Interaction method and system based on water station system agent

Abstract

The application discloses an interactive method and system based on an intelligent agent of a water station system, and belongs to the technical field of intelligent agent interaction, and the technical solution points of the application include the following steps: obtaining water quality monitoring parameters corresponding to each sub-region in a water area; identifying an abnormal region according to the water quality monitoring parameters, wherein the abnormal region includes multiple sub-regions; and determining an optimal processing device in multiple processing devices corresponding to the multiple sub-regions according to risk coefficients corresponding to the multiple sub-regions.The application accurately identifies an abnormal region by using water quality detection data collected by a perception layer, the decision layer fuses Bayesian inference and a virtual self-game algorithm to quantize the comprehensive utility of processing devices corresponding to each sub-region, and when facing a new abnormal type, the model parameters can be adapted to unknown abnormal types by fine-tuning parameters with less data, and the application realizes the collaborative processing of cross-region water pollution through the cooperation of multiple intelligent agents.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent interaction technology, and more specifically to an interaction method and system based on intelligent agents in a water station system. Background Technology

[0002] In the field of cross-regional collaborative water pollution control, with the acceleration of industrialization and urbanization, the ecological environment of watersheds faces severe challenges. However, the existing technology system lacks real-time and comprehensive monitoring capabilities. After pollution occurs, pollutants have spread to multiple areas with the water flow, resulting in high pollution control costs. Cross-regional water pollution involves multiple administrative entities, and the determination of responsibility relies on manual consultation. There is a lack of scientific means to quantify responsibility, resulting in a low implementation rate. Furthermore, when facing new abnormal situations such as emerging pollutants, the development of disposal strategies relies on expert experience and lacks intelligent adaptation capabilities. This results in a long generation cycle for disposal solutions for new abnormal situations, which are prone to limitations. Consequently, disposal solutions are unable to cope with complex and ever-changing pollution risks. Therefore, existing technologies are insufficient and cannot meet the needs of efficient governance in complex pollution scenarios. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an interaction method and system based on intelligent agents of water station systems, which realizes efficient collaborative treatment of cross-regional water pollution through the multi-agent structure within the interaction system.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] This invention provides an interaction method based on an intelligent agent in a water station system, comprising:

[0006] Obtain water quality monitoring parameters for each sub-region within the water body;

[0007] Anomaly areas are identified based on the water quality monitoring parameters, and the anomaly areas include multiple sub-regions;

[0008] Based on the risk coefficients corresponding to the multiple sub-regions, the optimal processing device is determined from the multiple processing devices corresponding to the multiple sub-regions;

[0009] Among the multiple processing devices corresponding to the multiple sub-regions, the optimal processing device is determined, based on the risk coefficients corresponding to the multiple sub-regions, including:

[0010] Based on the risk coefficient, the scheduling priority vector of multiple processing devices corresponding to the multiple sub-regions is obtained;

[0011] The set of candidate processing devices is determined based on the scheduling priority vector;

[0012] Based on Bayes' theorem and the virtual game algorithm, the optimal processing device is determined from the set of candidate processing devices.

[0013] As a further improvement of the present invention, identifying abnormal areas based on the water quality monitoring parameters includes:

[0014] Based on the water quality monitoring parameters and the encoder, the feature vector corresponding to each sub-region is obtained;

[0015] Based on the feature vector and the multilayer perceptron model, the association weight between any two sub-regions is obtained;

[0016] The abnormal regions are identified based on the association weights.

[0017] As a further improvement of the present invention, the step of determining the risk coefficient includes:

[0018] Based on the feature vector and the preset anomaly type library, determine the anomaly type corresponding to the anomaly region;

[0019] The parameter set of the neural network model is determined according to the anomaly type. The parameter set is obtained by training based on historical data, and each anomaly type corresponds to a parameter set.

[0020] Based on the parameter set and the feature vector, the risk coefficients corresponding to the multiple sub-regions are obtained.

[0021] As a further improvement of the present invention, the step of determining the optimal processing device from the set of candidate processing devices based on Bayes' theorem and virtual game algorithm includes:

[0022] The initial probabilities corresponding to the virtual game algorithm are determined according to Bayes' theorem;

[0023] Based on the initial probability, the plurality of processing devices are divided into one main device and at least one auxiliary device;

[0024] The process involves performing an iterative step, which includes determining the current scheduling strategy of the master device based on the current belief value and utility function of the slave device, updating the current belief value of the slave device, and continuing until a preset termination condition is met, and then outputting the optimal processing device.

[0025] As a further improvement of the present invention, determining the current scheduling strategy of the master device based on the current belief value and utility function corresponding to the slave device includes:

[0026] Determine the expected utility based on the current belief value and the utility function;

[0027] The current scheduling strategy of the master device is determined based on the expected utility.

[0028] As a further improvement of the present invention, updating the current belief value corresponding to the secondary device includes:

[0029] Update the gain coefficient and credit score according to the current scheduling strategy;

[0030] The current belief value corresponding to the secondary device is updated based on the gain coefficient, credit score, current iteration number, and indicator function.

[0031] As a further improvement of the present invention, if the preset anomaly type library does not contain an anomaly type corresponding to the anomaly region, the step of determining the risk coefficient includes:

[0032] Based on the feature vector and the preset anomaly type library, determine the anomaly type group corresponding to the anomaly region;

[0033] Obtain the anomaly type with the highest similarity to the feature vector from the anomaly type group, and use the parameter group corresponding to the anomaly type with the highest similarity as the initial parameter group of the neural network model;

[0034] The initial parameter group is updated according to the anomaly type group to obtain the parameter group of the neural network model;

[0035] Based on the parameter set and the feature vector, the risk coefficients corresponding to the multiple sub-regions are obtained.

[0036] As a further improvement of the present invention, the initial parameter group is updated according to the anomaly type group to obtain the parameter group of the neural network model, including:

[0037] Remove the anomaly type that has the highest similarity to the feature vector from the anomaly type group to obtain the training set;

[0038] The gradient for each parameter is obtained based on the data corresponding to the training set.

