Method and system for dynamically predicting dehydration efficiency of river and lake bottom mud and automatically adjusting parameters

Abstract

The invention provides a river and lake sediment dehydration efficiency dynamic prediction and parameter self-adjustment method and system. The method comprises the following steps: step 1, collecting a training sample set; step 2, using an LSTM neural network to build a prediction model, using the sludge initial moisture content of the historical dehydration parameter data in the training sample set, the PAC dosage and the feeding pressure as input characteristics, and using the first standard reaching time when the sludge cake moisture content of the historical dehydration parameter data in the training sample set is less than or equal to a first threshold value as an output target, training an efficiency prediction model; and step 3, based on the efficiency prediction model, predicting second standard time according to the real-time dehydration parameter data, and if the second standard time is greater than a second threshold value, automatically adjusting the real-time dehydration parameter data through a gradient descent algorithm until the predicted new second standard time returns to the second threshold value. According to the river and lake sediment dehydration efficiency dynamic prediction and parameter self-adjustment method and system, manual intervention is not needed, and the dehydration standard reaching rate and stability are higher.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and control technology, and in particular to a method and system for dynamic prediction and parameter self-adjustment of dewatering efficiency of river and lake sediments. Background Technology

[0002] With the acceleration of urbanization and the increasing urgency of water environment management, the dredging and harmless disposal of river and lake sediments have become a crucial link in water environment restoration projects. Sediment dewatering, as a core pre-treatment step in sludge disposal, directly affects the feasibility of subsequent resource utilization and disposal costs. Currently, mainstream dewatering processes mostly employ mechanical filter pressing or centrifugal dewatering, supplemented by flocculants such as polyaluminum chloride (PAC) to improve dewatering performance. However, these dewatering processes lack dynamic predictability. Traditional control methods rely on operator experience to set PAC dosage and feed pressure, or operate with fixed parameters, failing to consider the spatiotemporal fluctuations in the initial moisture content of the sediment. Due to the complex composition and large fluctuations in moisture content of river and lake sediments, the dewatering cycle is unstable, and the moisture content of the sludge cake is prone to exceeding the standard or over-dewatering, resulting in significant fluctuations in the compliance rate.

[0003] In view of this, there is an urgent need for a method and system for dynamic prediction of river and lake sediment dewatering efficiency and parameter self-adjustment, in order to at least address the above-mentioned shortcomings. Summary of the Invention

[0004] One objective of this invention is to provide a method and system for dynamic prediction and parameter self-adjustment of dewatering efficiency for river and lake sediments. It introduces a prediction model based on an LSTM neural network, training the model using historical dewatering parameter data. Based on this model, the real-time dewatering parameters are automatically adjusted using a gradient descent algorithm, ensuring the predicted time to achieve the target is within the specified requirements, without manual intervention, thus improving the dewatering success rate and stability.

[0005] The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency provided in this invention includes:

[0006] Step 1: Collect historical dehydration parameter data to form a training sample set;

[0007] Step 2: Use an LSTM neural network to build a prediction model. Use the initial moisture content of sludge, PAC dosage and feed pressure of historical dewatering parameter data in the training sample set as input features, and the first time when the moisture content of sludge cake is less than or equal to the preset first threshold as the output target to train the efficiency prediction model.

[0008] Step 3: Based on the efficiency prediction model, predict the second time to reach the target based on the real-time dehydration parameter data. If the second time to reach the target is greater than the preset second threshold, automatically adjust the real-time dehydration parameter data through the gradient descent algorithm until the predicted new second time to reach the target returns to within the second threshold.

[0009] Preferably, step 1: Collect historical dehydration parameter data to form a training sample set, including:

[0010] Historical operational data is automatically collected from industrial automation systems in wastewater treatment plants and / or river and lake sediment disposal centers.

[0011] Based on preset data cleaning rules, historical operational data is cleaned to obtain historical dehydration parameter data. The preset data cleaning rules are: to remove invalid dehydration cycle data and abnormal physical value data.

[0012] Based on each effective dehydration cycle, historical dehydration parameter data are used to form independent training samples, which are then divided into training set, test set, and validation set according to a preset division ratio.

[0013] Preferably, the real-time dehydration parameter data is automatically adjusted using a gradient descent algorithm until the predicted new second target time returns to within the second threshold, including:

[0014] Based on real-time dehydration parameter data and the second time to achieve the target, a target loss function is constructed. The target loss function is as follows:

[0015]

[0016] in, The target loss function is the ReLU function. Output when greater than 0 Output 0 when less than or equal to 0; This is the input feature vector extracted from real-time dehydration parameter data; The second time to achieve the target, as predicted by the efficiency prediction model; T lim The second threshold;

[0017] Calculate the gradient of the target loss function And determine the direction of real-time dehydration parameter adjustment;

[0018] Based on the gradient of the target loss function and the learning rate, gradient descent is applied to update the input feature vector. The updated input feature vector is as follows:

[0019]

[0020] in, The updated input feature vector. η is the input feature vector extracted from the real-time dehydration parameter data at the current moment, and η is the learning rate.

[0021] The real-time dehydration parameters are automatically adjusted based on the updated input feature vector.

