the output of the low-level recognition network, manually mark the occluded microbial area in each image. These annotation information will serve as the supervision signal of the supplementary recognition network. Construct the model architecture of the supplementary recognition network, including the input layer, attention mechanism module, feature pyramid module and deconvolution layer. Use the training set to perform end-to-end supervised training on the supplementary recognition network. The training goal is to minimize the prediction loss of the occluded microbial area. Evaluate the performance of the trained supplementary recognition network on the validation set, and tune the network structure and hyperparameters until the model achieves ideal performance on the validation set. Use the test set to evaluate the final trained supplementary recognition network to ensure that it can maintain good generalization performance on unseen data. The design of the attention mechanism module can be inherited from the low-level recognition network to further enhance the ability to capture occluded microorganisms. The feature pyramid module can fuse feature information of different scales to obtain a richer microbial distribution matrix. The deconvolution layer can upsample the low-level feature map to generate a distribution matrix of occluded microorganisms. During training, the MSE loss function can be used to optimize the difference between the predicted microbial distribution matrix and the actual distribution matrix.

[0180] 4. Training of the advanced recognition network: Use the trained supplementary recognition network to infer the training set and obtain the actual distribution matrix of each microorganism in each image. Construct the model architecture of the advanced recognition network, including the input layer, deep convolutional network, dilated convolution and feature pyramid module. Use the training set to perform end-to-end supervised training on the advanced recognition network. The training goal is to minimize the prediction loss of the feature vector describing the microbial status of the filter. Evaluate the performance of the trained advanced recognition network on the validation set, and tune the network structure and hyperparameters until the model achieves ideal performance on the validation set. Use the test set to evaluate the final trained advanced recognition network to ensure that it can maintain good generalization performance on unseen data. In the deep convolutional network, classic backbone networks such as ResNet or VGG can be used to extract powerful semantic features. Dilated convolution can be used in the later layers of the backbone network to expand the receptive field and capture broader contextual information. The feature pyramid module can fuse feature maps of different depth layers to obtain fine-grained and coarse-grained features. The feature fusion layer can splice and fuse microscopic features and macroscopic features to generate a comprehensive feature vector that describes the microbial status of the filter.

[0181] During training, the MSE loss function can be used to optimize the difference between the predicted feature vector and the actual feature vector.

[0182] Through hierarchical network design, the multi-scale visual features of microorganisms can be gradually extracted and integrated, and finally comprehensive information such as the types, distribution and quantity of microorganisms attached to the filter can be obtained. This information provides an important basis for subsequent microbial reproduction modeling and health assessment.

[0183] The specific implementation of step S60 is: according to the microorganism image and microorganism type, establish different types of microorganism propagation and diffusion models. Specifically including the following sub-steps:

[0184] 1. Using the actual distribution matrix of each microorganism obtained in step S50, combined with the corresponding timestamp information, analyze the dynamic change pattern of each type of microorganism on the filter.

[0185] 2. For different types of microorganisms, mathematical models of their reproduction and diffusion are established. These models can refer to the population dynamics models commonly used in biological research, such as the Logistic model and the Ricker model, and the parameters can be fitted in combination with actual observation data.

[0186] Taking the Logistic model as an example, its mathematical expression is:

[0187] N(t+1)=N(t)+r*N(t)*(1-N(t) / K);

[0188] Among them, N(t) represents the number of microbial population at time t, r is the intrinsic growth rate, and K is the saturation value of the population. By fitting the two parameters r and K, the natural growth law of a certain microorganism on the filter can be described.

[0189] 3. For different types of microorganisms, their reproduction and diffusion models are established respectively. The parameters of these models can be estimated by fitting with the actual distribution data obtained in step S50 through optimization algorithms such as least squares method or gradient descent method.

[0190] 4. Combine the established microbial reproduction models to form a comprehensive microbial community evolution model to describe the changing dynamics of the entire filter microbial community over time.

[0191] By establishing a microbial reproduction and diffusion model, the future distribution trend of various microorganisms on the filter can be predicted. These model parameters reflect the growth characteristics of different microbial species and provide an important basis for the subsequent evaluation of the health of the filter. The accuracy and reliability of the model directly affect the accuracy of the health assessment, so it is necessary to combine actual observation data for fine modeling and parameter tuning.

[0192] The specific implementation of step S70 is: according to the proliferation and diffusion model of different microorganisms in step S60, the total volume proportion of microorganisms on the filter at each time point is calculated as the health of the filter. It specifically includes the following sub-steps:

[0193] 1. Using the microbial propagation and diffusion model established in step S60, predict the distribution matrix of each microorganism at a certain time point t+1 in the future.

[0194] 2. For each time point t+1, add up the area or volume of each type of microorganism on the filter to get the total microbial proportion:

[0195]

[0196] Where N is the number of microbial species, A i (t+1) is the area occupied by the i-th microorganism at time t+1, A total is the total area of ​​the entire filter.

[0197] 3. By continuously calculating the total proportion of microorganisms H(t+1) at different time points t+1, we can obtain the dynamic change trend of the microorganisms on the filter during the entire monitoring period.