[0039] The initial parameter set is updated based on the gradient to obtain the parameter set of the neural network model.

[0040] This invention provides an interactive system based on a water station system intelligent agent, used to implement the aforementioned interactive method based on a water station system intelligent agent. The interactive system includes:

[0041] The perception layer agent is used to acquire water quality monitoring parameters corresponding to each sub-region in the water area, and to identify abnormal regions based on the water quality monitoring parameters. The abnormal regions include multiple sub-regions.

[0042] The decision-making agent is used to determine the optimal processing device among multiple processing devices corresponding to multiple sub-regions within the abnormal region, based on the risk coefficients corresponding to multiple sub-regions within the abnormal region.

[0043] The meta-agent is used to update the initial parameter group according to the anomaly type group to obtain the parameter group of the neural network model when the anomaly type corresponding to the anomaly region does not exist in the preset anomaly type library.

[0044] This invention achieves efficient collaborative treatment of cross-regional water pollution through a multi-layered intelligent agent system structure. Specifically, the invention accurately identifies abnormal areas by collecting water quality detection data through the perception layer intelligent agent, and quantifies the comprehensive effectiveness of the treatment equipment corresponding to each sub-region by integrating Bayesian inference and virtual self-game algorithm through the decision layer intelligent agent. This enables pollution control by calling fewer treatment devices when cross-regional water pollution occurs. Furthermore, when facing new anomaly types, the meta-intelligent agent fine-tunes the parameters with less data, so that the model parameters can adapt to unknown anomaly types. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0046] Figure 2 A schematic diagram of the sub-region;

[0047] Figure 3 This is a schematic diagram illustrating the steps of an iterative operation.

[0048] Figure 4 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0049] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0050] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0051] like Figure 1 As shown in the figure, this application provides an interaction method based on a smart agent in a water station system, including:

[0052] Obtain water quality monitoring parameters for each sub-region within the water body;

[0053] Anomaly areas are identified based on water quality monitoring parameters. These anomaly areas include multiple sub-regions.

[0054] Based on the risk coefficients corresponding to multiple sub-regions, the optimal processing device is determined among the multiple processing devices corresponding to the multiple sub-regions.

[0055] Specifically, such as Figure 2 As shown, the water area in this embodiment includes multiple sub-regions (A, B, C, D, E, F). The treatment of the entire water area is performed through an interactive system based on the water station system's intelligent agents. The interactive system includes a perception layer intelligent agent, a decision-making layer intelligent agent, and a meta-intelligent agent. The decision-making layer intelligent agent includes multiple sub-intelligent agents, each corresponding to a sub-region. When the water area is large, a treatment device is configured for each sub-region to complete the pollution treatment within that sub-region. Each sub-region is equipped with a buoy, which is used to collect water quality monitoring parameters corresponding to each sub-region, including water temperature, flow rate, pH value, turbidity, and COD. If the buoy is equipped with a specific sensor, it also includes the specific pollutant type detected, such as heavy metals and specific pesticides, and records the time and latitude and longitude of each collection. After the sensing agent obtains the water quality monitoring parameters collected by the buoy, it identifies abnormal areas based on the water quality monitoring parameters. Since the abnormal area involves multiple sub-regions and each sub-region has its corresponding processing equipment, this embodiment determines the optimal processing equipment among multiple processing equipment corresponding to multiple sub-regions by combining the risk coefficient corresponding to each sub-region with Bayesian inference and virtual self-game algorithm.

[0056] This embodiment sets up a multi-layer intelligent agent system structure. The water quality detection data collected by the perception layer intelligent agent accurately identifies abnormal areas. The decision layer intelligent agent integrates Bayesian inference and virtual self-game algorithm to quantify the comprehensive effectiveness of the processing equipment corresponding to each sub-region in the abnormal area, and obtains the optimal processing equipment. This avoids the duplication of work and waste of resources caused by starting multiple devices at the same time, and solves the problem that it is difficult for sub-regions to reach a consensus when cross-regional governance, thus realizing efficient collaborative treatment of cross-regional water pollution.

[0057] Furthermore, this embodiment provides a step for determining the optimal processing device among multiple processing devices corresponding to multiple sub-regions based on risk coefficients, including:

[0058] Based on the risk coefficient, the scheduling priority vector of multiple processing devices corresponding to multiple sub-regions is obtained;

[0059] The set of candidate processing devices is determined based on the scheduling priority vector;

[0060] Based on Bayes' theorem and the virtual game algorithm, the optimal processing device is determined from the set of candidate processing devices.

[0061] Specifically, each sub-region within the abnormal region corresponds to a processing device, and for each processing device, its corresponding scheduling priority vector can be calculated. For example, assuming there are multiple processing devices within the abnormal region... The sub-regions, for the first Each sub-region, and the scheduling priority vector of its corresponding processing device. for:

[0062]

[0063] in, For the first The risk coefficient corresponding to each sub-region.

[0064] Then, the value corresponding to each scheduling priority vector is used as the scheduling priority of the processing equipment corresponding to each sub-region. In this embodiment, each sub-region corresponds to a dedicated processing equipment (such as an unmanned vessel), and is equipped with a corresponding pollutant treatment agent. For example, if the buoy can only detect conventional numerical monitoring parameters and cannot obtain specific pollutant types, such as a sudden increase in COD but unknown composition, a general-purpose agent can be used for treatment first. For example, a combination of oxidant and adsorbent can be added, which can oxidize some organic matter and adsorb possible toxic substances. At the same time, the equipment used for deployment can quickly collect water samples for further analysis and treatment. If the buoy can obtain specific pollutant types, then the agent corresponding to the specific pollutant is added for treatment. This embodiment will not be elaborated on here.

[0065] Based on the water quality monitoring parameters and area of ​​each sub-region, the amount of reagent to be applied can be determined. The treatment equipment is a dedicated device pre-configured for each sub-region, requiring no additional application. Only the corresponding equipment needs to be scheduled to start operation. Since the abnormal area contains multiple sub-regions, the reagent cost and equipment wear cost can be shared by the corresponding sub-regions after equipment scheduling. This embodiment does not impose any restrictions on this.