[0022] The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency provided in this embodiment of the invention further includes:

[0023] After determining the direction of adjustment of real-time dehydration parameters, the abnormality is attributed by comparing the direction of adjustment of real-time dehydration parameters with the direction of adjustment of standard dehydration parameters.

[0024] Preferably, after determining the direction of real-time dehydration parameter adjustment, anomaly attribution is performed by comparing the real-time dehydration parameter adjustment direction with the standard dehydration parameter adjustment direction, including:

[0025] Adjust the direction based on real-time dehydration parameters and construct a real-time adjustment direction vector;

[0026] Construct a knowledge base for adjusting standard dehydration parameters;

[0027] Compare the real-time adjustment direction vector with the standard adjustment direction vector in the standard dehydration parameter adjustment direction knowledge base to obtain the comparison results;

[0028] Based on the comparison results and the anomaly cause database, determine the corresponding anomaly causes and handling recommendations, and respond to the handling recommendations.

[0029] Preferably, a knowledge base for adjusting standard dehydration parameters is constructed, including:

[0030] Normal dehydration parameter data were selected from historical dehydration parameter data;

[0031] Small perturbations were applied to the normal dehydration parameter data, and changes in the predicted time to reach the target were observed based on the efficiency prediction model;

[0032] Based on the applied small perturbation and the change in the time to reach the target, calculate the gradient direction under each perturbation and determine the gradient direction vector;

[0033] The gradient direction vectors corresponding to the normal dehydration parameter data are aggregated to form a standard directional probability distribution.

[0034] Cluster analysis is performed based on the standard directional probability distribution to extract typical adjustment patterns and their corresponding standard adjustment direction vectors to construct a knowledge base for standard dehydration parameter adjustment directions.

[0035] Preferably, the steps for determining the learning rate include:

[0036] The efficiency prediction model is embedded into the digital twin control system, and closed-loop control is performed by simulating the real dehydration process based on historical dehydration parameter data.

[0037] In the closed-loop control process, candidate learning rates from a preset candidate learning rate sequence are randomly applied;

[0038] An evaluation distribution is generated based on the candidate learning rate being applied, and the evaluation distribution is mapped onto the closed-loop control scenario corresponding to the candidate learning rate. The evaluation distribution includes: the convergence speed of the closed-loop control, parameter volatility, and safety compliance rate.

[0039] Based on the path of the scene elements of interest to the experimenters in the closed-loop control scenario, determine the development trend of the evaluation distribution of the current mapping;

[0040] Based on the development trend, determine the traversal direction of the candidate learning rate sequence;

[0041] Based on the traversal direction, the application steps of the candidate learning rate are repeated to determine the source candidate learning rate whose trend increment of the corresponding evaluation distribution is less than the preset third threshold as the final determined learning rate.

[0042] Preferably, based on the paths of the scene elements of interest to the experimenters in the closed-loop control scenario, the development trend of the evaluation distribution of the current mapping is determined, including:

[0043] Based on the historical adjustment patterns of performance metrics, determine the system log records prior to the adjustment;

[0044] Determine the data extraction dimensions from the system logs;

[0045] If the data extraction dimensions match the visual data extraction dimensions, determine the timestamp, visual subject, and visual data features of the corresponding record data;

[0046] Visual data features are marked on the visual subject and sorted according to the time sequence of the timestamp records to obtain the standard scene element path. The standard scene element path is then associated with the performance index trend corresponding to the historical adjustment situation.

[0047] We will focus on matching scene element paths with standard scene element paths. If the scene element paths match, we will summarize the corresponding performance index trends to obtain the development trend of the evaluation distribution.

[0048] The river and lake sediment dewatering efficiency dynamic prediction and parameter self-adjustment system provided in this embodiment of the invention includes:

[0049] The sample acquisition module is used to collect historical dehydration parameter data to form a training sample set.

[0050] The model training module is used to build a prediction model using an LSTM neural network. It takes the initial moisture content of sludge, PAC dosage and feed pressure of historical dewatering parameter data in the training sample set as input features, and the first time when the moisture content of sludge cake is less than or equal to the preset first threshold as the output target to train the efficiency prediction model.

[0051] The parameter self-adjustment module is used to predict the second target time based on the efficiency prediction model and real-time dehydration parameter data. If the second target time is greater than the preset second threshold, the real-time dehydration parameter data is automatically adjusted through the gradient descent algorithm until the predicted new second target time returns to within the second threshold.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention introduces a prediction model based on an LSTM neural network, which is trained using historical dehydration parameter data. Based on the efficiency prediction model, the real-time dehydration parameter data is automatically adjusted through a gradient descent algorithm, ensuring that the predicted time to achieve the target meets the specified requirements without manual intervention, thereby improving the dehydration success rate and dehydration stability.

[0054] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a schematic diagram of the dynamic prediction and parameter self-adjustment method for dewatering efficiency of river and lake sediments in an embodiment of the present invention;

[0058] Figure 2 This is a schematic diagram of the dynamic prediction and parameter self-adjustment system for river and lake sediment dewatering efficiency in an embodiment of the present invention. Detailed Implementation

[0059] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0060] This invention provides a method for dynamic prediction and parameter self-adjustment of dewatering efficiency of river and lake sediments, such as... Figure 1 As shown, it includes:

[0061] Step 1: Collect historical dehydration parameter data to form a training sample set.

