Deep neural network-based energy-saving control method for sedimentation tank sludge scraper

Abstract

A deep neural network-based energy-saving control method for a sedimentation tank sludge scraper, comprising: acquiring turbidity information, operating parameters and temperature information in a sedimentation tank, and inputting same into a deep learning neural network model to acquire prediction results of a sludge level and starting time of a sludge scraper, thereby controlling the starting and stopping of the sludge scraper. The deep learning neural network model comprises a convolutional neural network sub-model and a long short-term memory network sub-model, the convolutional neural network sub-model is used for adaptively extracting local features of water quality on the basis of received turbidity information, operating parameters and temperature information, and the long short-term memory network sub-model is used for performing secondary feature extraction on the extracted local features of the water quality to acquire time-series information of the features, and ultimately outputting the prediction results. The method allows for identification of a sludge level and prediction of starting time of a sludge scraping plate, has a certain anti-interference capability, and ultimately achieves automatic control of the starting and stopping of the sludge scraping plate, saving energy.

Description

An energy-saving control method for sedimentation tank scraper based on deep neural network Technical Field

[0001] The present invention relates to the technical field of sedimentation tank control, and in particular to an energy-saving control method for a sedimentation tank scraper based on a deep neural network. Background Art

[0002] A sedimentation tank is a structure that uses sedimentation to remove suspended solids from water, thereby purifying the water. In sewage treatment processes, activated sludge sedimentation and separation is accomplished in a sedimentation tank. Once the physicochemical or biochemical sludge settles to the bottom of the tank, a scraper is used to scrape the sediment from the bottom into a hopper. For example, a sedimentation tank scraper is disclosed in utility model patent application number CN212523134U.

[0003] The existing control method for sedimentation tank scrapers is to drive the scraper continuously to scrape mud and sand, maintaining a high-load operation at all times. However, in sedimentation tanks, mud and sand are generated at an uneven rate, and the sludge level in the sedimentation tank collection pit increases unevenly. Keeping the scraper running at all times is not an efficient solution for mud and sand removal and increases energy consumption.

[0004] Summary of the Invention

[0005] The purpose of the present invention is to provide an energy-saving control method for a sedimentation tank scraper based on a deep neural network in order to overcome the defect of the above-mentioned prior art that the scraper always maintains a high-load operation state and increases energy consumption.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A method for energy-saving control of a sedimentation tank scraper based on a deep neural network comprises the following steps:

[0008] Obtain turbidity information, operating parameters, and temperature information from the sedimentation tank, and input the obtained turbidity information, operating parameters, and temperature information into a pre-built and trained deep learning neural network model in time sequence to obtain prediction results for mud level and scraper start time, thereby controlling the opening and closing of the scraper;

[0009] The deep learning neural network model includes a convolutional neural network sub-model and a long short-term memory network sub-model connected in sequence. The convolutional neural network sub-model is used to adaptively extract local features of water quality based on the received turbidity information, operating parameters and temperature information. The long short-term memory network sub-model is used to perform secondary feature extraction on the local features of water quality extracted by the convolutional neural network sub-model, obtain the time series information of the features, and finally output the prediction results.

[0010] Furthermore, the deep learning neural network model ultimately reduces the feature dimension of the output of the long short-term memory network sub-model through a fully connected layer and outputs the prediction result.

[0011] Furthermore, during the training process of the deep learning neural network model, iterative optimization is performed through stochastic gradient descent, adaptive momentum estimation and / or Nesteev accelerated gradient.

[0012] Furthermore, during the training process of the deep learning neural network model, mean absolute error, root mean square error, and goodness of fit indicators are used to verify the predictive performance of the model.

[0013] Furthermore, the training data input into the deep learning neural network model during the training process includes turbidity information, operating parameters, temperature information and mud level information. The mud level information is obtained through mud level sensors arranged in the sedimentation tank, and the mud level sensors are arranged at the head, middle and tail of the sedimentation tank.