[0066] After obtaining the scheduling priority, multiple high-priority processing devices can be selected as a candidate set based on a preset scheduling threshold. This avoids low-priority devices from participating in the game and ensures the accuracy of the results. After determining the candidate set, the optimal processing device needs to be selected through autonomous game theory. The core of the game is to make each processing device in each sub-region an independent player, autonomously choosing its own optimal strategy based on its prediction of the strategies of other devices, while constraining the possibility of strategy masquerading. Through multiple rounds of interaction, a stable result is reached, and finally, the optimal processing device with the highest global utility is selected. The game between processing devices in each sub-region is executed by the sub-agent in each sub-region. Strategy masquerading refers to a processing device, as an independent player, deliberately hiding its actual capabilities and choosing a low-cost strategy that does not match its own capabilities for its own benefit (such as cost saving). This embodiment does not limit the specific value of the scheduling threshold.

[0067] This embodiment obtains the initial probability based on Bayes' theorem, simulates the independent decision-making process of each processing device through autonomous game, quantifies the processing utility of the devices under multi-party interaction, avoids the limitations of preset rules, and achieves optimal device selection that is more in line with the actual scenario through matching constraints between strategies and capabilities.

[0068] Furthermore, this embodiment provides a step for identifying abnormal areas based on water quality monitoring parameters, including:

[0069] Based on the water quality monitoring parameters and the encoder, the feature vector corresponding to each sub-region is obtained;

[0070] Based on the feature vectors and the multilayer perceptron model, the association weights between any two sub-regions are obtained;

[0071] Identify abnormal regions based on association weights.

[0072] The step of identifying abnormal areas based on water quality monitoring parameters is executed by the sensing layer agent.

[0073] Specifically, each buoy has the same collection time and collection interval, so it will receive the water quality parameters corresponding to each sub-region at the current moment. After the sensing layer agent receives the parameters collected by the buoy, it also needs to obtain the risk label and historical exceedance frequency corresponding to each sub-region. The risk label and historical exceedance frequency are obtained according to the historical water quality monitoring parameters corresponding to the sub-region. For example, the risk label of the sub-region can be set to high risk, medium risk and low risk according to the historical water quality monitoring parameters, and the historical exceedance frequency corresponding to the sub-region can be set to the frequency of water quality monitoring parameters exceeding the standard in the past 30 days. This is a technical means that can be implemented by those skilled in the art, and will not be described in detail in this embodiment.

[0074] Next, the water quality testing parameters, risk labels, and historical exceedance frequencies are preprocessed, such as by normalization, to eliminate the influence of dimensions between data points and to convert text data into numerical data. Then, for each sub-region, the normalized data corresponding to that sub-region is input into a linear embedding layer. The linear embedding layer maps each data point to a vector with the same dimension. For example, if the water quality parameters include water temperature, flow rate, and pH value, inputting each normalized data point into the linear embedding layer will output a vector with five identical dimensions. Then, each output vector is concatenated, and the concatenated result is input into the encoder. After passing through an attention mechanism and a feedforward neural network, the feature vector corresponding to that sub-region is output, thus obtaining the feature vector corresponding to each sub-region. For example, the encoder can be a Transformer encoder; this embodiment is not limited to this.

[0075] After obtaining the feature vector corresponding to each sub-region, it is necessary to calculate the association weight between any two sub-regions. For example, for the first sub-region... The and the first The association weights corresponding to each sub-region for:

[0076]

[0077] in, Represents an exponential function. Indicates the first The set of sub-regions within a preset range corresponding to each sub-region For the first The feature vectors corresponding to each sub-region For the first The feature vectors corresponding to each sub-region Used as an index for traversal. Each sub-region in, For the first The feature vectors corresponding to each sub-region Indicates will and splicing the beginning and end together, This is a multilayer perceptron model used to map the concatenated results to an association score. In this embodiment, the size of the preset range is not limited; for example, the preset range is a circular area centered at the center of the sub-region, with a radius of 500m. Furthermore, the association weight... Used to indicate the first The sub-regions for the first The influence of each sub-region, and the preset range corresponding to each sub-region is different, therefore and different.

[0078] Next, an initial set of abnormal sub-regions is determined based on the water quality monitoring parameters and preset standards corresponding to each sub-region. This embodiment does not limit the preset standards; those skilled in the art can determine them based on experience or historical data. Then, for each sub-region in the initial set of abnormal sub-regions, the following steps are performed: First, for one sub-region, determine all its corresponding correlation weights. For example, for the first sub-region... Each sub-region has a corresponding set of association weights. , Based on the association weight and preset threshold, select the first... Within a preset range corresponding to each sub-region, sub-regions with association weights greater than a threshold are taken as a diffusion region set. After performing the above steps on each sub-region in the initial abnormal sub-region set, a diffusion region set corresponding to each sub-region can be obtained. Then, for each sub-region in the diffusion region set, all its corresponding association weights are filtered again, and sub-regions with association weights greater than the threshold are selected according to the threshold. This results in a source region set corresponding to each sub-region in the diffusion region set. Each source region set is merged, and duplicate sub-regions are deleted to obtain the final abnormal region.

[0079] This embodiment first calculates the correlation weights using feature vectors and determines the initial set of abnormal sub-regions. However, the initial set of abnormal sub-regions can only reflect the local anomalies of each sub-region and cannot capture the diffusion trend and potential impact range of pollution. Therefore, this embodiment further traces the pollution diffusion path downstream based on the correlation weights to obtain a set of diffusion areas. However, it is difficult to determine the pollution source area using only the set of diffusion areas. Therefore, this embodiment further uses the correlation weights to trace the potential pollution source in reverse, and finally obtains a more accurate and complete set of abnormal areas.

[0080] Furthermore, this embodiment provides a step for determining the risk coefficient, including:

[0081] Based on the feature vector and the preset anomaly type library, determine the anomaly type corresponding to the anomaly region;

[0082] The parameter set of the neural network model is determined based on the anomaly type. The parameter set is obtained by training based on historical data, and each anomaly type corresponds to a parameter set.