[0198] 4. Based on this trend, the health index of the filter can be defined. For example, the total proportion of microorganisms at a certain moment can be used as the instantaneous health, and the average proportion of microorganisms during the life cycle of the filter can be used as the comprehensive health index.

[0199] 5. You can further set health thresholds. When the total proportion of microorganisms is lower than a certain threshold, the filter is considered to be in a healthy state; when it exceeds another higher threshold, the filter is considered to be in an unhealthy state. These thresholds can be set based on actual experience or industry standards.

[0200] 6. In addition to the total volume ratio, other indicators can also be considered to comprehensively evaluate the health of the filter, such as the balance of various microorganisms, the changing trend of dominant bacteria, etc. This needs to be defined in combination with specific application scenarios and needs.

[0201] Through the above steps, the dynamic changes of microorganisms on the filter in the future can be predicted based on the microbial reproduction and diffusion model established in the early stage, and the health index of the filter can be calculated accordingly. This index can intuitively reflect the current operating status of the filter and provide a basis for subsequent maintenance.

[0202] The whole health assessment process mainly includes three key steps: 1) using the microbial recognition model to obtain the real-time distribution information of microorganisms on the filter; 2) establishing a microbial reproduction and diffusion model based on the principle of biodynamics; 3) calculating the comprehensive health index according to the dynamic change trend of microorganisms. This method based on the combination of computer vision and biodynamics can realize the intelligent and quantitative assessment of the health status of the plankton pump filter.

[0203] It should be noted that in the specific implementation process, each step needs to be reasonably set and tuned according to the actual application scenarios and needs. For example, the network structure and hyperparameters of the microbial recognition model, the mathematical form and parameters of the reproduction and diffusion model, and the definition of the health index all need to be repeatedly verified and optimized through a large amount of experimental data to ensure the accuracy and robustness of the final results.

[0204] In general, this filter health assessment method based on machine vision and biodynamics makes full use of the cutting-edge technological achievements in the two fields of computer vision and biological modeling, realizes intelligent analysis and monitoring of the operating status of the plankton pump, and provides strong support for improving the operation and maintenance management level of the pump station.

[0205] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for determining the health of a filter of a plankton pump.

[0206] A third aspect of the present invention provides a system for determining the health of a filter screen of a plankton pump, which includes the above-mentioned computer-readable storage medium.

[0207] In order to better understand and implement the present invention, a specific embodiment 1 of the present invention in a computer readable storage medium or a computer system for a computer program is provided below: The specific implementation methods of the steps included in this embodiment 1 are as follows:

[0208] The specific implementation of step S10 is as follows:

[0209] In this step, it is necessary to continuously obtain multiple high-definition images of the plankton pump filter screen that is positioned for a long time. Specifically, the following methods can be used:

[0210] First, a high-definition camera needs to be installed on the filter of the plankton pump and positioned. This ensures that the camera can be stably aimed at the filter in the same area for a long time.

[0211] Secondly, the camera needs to be set up to automatically take a high-definition image at a certain time interval Δt. The image taken at time t is recorded as I(t). In this way, the image change process of the filter over a long period of time can be continuously obtained to form an image sequence {I(t)}.

[0212] In order to ensure the image quality and the accuracy of subsequent analysis, the following points should be noted during the shooting process:

[0213] 1. Ensure the camera's focal length f, aperture A and exposure time T e The parameters should be set reasonably to avoid image blur or distortion caused by heat, vibration or ambient light. The reasonable exposure parameters can be calculated by the following formula:

[0214]

[0215] Among them, E represents the image exposure, which needs to be kept within an appropriate range.

[0216] 2. Maintain the stability of the shooting angle θ and distance d to avoid geometric distortion caused by position changes. The relationship between camera parameters and image coordinates can be described by the following coordinate transformation formula:

[0217]

[0218] Among them, f x ,f y is the focal length, c x ,c y are the principal point coordinates, R,t is the rotation and translation matrix of the camera.

[0219] 3. For the collected image data set, necessary preprocessing operations are required, such as denoising, white balance, histogram equalization, etc., to improve the effect of subsequent image processing and analysis. These preprocessing operations can use classic image processing algorithms, such as median filtering, bilateral filtering, etc.

[0220] In general, by taking pictures automatically at a fixed time, the dynamic changes of the filter status can be continuously recorded, providing basic data for analyzing the proliferation and spread of microorganisms over time. At the same time, corresponding measures need to be taken to ensure image quality and lay a good foundation for subsequent computer vision analysis.

[0221] The specific implementation of step S20 is as follows:

[0222] In this step, each high-definition image I(t) needs to be preprocessed to obtain the preprocessed image I p (t). Specifically include the following operations:

[0223] 1. Scale normalization:

[0224] Since the images captured by the camera may have different sizes, they need to be scaled to a fixed size, such as 224×224 pixels. This can be achieved by the following scaling transformation:

[0225] I p (t) = Re size (I (t), 224, 224);

[0226] Among them, Re size (·) represents the image scaling operation.