[0014] Furthermore, the process of obtaining the turbidity information, operating parameters and temperature information in the sedimentation tank includes:

[0015] A temperature sensor, a turbidity sensor and a flow sensor are respectively arranged in the sedimentation tank to obtain temperature information, turbidity information and flow information.

[0016] Furthermore, the temperature sensor is arranged at the water inlet of the sedimentation tank, and the flow sensor is arranged at the water pipe of the sedimentation tank.

[0017] Furthermore, the scraper is connected to a PLC system, and the opening and closing of the scraper is controlled by the PLC system.

[0018] Furthermore, after obtaining the turbidity information, operating parameters and temperature information, each piece of information is first time-synchronized, and then the time-synchronized information is input into the deep learning neural network model.

[0019] Furthermore, the information input into the deep learning neural network model is a multi-dimensional time series signal based on turbidity information, operating parameters and temperature information.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] (1) The present invention is based on a deep learning neural network model, which uses the turbidity sensor information at the entrance to online learn and train the neural network model to output its mud level and scraper start time, and to control the start of the scraper in real time to achieve the purpose of intelligent water and energy saving; According to research, the mud level recognition and control system itself is a nonlinear system with a certain degree of complexity, which is difficult to describe with an accurate digital model, while the neural network can realize nonlinear mapping of input and output through adaptive learning without the aid of any known description. Therefore, the mud level can be identified and the scraper start time can be predicted based on the deep neural network, and it has a certain anti-interference ability, and finally realizes automatic control of the start and stop of the scraper to achieve the purpose of energy saving.

[0022] (2) The present invention takes into account that the convolutional neural network sub-model usually relies on the convolution kernel on the convolution layer to extract features, and then relies on the pooling layer to refine the features, thereby retaining important information of the data. However, the existence of the convolution kernel limits the "long-term dependence problem" of the convolutional neural network sub-model when processing time series signal data. Therefore, the long short-term memory network sub-model is introduced to solve this problem. The long short-term memory network sub-model continues to perform secondary feature extraction on the feature quantity extracted by the convolutional neural network sub-model to obtain the time series information of the features, ensuring that the deep spatial features and time dimension feature information of the input source can be effectively extracted. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] FIG1 is a schematic flow chart of an energy-saving control method for a sedimentation tank scraper based on a deep neural network according to an embodiment of the present invention;

[0024] FIG2 is a structural diagram of a convolutional neural network provided in an embodiment of the present invention;

[0025] FIG3 is a schematic diagram of input and output of a recurrent neural network provided in an embodiment of the present invention;

[0026] FIG4 is a schematic diagram of input and output of a long short-term memory network provided in an embodiment of the present invention;

[0027] Figure 5 is a schematic diagram of the overall framework of a deep learning neural network model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0029] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0030] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0031] Example 1

[0032] This embodiment provides an energy-saving control method for a sedimentation tank scraper based on a deep neural network, comprising the following steps:

[0033] S1: Construct a deep learning neural network model. The deep learning neural network model includes a convolutional neural network sub-model and a long short-term memory network sub-model connected in sequence. The convolutional neural network sub-model is used to adaptively extract local features of water quality based on the received turbidity information, operating parameters and temperature information. The long short-term memory network sub-model is used to perform secondary feature extraction on the local features of water quality extracted by the convolutional neural network sub-model to obtain the temporal information of the features. Finally, the feature dimension of the output of the long short-term memory network sub-model is reduced through the fully connected layer to output the prediction result.

[0034] S2: Obtain turbidity information, operating parameters, temperature information, and mud level information in the sedimentation tank, and input them into the deep learning neural network model for model training to obtain the trained deep learning neural network model;

[0035] During training, iterative optimization is performed using stochastic gradient descent, adaptive momentum estimation, and / or Nesteev accelerated gradient descent. The predictive performance of the model is verified using mean absolute error, root mean square error, and goodness-of-fit metrics.