[0062] In this embodiment, the historical dewatering parameter data refers to various operational parameters (e.g., initial sludge moisture content, PAC dosage, and feed pressure) and result data (e.g., cake moisture content and time to reach the target) recorded historically during the sludge dewatering process. The training sample set is a dataset compiled from the historical dewatering parameter data and used to train the machine learning model. Step 1 specifically includes:

[0063] Step 11: Automatically collect historical operating data from the industrial automation systems of wastewater treatment plants and / or river and lake sediment disposal centers;

[0064] Step 12: Clean the historical operation data based on the preset data cleaning rules to obtain historical dehydration parameter data; the preset data cleaning rules are: remove invalid dehydration cycle data and abnormal physical value data;

[0065] Step 13: Based on each effective dehydration cycle, construct independent training samples from historical dehydration parameter data, and divide the independent training samples into training set, test set and validation set according to a preset division ratio.

[0066] Step 2: Build a prediction model using an LSTM neural network. Use the initial moisture content of sludge, PAC dosage, and feed pressure from the historical dewatering parameters in the training sample set as input features, and the first time when the moisture content of the sludge cake is less than or equal to a preset first threshold as the output target to train the efficiency prediction model.

[0067] In this embodiment, the first threshold is preset manually, for example, 45%. The first time to reach the target is the time taken from the start of dehydration until the moisture content of the cake first drops below the first threshold. The efficiency prediction model is a mathematical model that predicts the time required for the moisture content of the cake to reach less than or equal to the first threshold based on the input characteristics (initial moisture content, PAC amount, feed pressure).

[0068] Step 3: Based on the efficiency prediction model, predict the second time to reach the target based on the real-time dehydration parameter data. If the second time to reach the target is greater than the preset second threshold, automatically adjust the real-time dehydration parameter data through the gradient descent algorithm until the predicted new second time to reach the target returns to within the second threshold.

[0069] In this embodiment, the real-time dewatering parameter data refers to the various operational parameters recorded during the real-time sludge dewatering process. The second attainment time is the time, predicted by the efficiency prediction model, required to achieve a cake moisture content ≤ the first threshold under the current real-time operating parameters. The second threshold is preset manually, for example, 2 hours. The real-time dewatering parameter data is automatically adjusted using a gradient descent algorithm until the predicted new second attainment time returns to within the second threshold: by calculating the gradient (i.e., sensitivity) of the second attainment time with respect to input parameters (PAC quantity, pressure, etc.), the real-time dewatering parameter data is adjusted in reverse to reduce the predicted time to within the second threshold.

[0070] The working principle and beneficial effects of the above technical solution are as follows:

[0071] This invention introduces a prediction model based on an LSTM neural network, which is trained using historical dehydration parameter data. Based on the efficiency prediction model, the real-time dehydration parameter data is automatically adjusted through a gradient descent algorithm, ensuring that the predicted time to achieve the target meets the specified requirements without manual intervention, thereby improving the dehydration success rate and dehydration stability.

[0072] In one embodiment, the real-time dehydration parameter data is automatically adjusted using a gradient descent algorithm until the predicted new second target time returns to within the second threshold, including:

[0073] Step 31: Based on real-time dehydration parameter data and the second time to achieve the target, construct the target loss function, which is as follows:

[0074]

[0075] in, The target loss function is the ReLU function. Output when greater than 0 Output 0 when less than or equal to 0; This is the input feature vector extracted from real-time dehydration parameter data; The second time to achieve the target, as predicted by the efficiency prediction model; T lim This is the second threshold.

[0076] In this embodiment, the input feature vector is a numerical vector constructed based on the initial moisture content of the sludge, the amount of PAC added, and the feed pressure extracted from real-time dewatering parameter data.

[0077] Step 32: Calculate the gradient of the target loss function And determine the direction of real-time dehydration parameter adjustment.

[0078] In this embodiment, the direction of real-time dehydration parameter adjustment is the direction of parameter adjustment strategy derived from the gradient of the target loss function. If the gradient is greater than 0, increasing the parameter will prolong the time to reach the target, so the corresponding parameter should be decreased, with the direction being -1. If the gradient is less than 0, increasing the parameter will shorten the time to reach the target, so the corresponding parameter should be increased, with the direction being +1. If the gradient is approximately equal to 0, changing the parameter will have no significant impact on the time to reach the target, with the direction being 0.

[0079] Step 33: Based on the gradient of the target loss function and the learning rate, apply gradient descent to update the input feature vector. The updated input feature vector is as follows:

[0080]

[0081] in, The updated input feature vector. η is the input feature vector extracted from the real-time dehydration parameter data at the current moment, and η is the learning rate.

[0082] In this embodiment, the learning rate is used to control the step size of each parameter adjustment, and is determined based on simulation optimization, for example: 0.05.

[0083] Step 34: Determine the automatically adjusted real-time dehydration parameter data based on the updated input feature vector.