[0227] 2. Denoising:

[0228] Due to the complexity of the shooting environment, the original image I(t) may have various noise interferences, such as thermal noise, sensor noise, etc. Classic noise reduction algorithms such as median filtering and bilateral filtering can be used to effectively remove these noises and enhance image details:

[0229] I p (t) = Denoise(I(t));

[0230] Among them, Denoise(·) represents the image denoising operation.

[0231] 3. Histogram equalization:

[0232] By adjusting the image I p The grayscale histogram of (t) can improve the contrast of the image, making the microbial details more prominent and easier to identify:

[0233] I p (t) = HistogramEqualization (I p (t));

[0234] Among them, HistogramEqualization(·) represents the histogram equalization operation.

[0235] 4. Color space conversion:

[0236] The original RGB color space image I p (t) Convert to a color space that is more conducive to subsequent feature extraction, such as LAB, HSV, etc.:

[0237] I p (t) = ColorSpaceConversion(I p (t));

[0238] Among them, ColorSpaceConversion(·) represents the color space conversion operation.

[0239] 5. Other pretreatment:

[0240] If necessary, you can also p (t) Perform other preprocessing operations, such as sharpening and edge detection, to further highlight the texture features of microorganisms.

[0241] Through the above preprocessing steps, the quality and analyzability of the image can be effectively enhanced, providing high-quality input data for subsequent image analysis and microbial identification tasks. These preprocessing operations help improve the robustness and generalization ability of the algorithm, thereby obtaining more accurate microbial identification results.

[0242] The specific implementation of step S30 is as follows:

[0243] In this step, the frame difference method is used to process multiple pre-processed images {I p (t)} to obtain the changed area. The specific operations include the following:

[0244] 1. Calculate two consecutive images I p (t-1) and I p Frame difference of (t):

[0245] D(t)=|I p (t)-I p (t-1)|;

[0246] Here, ||·|| represents the absolute value operation, and the obtained D(t) reflects the degree of change between two consecutive images.

[0247] 2. Apply an empirical threshold T to the frame difference image D(t) to obtain the binary change area mask M(t):

[0248]

[0249] By setting a reasonable threshold T, the area with obvious changes in the image (such as microbial proliferation) can be effectively segmented from the background. The value of the threshold T can be adjusted according to the actual situation, usually 1% to 5% of the image gray value.

[0250] 3. For multiple consecutive images {I p (t)}, repeating steps 1-2, we can get the changed area mask {M(t)} of each frame. By superimposing these mask images, we can get the changed area in the entire image sequence.

[0251] 4. In order to eliminate the influence of noise, the change area mask {M(t)} can be subjected to morphological filtering, such as opening and closing operations, to further extract the continuous and stable change area.

[0252] The frame difference method is a simple and effective image change detection algorithm. By calculating the grayscale difference between adjacent frames, it can quickly locate the changed areas in the image. These changed areas are likely to correspond to the growth and reproduction process of microorganisms on the filter, providing important clues for subsequent microbial identification tasks.

[0253] The specific implementation of step S40 is as follows:

[0254] In this step, the target tracking method is needed to delete each pre-processed image I p (t) to obtain the image I to be identified c (t). Specifically include the following operations:

[0255] 1. Using the changed region mask M(t) obtained in step S30, p (t) and extract these changed areas to obtain a set of N target areas

[0256] 2. For each change region R n (t), use the target tracking algorithm (such as SORT, DEEP SORT, etc.) to track it and obtain the region in the continuous image sequence {I p (t)} Target tracking can be performed by utilizing visual features such as the location, size, and shape of the region, combined with Kalman filtering or association matching.

[0257] 3. According to the target tracking results, each preprocessed image I p (t) Perform region deletion to remove the detected changed region from the image and obtain a "clean" image to be identified I c (t):

[0258]

[0259] Through this step, we can get a set of "clean" images to be identified. c (t)}, there is no interference from the changing area caused by microbial growth activities. This creates more favorable conditions for the subsequent microbial species identification task.

[0260] The specific implementation of step S50 is as follows:

[0261] In this step, a pre-set microbial recognition model is used to identify the first image I c (1) Identify and obtain the image of microorganisms attached to the filter and the identified microorganism types. This microorganism identification model includes the following three sub-networks:

[0262] 1. Low-level recognition network:

[0263] The network is used to identify the image I c (1) The texture features, color features and shape features are used to obtain the types of microorganisms in the image to be identified and mark the types of microorganisms in each area. Specifically, it includes:

[0264] (1) Input layer: receiving the image I to be recognized c (1)

[0265] (2) Multi-scale convolutional layer: Use convolution kernels of different scales (such as 3×3, 5×5, 7×7) to process the input image in parallel to capture subtle features of different sizes.

[0266] (3) Spatial pyramid pooling (SPP) layer: The feature map is divided into regions of different scales (such as 1×1, 2×2, 4×4), and maximum pooling is performed in each region to extract local features at different scales.