[0083] Based on the parameter set and feature vector, the risk coefficients corresponding to multiple sub-regions are obtained.

[0084] This embodiment only addresses the case where there is only one type of anomaly in the abnormal region.

[0085] Specifically, firstly, water quality monitoring data corresponding to each sub-region in the abnormal area is obtained, and the main sub-regions in the abnormal area are selected through preset criteria. The purpose of selecting the main sub-regions is to reduce the feature vectors required to determine the anomaly type through the preset anomaly type library, reduce the amount of data, improve the calculation efficiency, and avoid the impact of regions with low correlation weight on the identification accuracy. Therefore, this embodiment does not limit the specific criteria. For example, a sub-region in the water quality monitoring data where one or more indicators exceed twice the preset standard can be identified as the main sub-region.

[0086] Next, the feature vector corresponding to each main sub-region is obtained, and the anomaly type corresponding to the anomaly region is determined according to a preset anomaly type library. For example, the anomaly type can include chemical pollution type, agricultural pollution type (such as caused by excessive use of chemical fertilizers or pesticides), and natural runoff type (such as water pollution caused by geological erosion and soil leaching). The anomaly type can also be determined according to the specific pollutant, such as benzene series pollution and pyrethroid pollution, etc. This embodiment does not limit this. Then, the feature vectors corresponding to each main sub-region are concatenated end to end, and the concatenated result is input into the classification model. For example, a trained support vector machine model can be used for classification, and finally the anomaly type corresponding to the anomaly region is output. Furthermore, since the number of main sub-regions is different for different collection times, the dimension of the result after each concatenation is different. Since fixed-dimensional vectors are usually used when training support vector machines, feature encoding, dimensionality reduction, or dimensionality increase steps can be added after the concatenation step to map the concatenated vector to the fixed dimension required by the support vector machine. This is a technical means that can be implemented by those skilled in the art, and this embodiment will not elaborate on it here.

[0087] Then, the risk coefficient is calculated based on the neural network model, for example, the first... Risk coefficient corresponding to each sub-region for:

[0088]

[0089] in, The activation function can be either the Sigmoid or Softmax function, used to compress the output of the neural network model to a range of 0-1. , and As weight, , and For bias, The number of weights and biases is set according to the structure of the neural network model, such as a 3-layer fully connected neural network which includes 3 weights and 3 biases. The formula provided in this embodiment is only an example and does not limit the specific number of layers. Taking a 3-layer fully connected neural network as an example, a parameter group includes 3 weights and 3 biases, for a total of 6 parameters. Each anomaly type corresponds to a parameter group. After determining the anomaly type, it is necessary to call its corresponding parameter group to determine the risk coefficient. The risk coefficient is used to quantify the degree of harm of the anomaly region.

[0090] This embodiment takes into account that different anomaly types have different pollution characteristics. For example, chemical pollution spreads quickly, while agricultural pollution spreads slowly. The treatment objectives are also different. For example, chemical pollution requires rapid interception, while agricultural pollution requires long-term remediation. Uniform parameters cannot adapt to the diverse anomaly patterns and are prone to misjudgment. Therefore, this embodiment trains the neural network model based on historical data to obtain the parameters corresponding to each anomaly type, thereby obtaining more accurate risk parameters and improving the rationality of subsequent resource allocation and cost allocation.

[0091] Furthermore, this embodiment provides a step for determining the risk coefficient if the anomaly type corresponding to the anomaly region does not exist in the preset anomaly type library, including:

[0092] Based on the feature vector and the preset anomaly type library, determine the anomaly type group corresponding to the anomaly region;

[0093] Obtain the anomaly type with the highest similarity to the feature vector from the anomaly type group, and use the parameter group corresponding to the anomaly type with the highest similarity as the initial parameter group of the neural network model;

[0094] The initial parameter set is updated based on the anomaly type group to obtain the parameter set of the neural network model;

[0095] Based on the parameter set and feature vector, the risk coefficients corresponding to multiple sub-regions are obtained.

[0096] Specifically, when using Support Vector Machines (SVMs) for classification to determine anomaly types, the SVM outputs a probability value for each anomaly type. The anomaly type with the highest probability value above a threshold is selected as the anomaly type for the current time step. If no type above the threshold exists, it means there is no anomaly type corresponding to the anomaly region at that moment.

[0097] At this point, the anomaly types are sorted from highest to lowest probability values, and the top few are selected as anomaly type groups. The anomaly type with the highest similarity (highest similarity) is then chosen, and its corresponding parameter set is obtained as the initial parameter set. This embodiment does not limit the number of anomaly types selected. The initial parameter set is then updated based on the anomaly type groups to obtain the parameter set of the neural network model. Based on the parameter set and feature vectors, risk coefficients corresponding to multiple sub-regions are obtained. The calculation steps for the risk coefficients are the same as described above and will not be repeated here.

[0098] Furthermore, this embodiment provides a step for updating the initial parameter set according to the anomaly type group to obtain the parameter set of the neural network model, including:

[0099] Remove the anomaly type with the highest similarity to the feature vector from the anomaly type group to obtain the training set;

[0100] The gradient for each parameter is obtained based on the data corresponding to the training set.

[0101] The initial parameter set is updated based on the gradient to obtain the parameter set of the neural network model.

[0102] Specifically, firstly, the anomaly type with the highest similarity to the feature vector is removed from the anomaly type group. The set of each remaining anomaly type is then used as the training set. Next, the data corresponding to the training set is obtained, namely, the feature vector corresponding to each anomaly type and the risk coefficient corresponding to that feature vector. For example, when determining the feature vector corresponding to each anomaly type, the feature vectors corresponding to each sub-region of that anomaly type within a preset time period can be obtained. Then, the mean of these feature vectors is used as the feature vector corresponding to that anomaly type, and the mean of the risk coefficients corresponding to these feature vectors is used as the risk coefficient corresponding to that anomaly type.