[0036] S3: Obtain turbidity information, operating parameters, and temperature information in the sedimentation tank in real time, and input the obtained turbidity information, operating parameters, and temperature information into the trained deep learning neural network model in time sequence to obtain the prediction results of the mud level and the scraper start time, thereby controlling the opening and closing of the scraper.

[0037] It should be noted that after obtaining turbidity information, operating parameters and temperature information, it is first necessary to synchronize the time of each piece of information, and then construct the time-synchronized information into a multi-dimensional time series signal and input it into the deep learning neural network model.

[0038] The following is a detailed description of the above scheme:

[0039] 1. Overall Overview

[0040] Based on the deep learning neural network model, wireless sensing of turbidity, flow and other information at the water inlet of the flower wall is used to train the mud level at the output end of the neural network model online, and the scraper is started in real time to control the operation and send information to the scraper to realize the opening and closing of the mud discharge function.

[0041] The specific functions are as follows:

[0042] 1) Monitor the turbidity information at the water inlet of the sedimentation tank wall, and learn and train the mud level at the output end of the neural network model online.

[0043] 2) Start the scraper to scrape mud;

[0044] 3) During the sludge scraping process, when the sludge in the sedimentation tank sludge pit reaches the set height, the sludge discharge valve opens to discharge the sludge;

[0045] 4) Monitor the turbidity value at the outlet of the mud discharge valve and close the mud discharge valve when the water turbidity is low.

[0046] 2. Device selection

[0047] 2.1 Device Type

[0048] Device types include turbidity sensors, temperature sensors, mud level sensors, flow sensors, PLC systems and control boxes.

[0049] The installation positions of the components in this embodiment are shown in Table 1.

[0050] Table 1 Mounting components

[0051] 2.2 Data Transmission

[0052] Signals from installed turbidity and temperature sensors will not be transmitted to the scraper system. Instead, the information will be directly transmitted to the cloud platform via 4G wireless. The cloud platform will obtain data such as turbidity, temperature, and flow, and the deep learning neural network model will be used to obtain the optimal scraper start and close time. This will not affect the scraper's operating system and parameters, and will achieve intelligent water-saving and energy-saving optimization control.

[0053] 3. Algorithm research content

[0054] 3.1 Research Overview

[0055] The main development content of this solution is based on a deep learning neural network model. It uses the turbidity sensor information at the inlet to online learn and train the neural network model to output its mud level or the scraper start time, and control the start of the scraper operation in real time to achieve the purpose of intelligent water and energy saving.

[0056] According to the survey, the mud level identification and control system itself is a nonlinear system with a certain degree of complexity, which is difficult to describe with an accurate digital model. However, the neural network can realize nonlinear mapping of input and output through adaptive learning without the aid of any known description. Therefore, the mud level can be identified and the scraper start-up time can be predicted based on the deep neural network, and it has a certain anti-interference ability.

[0057] 3.2 Research Methods

[0058] The neural network-based mud level identification and scraper start-up time prediction uses the collected sensor information as the input of the neural network model. Since water turbidity, water inflow, and operating temperature may all affect the prediction results, turbidity information, operating parameters, temperature and time can be input as variables. Through iteration and training of the model, the nonlinear relationship between input and output is fitted to achieve the prediction of mud level and scraper start-up time. At the same time, the prediction performance of the model can be verified using evaluation indicators such as mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R2). The neural network model can be built in Python or MATLAB environment.

[0059] 3.3 Model Selection

[0060] Neural network methods such as support vector machines and BP neural networks are used to extract and filter the evolving characteristics of variables for identification and prediction. However, these traditional artificial neural network methods are mostly based on shallow network learning and lack the ability to learn complex nonlinear relationships, resulting in limited accuracy in prediction results. To effectively and simultaneously utilize both spatial and temporal characteristics of sensor-generated signal data, we propose a deep learning neural network model for mud level identification and actuation time prediction based on a combination of a convolutional neural network (CNN) and a long short-term memory (LSTM) network. This network model uses preprocessed sensor data, such as turbidity, as input and extracts spatial and temporal features using a CNN and then an LSTM. It then identifies the corresponding relationships between water turbidity, sediment concentration, and scraper actuation time, thereby predicting both sludge concentration and scraper actuation time.