[0267] (4) Attention mechanism module: including channel attention and spatial attention branches, which enable the model to adaptively focus on the most important subtle regional features

[0268] (5) Fully connected layer: Classify microbial species based on extracted features

[0269] (6) Output layer: Output the microbial species label of each area

[0270] 2. Supplementary recognition network:

[0271] This network is used to identify the image I based on the output of the low-level recognition network. c (1) to supplement the obscured microorganisms and obtain the actual distribution matrix of each microorganism. Specifically include:

[0272] (1) Input layer: receives the output of the low-level recognition network

[0273] (2) Attention mechanism module: including channel attention and spatial attention branches, further enhancing the ability to capture occluded microorganisms

[0274] (3) Feature pyramid module: integrates feature information at different scales to obtain a richer microbial distribution matrix

[0275] (4) Deconvolution layer: Generates the distribution matrix of occluded microorganisms

[0276] (5) Output layer: Output the actual distribution matrix of each microorganism

[0277] 3. Advanced recognition network:

[0278] The network is used to generate a comprehensive feature vector describing the distribution and quantity of plankton on the filter based on the actual distribution matrix of each microorganism, reflecting the overall distribution and quantity information of microorganisms on the filter. Specifically, it includes:

[0279] (1) Input layer: receiving the actual distribution matrix of each microorganism

[0280] (2) Deep convolutional network: Use classic deep convolutional neural networks such as ResNet or VGG as the backbone network to extract powerful semantic features

[0281] (3) Dilated convolution: Dilated convolution is used in the last few layers of the backbone network to expand the receptive field and capture broader contextual information.

[0282] (4) Feature pyramid module: fuses feature maps of different depth layers to obtain fine-grained and coarse-grained features

[0283] (5) Feature fusion layer: concatenate and fuse micro features and macro features

[0284] (6) Output layer: Generate a comprehensive feature vector x describing the microbial status of the filter

[0285] Through hierarchical network design, the multi-scale visual features of microorganisms can be gradually extracted and integrated, and finally comprehensive information such as the types, distribution and quantity of microorganisms attached to the filter can be obtained.

[0286] The specific implementation of step S60 is as follows:

[0287] In this step, it is necessary to establish different types of microbial reproduction and diffusion models based on the microbial images and recognition results obtained in step S50. Specifically, the following operations are included:

[0288] 1. Using the actual distribution matrix m of each microorganism obtained in step S50 i (t), combined with the corresponding timestamp t, the dynamic change law of each type of microorganism on the filter is analyzed. The method of biodynamic modeling can be used to establish a mathematical model of microbial reproduction and diffusion.

[0289] 2. For the i-th microorganism, its population size at time t can be expressed as N i (t) represents. Commonly used population dynamics models such as the Logistic model or the Ricker model can be established to describe its reproduction and diffusion process:

[0290] Logistic Model:

[0291] N i (t+1)=N i (t)+r i ·Ni (t)·(1-N i (t) / K i );

[0292] Among them, r i is the intrinsic growth rate of the i-th microorganism, K i is the saturation value of its population.

[0293] Ricker Model:

[0294] N i (t+1)=N i (t)·exp(r i -α i ·N i (t));

[0295] Among them, α i is the density dependence coefficient.

[0296] By fitting the parameters r of these models i ,K i ,α i Etc., can describe the natural growth pattern of each microorganism on the filter.

[0297] 3. For different types of microorganisms i = 1, 2, ..., M, their reproduction and diffusion models are established respectively. The parameters of these models can be optimized by the least squares method or gradient descent method, combined with the actual distribution data {m i (t)} for fitting estimation.

[0298] 4. Combine the established microbial reproduction models to form a comprehensive microbial community evolution model to describe the dynamic changes of the entire filter microbial community over time. This comprehensive model is recorded as:

[0299] {N(t+1)=F(N(t))};

[0300] in, is the number vector of each microbial population at time t, and F(·) represents the community evolution function composed of various microbial models.

[0301] By establishing a microbial reproduction and diffusion model, the future distribution trend of various microorganisms on the filter can be predicted. These model parameters reflect the growth characteristics of different microbial species and provide an important basis for the subsequent evaluation of the health of the filter. The accuracy and reliability of the model directly affect the accuracy of the health assessment, so it is necessary to combine actual observation data for fine modeling and parameter tuning.

[0302] The specific implementation of step S70 is as follows:

[0303] In this step, it is necessary to calculate the total volume percentage of microorganisms on the filter at each time point as the filter health degree according to the proliferation model of different microorganism species established in step S60. The specific operations include:

[0304] 1. Using the comprehensive microbial community evolution model {N(t+1)=F(N(t))} established in step S60, predict the number of each microbial population at a certain time point t+1 in the future

[0305] 2. For each time point t+1, add up the area or volume of each type of microorganism on the filter to obtain the total microbial proportion H(t+1):

[0306]

[0307] Among them, A i (t+1) is the area or volume occupied by the i-th microorganism at time t+1, A total It is the total area or volume of the entire filter.