[0103] Next, a loss function is preset to calculate the gradient. An exemplary loss function can be set as mean squared error, which measures the difference between the predicted risk coefficient and the risk coefficient corresponding to the anomaly type. Then, based on the loss function and the training set, the parameters are updated using gradient descent to obtain the final parameter set. The step of updating the parameters using gradient descent is a technical step that can be implemented by those skilled in the art, and will not be described in detail here.

[0104] This embodiment targets scenarios where no matching anomalies exist in a preset anomaly type library. It selects multiple anomaly types with high probabilities to form an anomaly type group, and uses the parameter of the type with the highest similarity within the group as the initial parameter of the neural network to avoid the blindness of parameter initialization. Then, with the help of a small sample fine-tuning mechanism, the neural network is made applicable to unknown anomaly types, thereby accurately obtaining the risk coefficient of the sub-region. This breaks through the dependence of traditional models on known anomalies and realizes intelligent identification and quantification of risks in unknown anomaly scenarios.

[0105] Furthermore, this embodiment provides a step for determining the optimal processing device from the candidate processing device set based on Bayes' theorem and a virtual game algorithm, including:

[0106] Determine the initial probabilities corresponding to the virtual game algorithm based on Bayes' theorem;

[0107] Based on the initial probability, multiple processing devices are divided into one master device and at least one slave device;

[0108] Perform iterative steps, such as Figure 3 As shown, the iterative steps include determining the current scheduling strategy of the master device based on the current belief value and utility function of the slave device, updating the current belief value of the slave device, until the preset termination condition is reached, and outputting the optimal processing device.

[0109] Specifically, for each processing device, its corresponding initial probability needs to be calculated according to Bayes' theorem. For example, assuming that the candidate processing device set includes device 1, device 2, and device 3, corresponding to sub-region 1, sub-region 2, and sub-region 3 respectively, taking device 1 as an example, firstly, the feature vector corresponding to sub-region 1 of device 1 is obtained, denoted as the first feature. Then, the abnormal feature vector corresponding to the abnormal region is obtained. For example, the time when the abnormal region was considered an abnormal region in the past year can be queried and denoted as the abnormal time. The feature vector corresponding to each sub-region in the abnormal region at each abnormal time is obtained, thereby obtaining the feature group corresponding to each abnormal time, and thus obtaining the second feature set. The feature group includes the feature vector corresponding to each sub-region in the abnormal region at the abnormal time. The second feature set includes each feature group. The one-year time is only an example, and those skilled in the art can choose an appropriate time according to the actual situation.

[0110] Next, multiple feature vectors corresponding to sub-region 1 are determined from the feature groups included in the second feature set. The similarity between the first feature and each feature vector corresponding to sub-region 1 is calculated. The set of feature vectors with similarity exceeding a threshold is taken as the similarity set corresponding to the first feature. This embodiment does not limit the specific calculation method and threshold of similarity. For example, Euclidean distance or cosine similarity can be used, or the anomaly type corresponding to each feature vector in multiple feature vectors can be obtained, and the feature vectors with the same anomaly type as the current anomaly type are taken as the similarity set corresponding to the first feature. Then, each anomaly time corresponding to the similarity set is traversed to determine the number of times that device 1 participates in processing and the processing effect is good in these anomaly times. Then, the ratio of this number to the number of feature vectors in the similarity set is calculated as the conditional probability corresponding to device 1. Next, multiple feature vectors corresponding to sub-region 1 are traversed to determine the number of times that device 1 participates in processing and the processing effect is good in these anomaly times. The ratio of this number to the multiple feature vectors corresponding to sub-region 1 is taken as the prior probability corresponding to device 1. Similarly, the conditional probability and prior probability corresponding to each processing device in the candidate processing device set can be calculated. Whether the treatment effect of device 1 is excellent can be determined based on the water quality test parameters before and after treatment and the time consumed, etc., and this embodiment does not impose any restrictions on this.

[0111] Next, according to Bayes' theorem, the posterior probability corresponding to device 1 can be obtained. for:

[0112]

[0113]

[0114] in, , and These correspond to equipment 1, equipment 2, and equipment 3, respectively. This refers to the conditional probability calculated above for device 1, which is the probability that the processing effect for this type of anomaly is optimal, assuming device 1 participates in the processing. That is, the type of exception corresponding to the current abnormal region. That is, the prior probability corresponding to device 1 calculated above is the probability that the processing effect of all anomaly types is excellent, assuming that device 1 participates in the processing. Similarly, and Let be the conditional probabilities corresponding to device 2 and device 3. and Let be the prior probabilities corresponding to device 2 and device 3. The total probability represents the overall probability that this type of anomaly will be effectively handled (with excellent results), and is calculated using the total probability formula.

[0115] Next, the posterior probabilities of each device are compared, and the device with the highest posterior probability is designated as the primary device, while the remaining devices are designated as secondary devices. Furthermore, this embodiment utilizes a virtual game theory algorithm to determine the optimal processing device among the three devices. First, the initial belief value for each device is determined. Following the example above, a strategy library is defined for each device, including different strategy types and corresponding capability conditions. Since the primary device has a higher posterior probability, it is more adaptable to handling this type of anomaly. Therefore, to better handle the current anomaly and maintain its high adaptability, it needs to possess stricter capability conditions.

[0116] For example, each device's strategy library contains three strategy types: active processing, collaborative processing, and standby. For the primary device, the capability conditions for the active processing type can be: drug reserve ≥ 60L, power range ≥ 2.5h, and normal communication; the capability conditions for the collaborative processing type can be: drug reserve ≥ 30L, power range ≥ 1.5h, and normal communication; the standby type has no mandatory capability conditions. For the secondary device, the capability conditions for the active processing type can be: drug reserve ≥ 50L, power range ≥ 2h, and normal communication; the capability conditions for the collaborative processing type can be: drug reserve ≥ 20L, power range ≥ 1h, and normal communication; the standby type has no mandatory capability conditions. This embodiment does not limit the number of strategies corresponding to each subject; for example, 3-5 strategies can be set to reduce computational complexity while ensuring accuracy. The values ​​in the above capability conditions are only examples used to illustrate the numerical relationships between capability conditions corresponding to different types and different devices. This embodiment does not impose any limitations on this, and those skilled in the art can set them according to actual conditions.