[0061] 4. Basic theory and related algorithms

[0062] Machine learning (ML) is a new research direction that has evolved from deep learning (DL). Originally proposed by Hinton in 2006, the concept stems from the study of artificial neural networks. The idea is to stack multiple layers, with the output of each layer serving as the input for the next. The model consists of a multi-layer network consisting of an input layer, a hidden layer, and an output layer. Nodes in adjacent layers are connected; nodes in the same layer or across layers are disconnected.

[0063] 4.1 Convolutional Neural Network (CNN)

[0064] A convolutional neural network is a feedforward neural network structured around one or more convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layer is the core layer of the neural network, extracting information about the relevant features of a dataset. Convolution 1 and Convolution 2, shown in Figure 2, are two layers of convolution. The pooling layer, also known as the downsampling layer, is added because extracting features directly from the convolutional layer would be computationally intensive and prone to overfitting. The figure also shows two pooling layers. An activation function is also required. This adds an offset to the output value after the convolution operation to maintain the model's nonlinearity and enhance its ability to learn complex datasets.

[0065] 4.2 Recurrent Neural Network (RNN)

[0066] RNNs are called recurrent neural networks because they have a self-repeating, cyclical structure, which allows them to retain information. This means that the current output of a sequence of data is related to the output at both the current and previous moments. This memory function is achieved thanks to the RNN model structure, which allows it to pass previous information to subsequent hidden layers, thus achieving a memorization effect. RNNs can remember information processed at time t and perform calculations at subsequent times. Figure 3 shows the RNN model structure.

[0067] The RNN model structure is relatively simple, consisting of an input layer, a hidden layer, and an output layer. The input layer x and the hidden layer s (input weights), as well as the hidden layer s and the output layer o (output weights), are fully connected. The hidden layer s itself has a self-loop structure, as shown in the left figure (w). An expanded RNN model is shown in the right figure of Figure 4.

[0068] In the RNN expansion diagram, t-1, t, and t+1 represent the time sequence. Each arrow has a transformation, that is, it has a corresponding weight. St represents the hidden state of the data at time point t. Its calculation process is affected by the state calculated at the previous time point. See the following formula: S t =f(W*S t-1 +U*Xt )

[0069] Here, f is the activation function, and RNN generally uses Relu as the activation function. The input layer is represented by x, W is the weight matrix of the previous sample, U is the weight matrix of the current input sample, and V is the output weight matrix. Although the RNN algorithm handles time series problems well, when the time series is long, the memory value is small due to the long time. In other words, the gradient explosion or gradient vanishing problem is prone to occur, which reduces the accuracy of the RNN model.

[0070] 4.3 Long Short-Term Memory (LSTM)

[0071] The Long Short-Term Memory (LSTM) network is a variant of the RNN (recurrent neural network). However, unlike standard RNNs, the LSTM improves upon the traditional RNN structure by introducing a gating mechanism and cell units, enabling it to capture dependencies between sequential signals. While the RNN structure can cause exploding or vanishing gradients, the LSTM addresses these issues by improving the hidden layer with a gating structure. The LSTM is therefore well-suited for processing time series problems, as shown in Figure 4.

[0072] 4.4 Activation Function

[0073] The activation function is added to introduce a nonlinear structure to the network. Because the previous convolutional layers and fully connected layers are all linear operations, the linear results are obtained. However, for more complex data sets, simple linear fitting cannot achieve good prediction or classification. Therefore, nonlinear components must be added to the network. Common activation functions include Sigmoid, Tanh, Relu, etc. The Sigmoid function is the most widely used activation function in the early stage. The output value range is 0-1. The function formula is as follows:

[0074] The Sigmoid function is simple and has good nonlinear mapping. The following shortcomings can be seen from the function curve and derivative curve: (1) The problem of vanishing gradient. According to its derivative function graph, when the input is large or small, the derivative value of the function tends to 0. After the multi-layer network is calculated through the activation function, the output value will decrease rapidly, causing the gradient to vanish. (2) It cannot reach the zero mean state, and the output value of the Sigmoid function is always positive.