[0308] 3. By continuously calculating the total proportion of microorganisms H(t+1) at different time points t+1, we can obtain the dynamic change trend of microorganisms on the filter during the entire monitoring period {H(t+1)}.

[0309] 4. Based on this trend, the health index of the filter can be defined. For example, the total proportion of microorganisms at a certain moment H(t+1) can be used as the instantaneous health, and the average proportion of microorganisms in the life cycle of the filter can be used as As an indicator of overall health.

[0310] 5. You can further set a threshold for health. When the total proportion of microorganisms H(t+1) is lower than the lower threshold H low When the filter is considered to be in a healthy state, the filter is considered to be in a healthy state; when the filter exceeds the higher threshold value H high When the filter is in an unhealthy state, the filter is judged to be in an unhealthy state. These thresholds can be set based on actual experience or industry standards, such as H low =20%,H high =60%.

[0311] In addition to the total volume percentage, H(t+1), other indicators can also be considered to comprehensively evaluate the health of the filter, such as the balance of various microorganisms. The changing trend of dominant strains, etc. This needs to be defined in combination with specific application scenarios and requirements.

[0312] Through the above steps, the dynamic changes of microorganisms on the filter in the future can be predicted based on the microbial reproduction and diffusion model established in the early stage, and the health index of the filter can be calculated accordingly. This index can intuitively reflect the current operating status of the filter and provide a basis for subsequent maintenance.

[0313] In general, the whole health assessment process mainly includes three key steps: 1) using the microbial recognition model to obtain the real-time distribution information of microorganisms on the filter; 2) establishing a microbial reproduction and diffusion model based on the principle of biodynamics; 3) calculating the comprehensive health index according to the dynamic change trend of microorganisms. This method based on the combination of computer vision and biodynamics can realize the intelligent and quantitative assessment of the health status of the plankton pump filter.

[0314] In the specific implementation process, each step needs to be reasonably set and tuned according to the actual application scenarios and needs. For example, the network structure and hyperparameters of the microbial recognition model, the mathematical form and parameters of the proliferation model, and the definition of the health index all need to be repeatedly verified and optimized through a large amount of experimental data to ensure the accuracy and robustness of the final results.

[0315] Specifically, the principle of the present invention is:

[0316] 1. Continuously obtain multiple high-definition images of the plankton pump filter set for a long time. By installing a high-definition camera on the filter and setting a timed automatic shooting mode, the dynamic change process of the filter state can be continuously recorded to provide the required raw data support for subsequent microbial identification and health assessment. At the same time, corresponding measures need to be taken to ensure image quality, such as adjusting exposure parameters, fixing shooting angles, etc., to create good conditions for subsequent computer vision analysis.

[0317] 2. Preprocess each high-definition image, including scale normalization, denoising, histogram equalization, color space conversion and other steps. These preprocessing operations help enhance the quality and analyzability of the image and provide high-quality input data for subsequent microbial identification tasks. Through these operations, the robustness and generalization ability of the algorithm can be effectively improved, thereby obtaining more accurate microbial identification results.

[0318] 3. Use the frame difference method to process multiple pre-processed images to obtain the changed areas. By calculating the grayscale differences between adjacent frames, the changed areas in the image can be quickly located. These changed areas are likely to correspond to the growth and reproduction process of microorganisms on the filter. This step provides important clues for subsequent microbial identification.

[0319] 4. Use the target tracking method to delete the changing areas in each pre-processed image to obtain a "clean" image to be identified. The key role of this step is to eliminate the dynamic interference caused by microbial growth and obtain a set of relatively "static" images to be identified, creating more favorable conditions for subsequent microbial species identification tasks.

[0320] 5. Use the pre-set microbial recognition model to identify the first image to be identified, and obtain the image of microorganisms attached to the filter and the identified microbial species. The recognition model uses a hierarchical deep learning network, including a low-level recognition network, a supplementary recognition network, and a high-level recognition network. This hierarchical network design can effectively capture the multi-scale visual features of microorganisms and greatly improve the accuracy of microbial recognition in complex environments.

[0321] 6. According to the microbial images and microbial species, establish different types of microbial reproduction and diffusion models. These models can refer to the population dynamics models commonly used in biological research, such as the Logistic model and the Ricker model, and describe the natural growth laws of various microorganisms on the filter by fitting relevant parameters. These microbial reproduction models are combined to form a comprehensive microbial community evolution model to better reflect the dynamic changes of the entire microbial community over time.

[0322] 7. According to the proliferation model of different microbial species in step 6, calculate the total volume ratio of microorganisms on the filter at each time point as the filter health. This indicator can not only intuitively reflect the current operating status of the filter, but also divide the health status by setting a reasonable threshold. In addition, other indicators such as the balance of microbial species and the change trend of dominant strains can also be considered to further enrich the dimensions of health assessment.