[0084] The working principle and beneficial effects of the above technical solution are as follows:

[0085] This invention constructs a target loss function based on real-time dehydration parameter data and the second time to achieve the target. When the target loss function is greater than 0, it uses automatic differentiation technology to backpropagate the efficiency prediction model and calculate the gradient of the target loss function. This gradient vector precisely indicates the sensitivity weight of each parameter to dehydration efficiency. Parameters are updated in the normalized space, with continuous, micro-step, and optimal adjustments along the direction of fastest loss reduction to determine the updated input feature vector. This vector is then converted into a true physical quantity through inverse normalization. Each adjustment converges towards the target faster, resulting in higher efficiency for automatic parameter adjustment.

[0086] In one embodiment, automatically adjusting real-time dehydration parameter data using a gradient descent algorithm until the predicted new second target time returns to within the second threshold also includes:

[0087] Step 321: After determining the direction of real-time dehydration parameter adjustment, compare the direction of real-time dehydration parameter adjustment with the direction of standard dehydration parameter adjustment to attribute the cause of the anomaly, including:

[0088] Step 3211: Adjust the direction according to the real-time dehydration parameters and construct a real-time adjustment direction vector.

[0089] In this embodiment, the real-time adjustment direction vector is a multi-dimensional binary vector formed by processing the real-time dehydration parameter adjustment direction with a sign function. It is used to standardize the expression of the system's current suggested parameter adjustment intention, such as [0,+1,-1], which represents increasing PAC and decreasing pressure.

[0090] Step 3212: Construct a knowledge base for adjusting standard dehydration parameters.

[0091] In this embodiment, the standard dehydration parameter adjustment direction knowledge base is a set of reference patterns trained based on historical normal operation data, representing typical reasonable adjustment behaviors. Specifically, step 3212 includes:

[0092] Step 32121: Filter out normal dehydration parameter data from historical dehydration parameter data.

[0093] In this embodiment, the normal dehydration parameter data is a high-quality sample subset selected from historical data that meets the conditions for normal process operation. It represents the typical operating state of the system under fault-free, anomaly-free, and stable conditions. For example, historical dehydration parameter data with a historical compliance time of less than the second threshold.

[0094] Step 32122: Apply small perturbations to the normal dehydration parameter data and observe the changes in the predicted time to reach the target based on the efficiency prediction model.

[0095] In this embodiment, applying a small perturbation means systematically and controllably applying tiny, known, and unnatural numerical perturbations to the input parameters (PAC, pressure) based on normal dehydration parameter data. For example, increasing the perturbation by ±0.02 normalized units corresponds to a physical quantity of approximately ±0.15 kg / m³. 3 (PAC) and ±0.05MPa (pressure). Observe the changes in predicted achievement time: For each perturbation sample, input the efficiency prediction model, obtain its output predicted achievement time, and compare it with the achievement time of the original sample to quantify the impact of parameter fine-tuning on the achievement time, providing numerical basis for subsequent gradient calculation.

[0096] Step 32123: Based on the applied small perturbation and the change in the time to reach the target, calculate the gradient direction under each perturbation and determine the gradient direction vector.

[0097] In this embodiment, when calculating the gradient direction under each perturbation, the local gradient direction of the model at the original sample points is approximated using the finite difference method based on the change in the predicted achievement time caused by the perturbation. When determining the gradient direction vector, the aforementioned local gradient direction is converted into a binary symbol vector to obtain the gradient direction vector.

[0098] Step 32124: Aggregate the gradient direction vectors corresponding to the normal dehydration parameter data to form a standard directional probability distribution.

[0099] In this embodiment, the gradient direction vectors calculated from each of the normal dehydration parameter samples are aggregated into a unified statistical distribution to reveal the adjustment mode most frequently recommended by the system under normal operating conditions.

[0100] Step 32125: Perform cluster analysis based on the standard direction probability distribution, extract typical adjustment patterns and their corresponding standard adjustment direction vectors to construct a standard dehydration parameter adjustment direction knowledge base.

[0101] In this embodiment, each cluster center in the standard orientation probability distribution is used as a standard adjustment orientation vector, and its confidence level is calculated. The confidence level is determined based on the average Hamming distance between all samples in the cluster and the cluster center. Cluster centers with a confidence level of not less than 0.8 are retained to construct a standard dehydration parameter adjustment orientation knowledge base. A typical adjustment pattern is the adjustment pattern described by the standard adjustment orientation vector. Table 1 shows an example of a standard dehydration parameter adjustment orientation knowledge base.

[0102] Table 1 Examples of the knowledge base on the adjustment direction of standard dehydration parameters.

[0103]

[0104] Step 3213: Compare the real-time adjustment direction vector with the standard adjustment direction vector in the standard dehydration parameter adjustment direction knowledge base, and obtain the comparison results.

[0105] In this embodiment, the comparison result is the matching result between the real-time adjustment direction vector and the standard adjustment direction vector in the standard dehydration parameter adjustment direction knowledge base, including: a successfully matched standard adjustment direction vector or no successfully matched standard adjustment direction vector.

[0106] Step 3214: Based on the comparison results and the anomaly cause database, determine the corresponding anomaly causes and handling suggestions, and respond to the handling suggestions.