[0075] Tanh activation function has an output value of (-1, 1). Compared with the sigmoid function, the derivative range is larger, which can make the network converge better. The calculation expression of this function is shown in the following formula.

[0076] It can be seen that the Tanh function has positive and negative values, so compared with the sigmoid function, its zero mean problem is solved, but the change curve of the Tanh derivative function is similar to that of the sigmoid, and the gradient disappears.

[0077] The Relu function is the most commonly used activation function today, which can improve the gradient vanishing problem. The calculation formula of this function is as follows: Relu = max(0, x)

[0078] When the input is negative, the output value is zero after passing through the Relu function, and the left derivative is also zero. When the input is positive, it changes in linear proportion, and there is no maximum value limit, which solves the gradient vanishing problem. The derivatives are different when the input is positive and negative. When the input is positive, the derivative is a constant 1, thus avoiding the problem of gradient vanishing during continuous multiplication. When the input is negative, the output is always zero (the feature is masked). Because deep learning often processes large data samples, this feature can retain key information and adjust the sparse ratio.

[0079] 4.5 Evaluation Indicators

[0080] In order to evaluate the performance of the model, the mean absolute error (MAE), root mean square error (RMSE), and goodness of fit (R^2) are used to evaluate the prediction results. The calculation expressions are shown in the following formula:

[0081] in, is the predicted value of sludge concentration (or scraper start-up time) obtained by the model, is the mean sludge concentration (or scraper start time), and Yi is the true value of sludge concentration (or scraper start time). The smaller the MAE and RMSE values, the smaller the prediction error; the closer R2 is to 1, the better the model fit.

[0082] 4.6 Optimization Algorithm

[0083] The goal of a model algorithm is to obtain optimal model parameters through training. To find this optimal set of parameters, a loss function is typically used as the parameter estimation function. Through sufficient iterations, the loss value is continuously reduced to obtain a better weight parameter solution. Therefore, error optimization algorithms are introduced into model training to continuously minimize the loss value to approach or reach the optimal value. During the neural network training process, the optimization method selected to find hyperparameters has a significant impact on the model's performance. Common optimization algorithms can be summarized in four steps:

[0084] (1) Calculate the current gradient of the objective function:

[0085] (2) Calculate the first-order momentum and second-order momentum: V t =γ(g1,g2,…,g t )

[0086] (3) Calculate the gradient at the current moment:

[0087] (4) Update target value: ω t+1 =ω t -η t

[0088] Where ωt is the parameter to be optimized; α is the learning rate; and f(ωt) is the objective function. This study can use the modified Adam optimization algorithm, also known as the adaptive algorithm. This algorithm can calculate the adaptive learning rate for each parameter and introduces first-order momentum and second-order momentum. When the parameters vary, the algorithm automatically adjusts the learning rate using first-order or second-order moment estimates.

[0089] 4.6.1 SGB

[0090] Stochastic Gradient Descent (SGD) does not introduce the concept of momentum in the process of model optimization, which results in its slow convergence speed and easy to fall into local optimal values.

[0091] 4.6.2 Adam

[0092] Adaptive Moment Estimation (Adam) is another common neural network optimizer, which introduces first-order momentum and second-order momentum, namely: mt = β1·mt-1+(1-β1)·gtVt = β2·Vt-1+(1-β2)·gt 2 The Adam algorithm can make the model converge faster, and the parameter update process is relatively smooth, which is more suitable for training deeper and more complex networks.

[0093] 4.6.3 Nadam

[0094] Nesterov Accelerated Gradient (NAG) is an improvement to the first-order momentum: during the update process, the current gradient is corrected using the accumulated gradient. The gradient is defined as follows:

[0095] NAG utilizes the second-order derivative information of the objective function to make the hyperparameters change over time and ensure that the learning rate changes steadily.