[0323] In summary, the proposed method for determining the health of the plankton pump filter fully integrates the cutting-edge technologies in the fields of computer vision and biodynamic modeling, and realizes intelligent and quantitative monitoring and evaluation of the filter status. The method first uses a hierarchical deep learning network to highly accurately identify various types of microorganisms attached to the filter, laying the foundation for subsequent dynamic analysis. Then, a mathematical model is established based on the dynamic reproduction law of microorganisms to predict the distribution trend of microorganisms in the future. Finally, the overall health evaluation index of the filter is given by combining microscopic and macroscopic characteristics.

[0324] The following is an example 2 of the present invention in a specific application scenario: A sewage treatment plant in a certain district of a certain city is operating a medium-sized plankton pumping station, which treats 50,000 cubic meters of sewage per day. The pumping station adopts a typical A2 / O process, in which the plankton pump is one of the key equipment in the entire sewage treatment system. The main function of the plankton pump is to promote the full contact between dissolved oxygen and nutrients in the water and microorganisms by aerating and stirring the water body, thereby achieving the removal of organic matter and nitrogen and phosphorus in the sewage. The pumping station runs 18 hours a day, using intermittent aeration, and the average dissolved oxygen concentration is controlled at 2-4 mg / L.

[0325] The filter of the plankton pump, as a carrier of contact oxidation, is directly related to the stable operation of the entire biochemical treatment system. The types and number of microorganisms attached to the filter directly affect the decontamination effect of sewage. Therefore, accurately grasping the distribution of microorganisms on the filter is of great significance for optimizing operating parameters and improving treatment efficiency. At present, the pump station uses regular manual inspections to monitor the status of the filter, which has problems such as low inspection frequency, inability to track in real time, and strong subjectivity of evaluation results. This makes the operation and maintenance management of the pump station uncertain, and it is easy to cause the water quality of the effluent to deteriorate due to problems such as filter blockage.

[0326] In view of the above problems, the sewage treatment plant decided to introduce the plankton pump filter screen health determination method proposed in this invention to realize intelligent and quantitative monitoring and evaluation of the filter screen status. The specific implementation steps are as follows:

[0327] (I) Filter image acquisition

[0328] First, a high-definition industrial camera is installed on the filter screen of the plankton pump and its positioning and calibration are performed. The main parameters of the camera are as follows:

[0329] Shooting resolution: 3840×2160 (4K)

[0330] Photosensitive element: 1 / 1.7 inch CMOS

[0331] Lens focal length: 6mm

[0332] Aperture: F2.0

[0333] Exposure time: 1 / 100s

[0334] White balance: Automatic

[0335] The camera uses automatic shooting mode, automatically capturing a filter image every 5 minutes. In order to ensure image quality, the camera's installation position and angle have been carefully designed to ensure that the shooting angle and distance remain stable and avoid distortion caused by factors such as heat and vibration. At the same time, the camera's exposure, aperture and other parameters have also been optimized so that the filter texture details can be clearly presented.

[0336] After three months of continuous data collection, the system acquired a total of 12,600 filter images. All images were saved on a nearby industrial-grade solid-state drive for subsequent analysis and processing.

[0337] (II) Image preprocessing

[0338] For the original filter image collected, a series of preprocessing operations were first performed:

[0339] 1. Scale Normalization

[0340] Since the shooting distance of the camera may vary slightly depending on environmental factors, the scale of the original image may vary. To ensure the consistency of subsequent analysis, all images were scaled to 224×224 pixels.

[0341] 2. Image Noise Reduction

[0342] The original filter image will inevitably be disturbed by some thermal noise, sensor noise, etc., which will affect the recognition of microbial details. The bilateral filtering algorithm is used to reduce the noise of the image, which effectively removes these noise components while retaining the edge details of the image.

[0343] 3. Histogram Equalization

[0344] By adjusting the grayscale histogram of the image, the contrast of the image is significantly improved, making the texture features of the microorganisms on the filter more prominent, which lays a good foundation for subsequent microbial identification.

[0345] 4. Color space conversion

[0346] The original image uses the standard RGB color space, but the LAB color space is usually more advantageous for microbial feature extraction. Therefore, the image is converted from RGB to LAB space to highlight the characteristics of microorganisms in terms of color, saturation, etc.

[0347] After the above preprocessing, the quality and analyzability of the original filter image have been significantly improved, creating good conditions for subsequent computer vision analysis.

[0348] (III) Detection of areas with microbial changes

[0349] In the preprocessed filter image sequence, the frame difference method is used to detect the changed areas in the image. The specific steps are as follows:

[0350] 1. Calculate the difference between two adjacent frames

[0351] For the image I at time t and time t-1 p (t) and I p(t-1), calculate the absolute difference between them:

[0352] D(t)=|I p (t)-I p (t-1)|

[0353] 2. Binarization change area

[0354] Binarize the difference map D(t) to obtain the mask map M(t) of the changed area:

[0355]

[0356] The empirical threshold set here is 30, that is, when the grayscale difference between two frames of images exceeds 30, it is considered to be an area where significant changes have occurred.

[0357] 3. Timing Filtering

[0358] In order to eliminate noise interference, the change region mask {M(t)} of multiple consecutive frames is temporally smoothed. Specifically, a 5×5 median filter kernel is used to remove sporadic changes that occur in a short period of time and only retain the continuous and stable change region.