[0117] Next, a credit score is initialized for each device. The function of the credit score is to punish spoofing behavior. Preferably, the initial credit score of the main device should be higher than that of the secondary device. For example, the initial credit score of the main device is 105 and the initial credit score of the secondary device is 100. Setting a higher initial credit score for the main device can prevent the main device from being distrusted by other devices due to a single mistake, thus ensuring the main device's position as the core decision-maker.

[0118] The initial belief value is the predicted probability of each processing device choosing each strategy for another processing device. Since there are three devices in this embodiment and the strategy library includes three strategy types, each processing device has 2 × 3 = 6 initial belief values. Taking the master device as an example, assuming device 1 is the master device, the three initial belief values ​​of device 1 for device 2 are as follows: , and , representing the probabilities that device 1 predicts device 2 will choose one of three strategy types: proactive processing, collaborative processing, and standby, respectively. For example, ,in This represents the posterior probability corresponding to device 2. Regarding the strategy allocation coefficient, this embodiment does not limit the specific value of the strategy allocation coefficient. Those skilled in the art can determine it based on the strategy tendencies of different devices. For example, since the master device is the core processor, it will initially assume that the task of the secondary device is collaborative processing. Therefore, the master device will predict that the probability of the secondary device choosing collaborative processing is higher than that of active processing. The corresponding strategy allocation coefficient should be less than The corresponding strategy allocation coefficients, such as The corresponding strategy allocation coefficient is 0.4. The corresponding strategy allocation coefficient is 0.5. The role weight coefficient represents the degree of attention different devices pay to other devices. This embodiment does not limit its specific value. For example, for devices 1 and 2, one is the primary device and the other is the secondary device. The primary device is the core, and the effectiveness of pollution control depends on the auxiliary strategies of the secondary device. The secondary devices provide equal assistance to each other, and the influence of their mutual strategies is far less than that of the primary device's strategy. Therefore, the role weight coefficient between the primary and secondary devices should be greater than the role weight coefficient between the secondary devices. For example, the role weight coefficient between the primary and secondary devices is 1.2, and the role weight coefficient between the secondary devices is 1. Similarly, each initial belief value can be calculated, which will not be elaborated upon in this embodiment.

[0119] Then, an iterative operation is performed, where each sub-agent sends its current belief value for its corresponding processing device to other sub-agents, and calculates the maximum expected utility of its own processing device. For example, for the master device, its maximum expected utility is:

[0120]

[0121] in, This indicates the policy corresponding to the master device (device 1). hour, , and The main device selects one of three strategies: proactive processing, collaborative processing, or standby. Similarly, and These represent the strategies corresponding to device 2 and device 3, respectively. At that time, the three strategies corresponding to devices 2 and 3 are respectively selected: proactive processing, collaborative processing, and standby. Indicates selection , and The utility function corresponding to time, because There are three possible values, therefore each strategy selected by the master device corresponds to nine utility function values. The expected utility is represented by the mean of nine utility function values ​​for each strategy chosen by the master device. This mean is used as the expected utility when the master device chooses that strategy. Therefore, the expected utility has three possible values. This represents the maximum expected utility, which is the largest of the three expected utilities. Similarly, for each secondary device, we can obtain the expected utility corresponding to each strategy it chooses, and thus obtain its corresponding maximum expected utility.

[0122] Specifically, the utility function can be determined based on revenue and cost. Taking the main equipment as an example, the utility function is:

[0123]

[0124] in, This indicates the risk coefficient of the sub-area corresponding to the main equipment. This represents the gain coefficient corresponding to the master device. The gain coefficient is used to reflect the role difference between the master device and the slave device. Since the master device is the core processor, setting a gain coefficient can match its core responsibility. Setting a gain coefficient can also enhance the benefits of the corresponding strategy and guide the master device to prioritize the active strategy. In this embodiment, the specific value of the gain coefficient is not limited. For example, it can be set to 1.1. Indicates the current strategy combination The corresponding utility coefficient is used to reflect the processing efficiency of different strategy combinations. When each device selects different strategies, its corresponding utility coefficient is different. This embodiment does not limit the specific value of the utility coefficient. Preferably, when there is cooperation between devices or when devices tend to choose active processing, the value of the utility coefficient is higher. For example, when the main device chooses active processing and the secondary devices choose cooperative processing, the utility coefficient is 1.8; when both the main device and the secondary devices choose active processing, the utility coefficient is 1.6; when both the main device and the secondary devices choose cooperative processing, the utility coefficient is 1.2; and when both the main device and the secondary devices choose standby, the utility coefficient is 0.5. This embodiment does not give the utility coefficients corresponding to all strategy combinations, and the values ​​provided in this embodiment are only examples and are only used to illustrate the size relationship of the utility coefficients corresponding to different strategy combinations. This represents the total cost corresponding to the main equipment (equipment 1), and can be set based on factors such as the distance the equipment travels to the abnormal area. This represents the cost of overlapping processing. When multiple devices simultaneously select active processing, it can easily lead to problems such as excessive dosage and resource waste. Therefore, this embodiment sets a quantifiable cost of overlapping processing to determine the degree of waste. This embodiment does not limit its value; for example, when multiple devices select active processing, a certain value can be set. When there are no multiple devices choosing to actively process, the settings are as follows: , and The weighting coefficients representing revenue and cost are not limited in this embodiment, but can be set by those skilled in the art according to actual circumstances.

[0125] Similarly, taking a secondary device (device 2) as an example, its corresponding utility function is:

[0126]

[0127] in, This represents the risk coefficient of the sub-region corresponding to device 2. This represents the overall cost of the device. The method for determining each parameter is the same as that for the main device, and will not be repeated here. Since the secondary device does not need to bear core responsibility, there is no need to set a gain coefficient. Similarly, the utility function corresponding to other secondary devices can be obtained.