[0107] In this embodiment, the abnormal cause library consists of real process or equipment abnormal causes corresponding to adjustment modes other than the standard adjustment direction vector summarized by experts. It also includes expert operation suggestions or control action instructions for the corresponding abnormalities, which are used to guide operation and maintenance personnel or automatic control systems to take the next step (handling suggestions).

[0108] The working principle and beneficial effects of the above technical solution are as follows:

[0109] After determining the adjustment direction of real-time dewatering parameters, there are situations where the adjustment behavior does not conform to typical reasonable adjustment behavior, such as: not adjusting the moisture content (uncontrollable), reducing the PAC dosage, or increasing the feed pressure. When the above-mentioned non-typical reasonable adjustment behavior occurs, it often implicitly indicates an abnormal situation (e.g., AI misadjustment caused by sensor drift). If the adjustment continues in the incorrect direction, it will lead to a decrease in dewatering efficiency. Therefore, this embodiment of the invention constructs a knowledge base for standard dewatering parameter adjustment directions.

[0110] In the specific construction process, normal dewatering parameter data that meets the conditions for normal process operation is selected from historical dewatering parameter data (e.g., historical compliance time is less than the second threshold). Based on the normal dewatering parameter data and efficiency prediction model, small perturbations are artificially applied, and gradient direction vectors are determined based on the applied perturbations and changes in compliance time. The gradient direction vectors corresponding to the normal dewatering parameter data are aggregated to form a standard direction probability distribution. Cluster analysis is performed based on the standard direction probability distribution to extract typical adjustment patterns and their corresponding standard adjustment direction vectors to construct a standard dewatering parameter adjustment direction knowledge base. Selecting reference patterns of typical and reasonable adjustment behaviors from reliable normal dewatering parameter data improves the standardization of the construction of the standard dewatering parameter adjustment direction knowledge base.

[0111] Next, based on the comparison results between the real-time adjustment direction vector and the standard adjustment direction vector in the standard dehydration parameter adjustment direction knowledge base, it is determined whether the real-time adjustment mode is not a standard mode (there is no successfully matched standard adjustment direction vector). If so, it connects to the anomaly cause database and takes corresponding actions based on the handling suggestions of experts in the anomaly cause database (for example, if the anomaly cause is determined to be sensor drift, the handling suggestion is to perform sensor calibration), which further improves the dehydration efficiency.

[0112] In one embodiment, the step of determining the learning rate includes:

[0113] Step 331: Embed the efficiency prediction model into the digital twin control system and perform closed-loop control by simulating the real dehydration process based on historical dehydration parameter data.

[0114] In this embodiment, the digital twin control system is a high-fidelity, real-time synchronous, two-way interactive virtual system built based on a dewatering device for sediment dewatering. Closed-loop control is performed by simulating the actual dewatering process using historical dewatering parameter data: based on an efficiency prediction model, the historical dewatering parameter data is automatically adjusted using a gradient descent algorithm to bring the predicted time to achieve the target back to within a second threshold.

[0115] Step 332: In the closed-loop control process, randomly apply the candidate learning rate from the preset candidate learning rate sequence.

[0116] In this embodiment, the preset candidate learning rate sequence is manually pre-set, for example: [0.001, 0.005, 0.01, 0.02, 0.05, 0.1].

[0117] Step 333: Generate an evaluation distribution based on the candidate learning rate being applied, and map the evaluation distribution onto the closed-loop control scenario corresponding to the candidate learning rate. The evaluation distribution includes: the convergence speed of the closed-loop control, parameter volatility, and safety compliance rate.

[0118] In this embodiment, the evaluation distribution is the statistical distribution of performance metrics corresponding to candidate learning rates. These performance metrics include: convergence speed of closed-loop control, parameter volatility, and safety compliance rate. The closed-loop control scenario is a dynamic dehydrated twin scenario with corresponding candidate learning rates set.

[0119] Step 334: Based on the path of the scene elements of interest to the experimenters in the closed-loop control scenario, determine the development trend of the evaluation distribution of the current mapping.

[0120] In this embodiment, the path of interest for scene elements is: the viewing path of scene elements obtained by sorting the scene elements that the experimenter is interested in in the twin closed-loop control scenario according to the order in which they are interested. For example, the path of interest for scene elements is: the identifier elements corresponding to the feed pressure values ​​arranged in order of decreasing order; the identifier elements corresponding to the system time arranged in order of increasing order. The development trend is: the optimization trend of each performance index that the experimenter prefers based on the path of interest for scene elements. For example, if the path of interest for scene elements is the identifier elements corresponding to the feed pressure values ​​arranged in order of decreasing order, it indicates that the experimenter is interested in parameter fluctuations and prefers to reduce parameter fluctuations; or if the path of interest for scene elements is the identifier elements corresponding to the system time arranged in order of increasing order, it indicates that the experimenter is interested in the convergence speed of the closed-loop control and prefers to accelerate the convergence speed. Specifically, step 334 includes:

[0121] Step 3341: Based on the historical adjustment of performance indicators, determine the system log records before the adjustment.

[0122] In this embodiment, the historical adjustment of performance indicators refers to the situation in which the convergence speed, parameter volatility, and safety compliance rate of closed-loop control are adjusted manually in the past. The system log is the running data stream that the system automatically collects and stores for a period of time (e.g., 5 minutes) before each adjustment occurs.