[0359] Through the above steps, the areas with significant changes in the filter image sequence were successfully detected. These changes are likely to be caused by the growth and reproduction of microorganisms on the filter. Subsequent microbial identification analysis will be carried out on these changed areas.

[0360] (IV) Microbial identification

[0361] For the detected microbial change areas, a hierarchical deep learning network is used for accurate identification. The recognition network consists of three submodules:

[0362] 1. Low-level recognition network

[0363] The main function of the network is to perform preliminary classification of microorganisms based on the texture, color and shape features of the image. The specific structure of the network is as follows:

[0364] Input layer: receives the preprocessed filter image block, with a size of 224×224×3

[0365] Multi-scale convolution layer: Use convolution kernels of three different scales, 3×3, 5×5, and 7×7, to process the input image in parallel and extract visual features of different granularities.

[0366] Spatial Pyramid Pooling (SPP) layer: The feature map is divided into three scales of 1×1, 2×2, and 4×4, and maximum pooling is performed in each area to capture local features at different scales.

[0367] Attention mechanism module: including channel attention branch and spatial attention branch, which enables the network to adaptively focus on the most important detail areas

[0368] Fully connected layer: classify microbial species based on the extracted features and output the microbial species label for each area

[0369] 2. Supplementary recognition network

[0370] The goal of this network is to supplement the recognition of occluded microorganisms that are missed by the low-level recognition network and obtain the actual distribution matrix of each microorganism. The network structure is as follows:

[0371] Input layer: receives the output features of the low-level recognition network

[0372] Attention mechanism module: including channel attention branch and spatial attention branch, further enhancing the ability to capture obscured microorganisms

[0373] Feature pyramid module: integrates feature information at different scales to obtain a richer expression of microbial distribution

[0374] Deconvolution layer: Generates the distribution matrix of occluded microorganisms

[0375] Output layer: Output the actual distribution matrix of each microorganism

[0376] 3. Advanced recognition network

[0377] The role of this network is to comprehensively utilize microscopic and macroscopic features to generate a comprehensive feature vector that describes the microbial status of the filter. The network structure is as follows:

[0378] Input layer: Receive the actual distribution matrix of various microorganisms

[0379] Backbone network: VGG-16 is used as the backbone of the deep convolutional network to extract powerful semantic features

[0380] Dilated convolution module: Dilated convolution is used in the last few layers of the backbone network to expand the receptive field and capture broader contextual information.

[0381] Feature pyramid module: Fuse feature maps of different depth layers to obtain fine-grained and coarse-grained feature expressions

[0382] Feature fusion layer: concatenate and fuse micro features and macro features

[0383] Output layer: Generates a 32-dimensional feature vector x describing the microbial status of the filter

[0384] The three network modules mentioned above jointly achieve accurate identification of filter microorganisms through end-to-end training. The low-level network is responsible for preliminary classification, the supplementary network fills in the occlusion information, and the high-level network integrates all features and outputs a feature vector that describes the overall microbial status of the filter.

[0385] The three networks were jointly trained using a large number of labeled filter images, and ultimately achieved a 92.7% accuracy rate in microbial species recognition on the test set.

[0386] 5. Microbial dynamic modeling

[0387] With the above microbial identification results as a basis, the corresponding mathematical model was established for the dynamic change law of each microorganism.

[0388] 1. Single microorganism model

[0389] Taking a certain dominant microorganism "nitrifying bacteria" as an example, the Logistic model is used to describe its growth process on the filter:

[0390] N b (t+1)=N b (t)+r b ·N b (t)·(1-N b (t) / K b )

[0391] Among them, N b (t) represents the population of nitrifying bacteria at time t, r b Its intrinsic growth rate, K b is the saturated population number. By fitting the measured data, we get r b =0.18,K b =2.5×10 6 Pieces / cm^2.

[0392] 2. Microbial community model

[0393] In addition to the models of single microorganisms, comprehensive community models containing multiple microbial species were also established:

[0394] N(t+1)=F(N(t))

[0395] in, is the number vector of each microbial population at the tth moment, and F(·) represents the community evolution function composed of each single microbial model. Through this model, the dynamic changes of various microorganisms on the filter in the future can be predicted.

[0396] It should be noted that when establishing these microbial dynamic models, the aforementioned microbial identification results were fully utilized, and the model parameters were repeatedly fitted and calibrated in combination with actual monitoring data to ensure the accuracy and reliability of the model predictions.

[0397] (VI) Filter health assessment

[0398] Based on the results of the above microbial identification and dynamic modeling, a comprehensive assessment of the health of the plankton pump filter was finally achieved. The specific steps are as follows:

[0399] 1. Calculate the proportion of total microbial volume

[0400] Using the microbial community model, we can predict the distribution of various microorganisms N(t+1) at a certain time t+1 in the future. Further, based on the average volume of each microorganism, we can calculate the total volume proportion of microorganisms on the filter at time t+1 H(t+1):

[0401]

[0402] Among them, V i is the average volume of the i-th microorganism, V total is the total volume of the entire filter.