[0128] Once each device has obtained its corresponding maximum expected utility, the strategy corresponding to that device within the maximum expected utility is taken as the current scheduling strategy for that device. Next, the matching degree between each device's current scheduling strategy and its actual capabilities needs to be calculated. For example, taking a secondary device as an example, assuming its current scheduling strategy is active processing, according to the above description, each device's capability conditions include drug reserves, power range, and communication status. For each condition, first calculate its corresponding individual matching degree. Assuming its current actual situation is 55L of drug reserves, 2.4h of power range, and communication anomalies, for example, for the drug reserve condition, the corresponding individual matching degree is 55 / 60≈0.92; for power range, the corresponding individual matching degree is 2.4 / 2.5=0.96; and for the communication condition, the corresponding individual matching degree is 0 (if communication is normal, the corresponding individual matching degree is 1). Then, take the minimum value among the three individual matching degrees as the matching degree between the device's current scheduling strategy and its actual capabilities. Similarly, the matching degree between the current scheduling strategy of each other device and its actual capabilities can be obtained, which will not be elaborated here in this embodiment.

[0129] Then, the gain coefficient and credit score are updated based on the matching degree. For example, a threshold can be set for the matching degree, and an update ratio can be set for the gain coefficient and credit score. When the matching degree is less than the threshold, the current gain coefficient and credit score are multiplied by the corresponding update ratio to obtain the updated credit score and gain coefficient, which are used for subsequent calculations. This embodiment does not limit the threshold and update ratio. For example, it can be set to 20%. If the matching degree is greater than or equal to the threshold, the gain coefficient and credit score remain unchanged.

[0130] Then, based on the updated credit score, the current belief value for each device is updated. Since this is the first iteration, it is equivalent to updating the initial belief value to adjust the current belief value. Taking an update as an example, the updated value is:

[0131]

[0132] in, Indicates the current iteration number. This indicates the updated credit score. This is an indicator function; when the current scheduling policy of device 2 is active processing, ,otherwise Similarly, each current belief value corresponding to each device can be updated. This embodiment will not elaborate on this. Then, the above steps are repeated multiple times based on the updated belief values ​​until the preset termination condition is reached.

[0133] The preset termination condition can be set as follows: in multiple consecutive iterations, the current scheduling strategy corresponding to each device does not change (this embodiment does not limit the specific number of iterations); or the maximum expected utility of a certain device is significantly higher than that of all other devices for multiple consecutive rounds (this embodiment does not limit the standard for significant higher utility or the specific number of iterations); or the matching degree of the master device is lower than the corresponding threshold for multiple consecutive rounds (this embodiment does not limit the specific number of iterations); or the iteration count reaches the preset iteration count, and the iteration is terminated when any one of the conditions is met.

[0134] Specifically, when the current scheduling policy for each device remains unchanged across multiple iterations, the current scheduling policy for each device is output. The device that actively processes the data is selected as the optimal scheduling device, and the remaining devices execute their corresponding current scheduling policies. If the maximum expected utility of a device is significantly higher than that of all other devices for multiple consecutive rounds, the policy corresponding to the maximum expected utility of that device is obtained. If it is active collaboration, it is selected as the optimal scheduling device, and the remaining devices execute their corresponding current scheduling policies. If it is collaborative processing or standby and the device is a secondary device, it indicates that the primary device is in an abnormal state. In this case, even if the device does not choose to actively process the data, it is still selected as the optimal scheduling device, and the remaining devices execute their corresponding current scheduling policies. If the matching degree of the primary device is lower than the corresponding threshold for multiple consecutive rounds, it also indicates that the primary device is in an abnormal state. In this case, the device with the largest maximum expected utility value among the secondary devices is selected as the optimal scheduling device, and the remaining devices execute their corresponding current scheduling policies. If the number of iterations reaches the preset number of iterations, each device executes its own corresponding current scheduling policy.

[0135] If all devices except the optimal scheduling device are selected to standby, the optimal scheduling device will carry the reagents to carry out all pollution control work. If there are other devices besides the optimal scheduling device that are not selected to standby, the optimal scheduling device will perform active processing, and the other devices will perform collaborative processing. However, this embodiment does not limit the specific allocation of work. For example, according to the actual situation of each device, the optimal scheduling device can carry 60% of the reagents, and the other devices can carry a total of 40% of the reagents.

[0136] Based on the above analysis, it can be summarized that the steps provided in this embodiment for determining the current scheduling strategy of the master device based on the current belief value and utility function corresponding to the slave device include:

[0137] Determine expected utility based on current belief value and utility function;

[0138] The current scheduling strategy for the master device is determined based on the expected utility.

[0139] This embodiment provides a step for updating the current belief value corresponding to a secondary device, including:

[0140] Update the gain coefficient and credit score according to the current scheduling strategy;

[0141] The current belief value corresponding to the secondary device is updated based on the gain coefficient, credit score, current iteration number, and indicator function.

[0142] This embodiment uses a strategy game based on credit score and matching degree to constrain the strategic spoofing behavior of devices. For example, if a device falsely reports its capability and chooses to actively process the problem, but the actual reagents or power are insufficient, the decrease in matching degree will trigger a simultaneous deduction of credit score. Other devices' belief values ​​will also be updated as the credit score weight decreases, thereby avoiding ineffective resource allocation caused by the main device's operation. Furthermore, this embodiment selects the main device based on posterior probability. If it is also the optimal processing device, it can directly complete efficient treatment in the most polluted core area, saving critical response time. However, when the main device's capability is insufficient, this embodiment allows other devices to act as the optimal processing device, avoiding the delay in core area treatment and resource waste caused by the main device's insufficient capability. This ensures that the allocation of equipment operations matches the actual capability and scenario requirements, while also adapting to the dynamic changes in the pollution scenario.

[0143] This embodiment uses Bayes' theorem to make reasonable inferences based on existing information, effectively determining the primary and secondary devices. Then, through the iterative steps of the virtual game algorithm, it solves the game problem under incomplete information conditions. Specifically, firstly, by updating the belief value, the game participants gradually correct their initial blind assumptions about the opponent's strategy to a precise grasp of the strategy selection rules, breaking through the limitation of information asymmetry. Secondly, based on utility function quantification and strategy selection, the complex comparison of the advantages and disadvantages of strategies is transformed into calculable utility values. Finally, through multiple rounds of iteration, the game converges from an unstable state of strategy oscillation to the optimal strategy result and processing device.