[0123] Step 3342: Determine the data extraction dimensions for the data recorded in the system log.

[0124] In this embodiment, the data extraction dimension is the data extraction method, such as: system monitoring, manual reading.

[0125] Step 3343: If the data extraction dimension matches the visual data extraction dimension, determine the timestamp, visual subject, and visual data features of the corresponding record data.

[0126] In this embodiment, the visual data extraction dimension is the data extraction dimension obtained by manual viewing. The timestamp of the recorded data is the recording time of the recorded data. The visual subject is the visual carrier of the visual data, such as the identifier frame of a certain identifier ID. The visual data characteristics are the data content patterns of the visual data, such as: variance greater than a preset variance threshold, an increasing trend, or a decreasing trend.

[0127] Step 3344: Mark the visual data features on the visual subject and sort them according to the time sequence of the timestamp records to obtain the standard scene element path, and associate the standard scene element path with the performance index trend corresponding to the historical adjustment situation.

[0128] In this embodiment, the performance indicator trend refers to the type and direction of performance indicator adjustments made in historical adjustment scenarios.

[0129] Step 3345: Match the scene element paths of the focus scene with the standard scene element paths. If the scene element paths match, summarize the corresponding performance index trends to obtain the development trend of the evaluation distribution.

[0130] In this embodiment, scene element path matching is defined as: whether the visible subjects are consistent and whether the feature matching degree of the visible data features (i.e., the similarity of data content patterns) is greater than the preset data content pattern similarity threshold.

[0131] Step 335: Determine the traversal direction of the candidate learning rate sequence based on the development trend.

[0132] In this embodiment, when the development trend tends to increase the convergence speed of closed-loop control, the traversal direction is to traverse in the direction of increasing candidate learning rate; when the development trend tends to decrease parameter volatility and / or increase safety compliance rate, the traversal direction is to traverse in the direction of decreasing candidate learning rate.

[0133] Step 336: Based on the traversal direction, repeat the application steps of the candidate learning rate to determine the source candidate learning rate whose trend increment of the corresponding evaluation distribution is less than the preset third threshold as the final determined learning rate.

[0134] In this embodiment, the application steps of the candidate learning rate are: steps 333-336. A trend increment less than the preset third threshold indicates that the evaluation distribution is in a steady state, and the experimenters are satisfied with the current model performance.

[0135] The working principle and beneficial effects of the above technical solution are as follows:

[0136] When performing gradient descent based on the gradient of the target loss function and the learning rate, the setting of the learning rate significantly impacts self-adjusting performance. A learning rate that is too high may lead to large parameter fluctuations and parameters exceeding safety compliance limits; a learning rate that is too low may result in slow convergence of closed-loop control and low adjustment efficiency. Therefore, this embodiment of the invention introduces a preset candidate learning rate sequence. A candidate learning rate is randomly selected from the sequence for Siamese simulation, and an evaluation distribution is generated based on the candidate learning rate being applied. During Siamese simulation, the experimenter will view the parameter self-adjusting closed-loop control scenario. This viewing will generate paths for elements of interest in the scenario, implicitly reflecting the experimenter's intention to adjust the performance indicators of the current closed-loop control scenario (the development trend of the currently mapped evaluation distribution).

[0137] Specifically, when determining the development trend of the current mapping evaluation distribution based on the path of the focus scene elements in the closed-loop control scenario, the system logs before historical performance index adjustments are identified. The timestamps, visual subjects, and visual data features of the recorded data are extracted from these logs. Visual data features are labeled on the visual subjects and sorted chronologically to obtain standard scene element paths. These standard scene element paths are then correlated with the performance index trends corresponding to historical adjustment scenarios. The focus scene element paths and standard scene element paths are matched, and the corresponding performance index trends that match are summarized to obtain the development trend of the evaluation distribution.

[0138] Next, based on the development trend, the adjustment direction of the candidate learning rate is further determined, and the application steps of the candidate learning rate are repeated until the evaluation distribution reaches a steady state. The best candidate learning rate is selected as the final learning rate, which realizes the synergy of parameter self-adjustment efficiency and stability, and the parameter self-adjustment process is more reasonable.

[0139] This invention provides a system for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency, such as... Figure 2 As shown, it includes:

[0140] Sample acquisition module 1 is used to collect historical dehydration parameter data to form a training sample set;

[0141] Model training module 2 is used to build a prediction model using an LSTM neural network. It takes the initial moisture content of sludge, the amount of PAC added, and the feed pressure of historical dewatering parameters in the training sample set as input features, and takes the first time when the moisture content of the sludge cake is less than or equal to the preset first threshold as the output target to train the efficiency prediction model.

[0142] The parameter self-adjustment module 3 is used to predict the second time to achieve the target based on the efficiency prediction model and real-time dehydration parameter data. If the second time to achieve the target is greater than the preset second threshold, the real-time dehydration parameter data is automatically adjusted through the gradient descent algorithm until the predicted new second time to achieve the target returns to the second threshold.