[0403] 2. Define health indicators

[0404] According to the changing trend of the total volume proportion of microorganisms H(t+1), the following two health indicators are defined:

[0405] (1) Instantaneous health H(t+1): reflects the distribution of microorganisms in the filter at a certain moment

[0406] (2) Overall health Indicates the average proportion of microorganisms in the filter during the entire monitoring period

[0407] By setting reasonable health thresholds, the filter's operating status can be divided into three categories: "healthy", "normal" and "unhealthy":

[0408] When H(t+1)<20%, the filter is considered to be in a healthy state.

[0409] When 20%≤H(t+1)≤60%, the filter is considered to be in a normal state.

[0410] When H(t+1)>60%, the filter is considered to be in an unhealthy state.

[0411] In addition, other indicators are also considered, such as the balance of various microorganisms, the changing trend of dominant species, etc., which further enrich the dimensions of health assessment.

[0412] 3. Health monitoring results

[0413] Based on the above health assessment method, the filter status of the plankton pumping station was dynamically monitored for 3 months. The monitoring results are shown in Table 1 below:

[0414]

[0415]

[0416] like Figure 2 As shown in the figure, the monitoring results show that in the first two months, the health status of the filter of the plankton pumping station was relatively stable, with the instantaneous health maintained below 20% and the comprehensive health within 40%, indicating that the filter operated well. However, starting from the third month, the total volume of microorganisms on the filter gradually increased, and in late April it exceeded 60%, reaching the "unhealthy" standard. This shows that the growth of microorganisms on the filter has seriously affected the normal operation of the entire system.

[0417] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for determining the health of a filter screen of a plankton pump, characterized in that: The following steps are involved: S10, continuously acquiring multiple high-definition images of the plankton pump filter screen that is positioned and set for a long time; S20, preprocessing each high-definition image to obtain a preprocessed image; S30, using a frame difference method to process the multiple pre-processed images to obtain a changed area; S40, using a target tracking method to delete the changed area in each pre-processed image to obtain an image to be identified; S50, using a pre-set microorganism recognition model to recognize the first image to be recognized, and obtaining an image of microorganisms attached to the filter and the recognized types of microorganisms; S60, establishing different types of microbial reproduction and diffusion models according to the microbial images and microbial types; S70, according to the reproduction and diffusion model of different microorganism types in step S60, the total volume ratio of microorganisms on the filter at each time point is calculated as the health of the filter; Wherein, the microorganism recognition model includes a low-level recognition network, a supplementary recognition network, and a high-level recognition network; The low-level recognition network is used to obtain the types of microorganisms in the image to be recognized based on the texture features, color features and shape features of the image to be recognized and mark the types of microorganisms in each area of ​​the image to be recognized; The supplementary recognition network is used to supplement the microorganisms blocked in the image to be recognized based on the output of the low-level recognition network to obtain the actual distribution matrix of each microorganism; The advanced recognition network is used to generate a comprehensive feature vector describing the distribution and quantity of plankton on the filter based on the actual distribution matrix of each organism, reflecting the overall distribution and quantity information of microorganisms on the filter.

2. A method for determining the health of a filter screen of a plankton pump according to claim 1, characterized in that: The low-level recognition network specifically includes: Input layer: receives the image to be recognized Multi-scale convolution layer: Use convolution kernels of different scales to process the input image in parallel to capture subtle features of different sizes; Spatial pyramid pooling layer, or SPP layer: divides the feature map into regions of different scales, performs maximum pooling in each region, and extracts local features at different scales; Attention mechanism module: including channel attention and spatial attention branches, which enable the model to adaptively focus on the most important subtle regional features; Fully connected layer: classify microbial species based on the extracted features; Output layer: Output the microbial species label of each area; The supplementary recognition network specifically includes: Input layer: receives the output of the low-level recognition network; Attention mechanism module: including channel attention and spatial attention branches, further enhancing the ability to capture occluded microorganisms; Feature pyramid module: integrates feature information of different scales to obtain a richer microbial distribution matrix; Deconvolution layer: Generates the distribution matrix of occluded microorganisms; Output layer: Output the actual distribution matrix of each microorganism.

3. A method for determining the health of a filter screen of a plankton pump according to claim 2, characterized in that: The advanced recognition network specifically includes: Input layer: receives the actual distribution matrix of each microorganism; Deep convolutional network: Use ResNet or VGG deep convolutional neural network as the backbone network to extract powerful semantic features; Dilated convolution: Dilated convolution is used in the last few layers of the backbone network to expand the receptive field and capture broader contextual information; Feature pyramid module: fuses feature maps of different depth layers to obtain fine-grained and coarse-grained features; Feature fusion layer: splice and fuse micro features and macro features; Output layer: Generates a comprehensive feature vector describing the microbial status of the filter.

4. A method for determining the health of a filter screen of a plankton pump according to claim 3, characterized in that: The microbial reproduction and diffusion model is generated based on a population dynamics model commonly used in biological research.

Images