[0096] 5. CNN-LSTM-based neural network model - overall model framework

[0097] Deep learning-based models are capable of automatically extracting features. CNN models typically rely on convolutional kernels in convolutional layers to extract features, and then rely on pooling layers to refine features, preserving important data information. However, the presence of convolutional kernels limits CNNs' ability to process time series signals, which presents a "long-term dependency problem." Therefore, the deep learning model LSTM was introduced to address this issue. The model combines CNN and LSTM to ensure that deep spatial and temporal features of the input source are effectively extracted.

[0098] The turbidity, water volume, operating parameters, and process parameters used in this study are multidimensional time series signals. When performing regression prediction, it is usually necessary to consider the numerical impact of consecutive time periods and the relationship between various variables. Therefore, this study combines the CNN and LSTM models to complete the prediction of emission concentration. The overall architecture of the hybrid model is shown in Figure 5.

[0099] According to the structure diagram, the CNN first adaptively extracts local water quality features from data such as turbidity, water volume, and operating parameters. Taking into account the temporal correlation between turbidity and mud level, the LSTM performs secondary feature extraction on the features extracted by the CNN to obtain temporal information about the features. A fully connected layer then reduces the feature dimensionality to the same number of target results, enabling prediction of mud level and scraper activation time. The model's predictive performance is then evaluated using selected metrics.

[0100] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A sedimentation tank scraper energy-saving control method based on deep neural network, characterized in that: The following steps are involved: Obtain turbidity information, operating parameters and temperature information in the sedimentation tank, and input the obtained turbidity information, operating parameters and temperature information into a pre-built and trained deep learning neural network model in time sequence to obtain the prediction results of the mud level and the start time of the scraper, so as to control the opening and closing of the scraper; The deep learning neural network model includes a convolutional neural network sub-model and a long short-term memory network sub-model connected in sequence. The convolutional neural network sub-model is used to adaptively extract local features of water quality based on received turbidity information, operating parameters and temperature information. The long short-term memory network sub-model is used to perform secondary feature extraction on the local features of water quality extracted by the convolutional neural network sub-model, obtain the time series information of the features, and finally output the prediction results.

2. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: The deep learning neural network model finally reduces the feature dimension of the output of the long short-term memory network sub-model through a fully connected layer and outputs the prediction result.

3. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: During the training process of the deep learning neural network model, iterative optimization is performed through stochastic gradient descent, adaptive momentum estimation and / or Nesteev accelerated gradient.

4. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: During the training process of the deep learning neural network model, mean absolute error, root mean square error, and goodness of fit indicators are used to verify the predictive performance of the model.

5. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: The training data input into the deep learning neural network model during the training process includes turbidity information, operating parameters, temperature information and mud level information. The mud level information is obtained through mud level sensors arranged in the sedimentation tank. The mud level sensors are arranged at the head, middle and tail of the sedimentation tank.

6. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: The process of obtaining the turbidity information, operating parameters and temperature information in the sedimentation tank includes: A temperature sensor, a turbidity sensor and a flow sensor are respectively arranged in the sedimentation tank to obtain temperature information, turbidity information and flow information.

7. The energy-saving control method for a sedimentation tank scraper based on a deep neural network according to claim 6 is characterized in that: The temperature sensor is arranged at the water inlet of the sedimentation tank, and the flow sensor is arranged at the water pipe of the sedimentation tank.

8. The energy-saving control method for a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: The scraper is connected to a PLC system, and the opening and closing of the scraper is controlled by the PLC system.

9. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: After obtaining the turbidity information, operating parameters and temperature information, each piece of information is first time synchronized, and then the time synchronized information is input into the deep learning neural network model.

10. The energy-saving control method of a sedimentation tank scraper based on a deep neural network according to claim 1 is characterized in that: The information input into the deep learning neural network model is a multi-dimensional time series signal based on turbidity information, operating parameters and temperature information.

Images