[0143] Among them, the parameter self-adjustment module 3 automatically adjusts the real-time dehydration parameter data through the gradient descent algorithm until the predicted new second target time returns to within the second threshold, including:

[0144] Based on real-time dehydration parameter data and the second time to achieve the target, a target loss function is constructed;

[0145] Calculate the gradient of the target loss function and determine the direction of real-time dehydration parameter adjustment;

[0146] Based on the gradient of the target loss function and the learning rate, gradient descent is applied to update the input feature vector;

[0147] The real-time dehydration parameters are automatically adjusted based on the updated input feature vector.

[0148] After determining the direction of real-time dehydration parameter adjustment, anomaly attribution is performed by comparing the real-time dehydration parameter adjustment direction with the standard dehydration parameter adjustment direction, including:

[0149] Adjust the direction based on real-time dehydration parameters and construct a real-time adjustment direction vector;

[0150] Construct a knowledge base for adjusting standard dehydration parameters;

[0151] Compare the real-time adjustment direction vector with the standard adjustment direction vector in the standard dehydration parameter adjustment direction knowledge base to obtain the comparison results;

[0152] Based on the comparison results and the anomaly cause database, determine the corresponding anomaly causes and handling recommendations, and respond to the handling recommendations;

[0153] This includes constructing a knowledge base for adjusting standard dehydration parameters, including:

[0154] Normal dehydration parameter data were selected from historical dehydration parameter data;

[0155] Small perturbations were applied to the normal dehydration parameter data, and changes in the predicted time to reach the target were observed based on the efficiency prediction model;

[0156] Based on the applied small perturbation and the change in the time to reach the target, calculate the gradient direction under each perturbation and determine the gradient direction vector;

[0157] The gradient direction vectors corresponding to the normal dehydration parameter data are aggregated to form a standard directional probability distribution.

[0158] Cluster analysis is performed based on the standard direction probability distribution to extract typical adjustment patterns and their corresponding standard adjustment direction vectors to construct a knowledge base for the adjustment direction of standard dehydration parameters.

[0159] The steps for determining the learning rate include:

[0160] The efficiency prediction model is embedded into the digital twin control system, and closed-loop control is performed by simulating the real dehydration process based on historical dehydration parameter data.

[0161] In the closed-loop control process, candidate learning rates from a preset candidate learning rate sequence are randomly applied;

[0162] An evaluation distribution is generated based on the candidate learning rate being applied, and the evaluation distribution is mapped onto the closed-loop control scenario corresponding to the candidate learning rate. The evaluation distribution includes: the convergence speed of the closed-loop control, parameter volatility, and safety compliance rate.

[0163] Based on the path of the scene elements of interest to the experimenters in the closed-loop control scenario, determine the development trend of the evaluation distribution of the current mapping;

[0164] Based on the development trend, determine the traversal direction of the candidate learning rate sequence;

[0165] Based on the traversal direction, the application steps of the candidate learning rate are repeated to determine the source candidate learning rate whose trend increment of the corresponding evaluation distribution is less than the preset third threshold as the final determined learning rate.

[0166] Among them, based on the paths of the scene elements of interest to the experimenters in the closed-loop control scenario, the development trend of the evaluation distribution of the current mapping is determined, including:

[0167] Based on the historical adjustment patterns of performance metrics, determine the system log records prior to the adjustment;

[0168] Determine the data extraction dimensions from the system logs;

[0169] If the data extraction dimensions match the visual data extraction dimensions, determine the timestamp, visual subject, and visual data features of the corresponding record data;

[0170] Visual data features are marked on the visual subject and sorted according to the time sequence of the timestamp records to obtain the standard scene element path. The standard scene element path is then associated with the performance index trend corresponding to the historical adjustment situation.

[0171] We will focus on matching scene element paths with standard scene element paths. If the scene element paths match, we will summarize the corresponding performance index trends to obtain the development trend of the evaluation distribution.

[0172] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for dynamic prediction and parameter self-adjustment of dewatering efficiency of river and lake sediments, characterized in that, include: Step 1: Collect historical dehydration parameter data to form a training sample set; Step 2: Use an LSTM neural network to build a prediction model. Use the initial moisture content of sludge, PAC dosage and feed pressure of historical dewatering parameter data in the training sample set as input features, and the first time when the moisture content of sludge cake is less than or equal to the preset first threshold as the output target to train the efficiency prediction model. Step 3: Based on the efficiency prediction model, predict the second time to reach the target based on the real-time dehydration parameter data. If the second time to reach the target is greater than the preset second threshold, automatically adjust the real-time dehydration parameter data through the gradient descent algorithm until the predicted new second time to reach the target returns to within the second threshold.

2. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 1, characterized in that, Step 1: Collect historical dehydration parameter data to form a training sample set, including: Historical operational data is automatically collected from industrial automation systems in wastewater treatment plants and / or river and lake sediment disposal centers. Based on preset data cleaning rules, historical operational data is cleaned to obtain historical dehydration parameter data. The preset data cleaning rules are: to remove invalid dehydration cycle data and abnormal physical value data. Based on each effective dehydration cycle, historical dehydration parameter data are used to form independent training samples, which are then divided into training set, test set, and validation set according to a preset division ratio.

3. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 1, characterized in that, The real-time dehydration parameters are automatically adjusted using a gradient descent algorithm until the predicted new second target time returns to within the second threshold, including: Based on real-time dehydration parameter data and the second time to achieve the target, a target loss function is constructed; Calculate the gradient of the target loss function and determine the direction of real-time dehydration parameter adjustment; Based on the gradient of the target loss function and the learning rate, gradient descent is applied to update the input feature vector; The real-time dehydration parameters are automatically adjusted based on the updated input feature vector.

4. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 3, characterized in that, Also includes: After determining the direction of adjustment of real-time dehydration parameters, the abnormality is attributed by comparing the direction of adjustment of real-time dehydration parameters with the direction of adjustment of standard dehydration parameters.

5. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 4, characterized in that, After determining the direction of real-time dehydration parameter adjustment, anomaly attribution is performed by comparing the real-time dehydration parameter adjustment direction with the standard dehydration parameter adjustment direction, including: Adjust the direction based on real-time dehydration parameters and construct a real-time adjustment direction vector; Construct a knowledge base for adjusting standard dehydration parameters; Compare the real-time adjustment direction vector with the standard adjustment direction vector in the standard dehydration parameter adjustment direction knowledge base to obtain the comparison results; Based on the comparison results and the anomaly cause database, determine the corresponding anomaly causes and handling recommendations, and respond to the handling recommendations.

6. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 5, characterized in that, Construct a knowledge base for adjusting standard dehydration parameters, including: Normal dehydration parameter data were selected from historical dehydration parameter data; Small perturbations were applied to the normal dehydration parameter data, and changes in the predicted time to reach the target were observed based on the efficiency prediction model; Based on the applied small perturbation and the change in the time to reach the target, calculate the gradient direction under each perturbation and determine the gradient direction vector; The gradient direction vectors corresponding to the normal dehydration parameter data are aggregated to form a standard directional probability distribution. Cluster analysis is performed based on the standard directional probability distribution to extract typical adjustment patterns and their corresponding standard adjustment direction vectors to construct a knowledge base for standard dehydration parameter adjustment directions.

7. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 3, characterized in that, The steps for determining the learning rate include: The efficiency prediction model is embedded into the digital twin control system, and the real dehydration process is simulated and closed-loop control is performed based on historical dehydration parameter data. In the closed-loop control process, candidate learning rates from a preset candidate learning rate sequence are randomly applied; An evaluation distribution is generated based on the candidate learning rate being applied, and the evaluation distribution is mapped onto the closed-loop control scenario corresponding to the candidate learning rate. The evaluation distribution includes: the convergence speed of the closed-loop control, parameter volatility, and safety compliance rate. Based on the paths of the scene elements of interest to the experimenters in the closed-loop control scenario, determine the development trend of the evaluation distribution of the current mapping; Based on the development trend, determine the traversal direction of the candidate learning rate sequence; Based on the traversal direction, the application steps of the candidate learning rate are repeated to determine the source candidate learning rate whose trend increment of the corresponding evaluation distribution is less than the preset third threshold as the final determined learning rate.

8. The method for dynamic prediction and parameter self-adjustment of river and lake sediment dewatering efficiency as described in claim 7, characterized in that, Based on the paths of the scene elements of interest to the experimenters in the closed-loop control scenario, the development trend of the evaluation distribution of the current mapping is determined, including: Based on the historical adjustment patterns of performance metrics, determine the system log records prior to the adjustment; Determine the data extraction dimensions from the system logs; If the data extraction dimensions match the visual data extraction dimensions, determine the timestamp, visual subject, and visual data features of the corresponding record data; Visual data features are marked on the visual subject and sorted according to the time sequence of the timestamp records to obtain the standard scene element path. The standard scene element path is then associated with the performance index trend corresponding to the historical adjustment situation. We will focus on matching scene element paths with standard scene element paths. If the scene element paths match, we will summarize the corresponding performance index trends to obtain the development trend of the evaluation distribution.

9. A dynamic prediction and parameter self-adjustment system for river and lake sediment dewatering efficiency, characterized in that, include: The sample acquisition module is used to collect historical dehydration parameter data to form a training sample set. The model training module is used to build a prediction model using an LSTM neural network. It takes the initial moisture content of sludge, PAC dosage and feed pressure of historical dewatering parameter data in the training sample set as input features, and the first time when the moisture content of sludge cake is less than or equal to the preset first threshold as the output target to train the efficiency prediction model. The parameter self-adjustment module is used to predict the second target time based on the efficiency prediction model and real-time dehydration parameter data. If the second target time is greater than the preset second threshold, the real-time dehydration parameter data is automatically adjusted through the gradient descent algorithm until the predicted new second target time returns to within the second threshold.

10. The dynamic prediction and parameter self-adjustment system for river and lake sediment dewatering efficiency as described in claim 9, characterized in that, The sample acquisition module collects historical dehydration parameter data to form a training sample set, including: Historical operational data is automatically collected from industrial automation systems in wastewater treatment plants and / or river and lake sediment disposal centers. Based on preset data cleaning rules, historical operational data is cleaned to obtain historical dehydration parameter data. The preset data cleaning rules are: to remove invalid dehydration cycle data and abnormal physical value data. Based on each effective dehydration cycle, historical dehydration parameter data are used to form independent training samples, which are then divided into training set, test set, and validation set according to a preset division ratio.

Images