[0144] Furthermore, such as Figure 4 As shown, this application provides an interactive system based on a water station system intelligent agent, including:

[0145] The perception layer agent is used to acquire water quality monitoring parameters corresponding to each sub-region in the water area, and to identify abnormal areas based on the water quality monitoring parameters. The abnormal areas include multiple sub-regions.

[0146] The decision-making agent is used to determine the optimal processing device among multiple processing devices corresponding to multiple sub-regions within the abnormal region based on the risk coefficients of the multiple sub-regions within the abnormal region.

[0147] The meta-agent is used to update the initial parameter set according to the anomaly type set when there is no anomaly type corresponding to the anomaly region in the preset anomaly type library, so as to obtain the parameter set of the neural network model.

[0148] Specifically, the perception layer agent sends the identified abnormal regions to the decision layer agent, which then determines the optimal processing device among the multiple processing devices corresponding to the multiple sub-regions. The system also includes a large model hub for storing all preset information, such as the parameter set corresponding to each abnormality type and the parameter set of the neural network model obtained by the meta-agent, for other agents to call at any time.

[0149] This invention achieves efficient collaborative treatment of cross-regional water pollution through a multi-layered intelligent agent system structure. Specifically, the invention accurately identifies abnormal areas by collecting water quality detection data through the perception layer intelligent agent, quantifies the comprehensive effectiveness of the treatment equipment corresponding to each sub-region by integrating Bayesian inference and virtual self-game algorithm through the decision layer intelligent agent, and fine-tunes the parameters with less data when facing new anomaly types through the meta-intelligent agent so that the model parameters can adapt to unknown anomaly types.

[0150] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0151] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An interaction method based on intelligent agents in a water station system, characterized in that, include: Obtain water quality monitoring parameters for each sub-region within the water body; Anomaly areas are identified based on the water quality monitoring parameters, and the anomaly areas include multiple sub-regions; Based on the risk coefficients corresponding to the multiple sub-regions, the optimal processing device is determined from the multiple processing devices corresponding to the multiple sub-regions; Among the multiple processing devices corresponding to the multiple sub-regions, the optimal processing device is determined, based on the risk coefficients corresponding to the multiple sub-regions, including: Based on the risk coefficient, the scheduling priority vector of multiple processing devices corresponding to the multiple sub-regions is obtained; The set of candidate processing devices is determined based on the scheduling priority vector; Based on Bayes' theorem and the virtual game algorithm, the optimal processing device is determined from the set of candidate processing devices; The step of determining the optimal processing device from the candidate processing device set based on Bayes' theorem and virtual game algorithm includes: The initial probabilities corresponding to the virtual game algorithm are determined according to Bayes' theorem; Based on the initial probability, the candidate processing device set is divided into one main device and at least one secondary device; The process involves performing an iterative step, which includes determining the current scheduling strategy of the master device based on the current belief value and utility function of the slave device, updating the current belief value of the slave device, and continuing until a preset termination condition is met, and then outputting the optimal processing device.

2. The interaction method based on the intelligent agent of a water station system according to claim 1, characterized in that, Identifying abnormal areas based on the aforementioned water quality monitoring parameters includes: Based on the water quality monitoring parameters and the encoder, the feature vector corresponding to each sub-region is obtained; Based on the feature vector and the multilayer perceptron model, the association weight between any two sub-regions is obtained; The abnormal regions are identified based on the association weights.

3. The interaction method based on the intelligent agent of a water station system according to claim 2, characterized in that, The steps for determining the risk coefficient include: Based on the feature vector and the preset anomaly type library, determine the anomaly type corresponding to the anomaly region; The parameter set of the neural network model is determined according to the anomaly type. The parameter set is obtained by training based on historical data, and each anomaly type corresponds to a parameter set. Based on the parameter set and the feature vector, the risk coefficients corresponding to the multiple sub-regions are obtained.

4. The interaction method based on a smart agent in a water station system according to claim 1, characterized in that, The step of determining the current scheduling strategy of the master device based on the current belief value and utility function corresponding to the slave device includes: Determine the expected utility based on the current belief value and the utility function; The current scheduling strategy of the master device is determined based on the expected utility.

5. The interaction method based on a smart agent in a water station system according to claim 4, characterized in that, Updating the current belief value corresponding to the secondary device includes: Update the gain coefficient and credit score according to the current scheduling strategy; The current belief value corresponding to the secondary device is updated based on the gain coefficient, credit score, current iteration number, and indicator function.

6. The interaction method based on a smart agent in a water station system according to claim 3, characterized in that, If the preset anomaly type library does not contain an anomaly type corresponding to the anomaly region, the step of determining the risk coefficient includes: Based on the feature vector and the preset anomaly type library, determine the anomaly type group corresponding to the anomaly region; Obtain the anomaly type with the highest similarity to the feature vector from the anomaly type group, and use the parameter group corresponding to the anomaly type with the highest similarity as the initial parameter group of the neural network model; The initial parameter group is updated according to the anomaly type group to obtain the parameter group of the neural network model; Based on the parameter set and the feature vector, the risk coefficients corresponding to the multiple sub-regions are obtained.

7. The interaction method based on a smart agent in a water station system according to claim 6, characterized in that, The initial parameter set is updated according to the anomaly type group to obtain the parameter set of the neural network model, including: Remove the anomaly type that has the highest similarity to the feature vector from the anomaly type group to obtain the training set; The gradient for each parameter is obtained based on the data corresponding to the training set. The initial parameter set is updated based on the gradient to obtain the parameter set of the neural network model.

8. An interactive system based on a water station system intelligent agent, used to implement the interactive method based on a water station system intelligent agent as described in any one of claims 1-7, characterized in that, The interactive system includes: The perception layer agent is used to acquire water quality monitoring parameters corresponding to each sub-region in the water area, and to identify abnormal regions based on the water quality monitoring parameters. The abnormal regions include multiple sub-regions. The decision-making agent is used to determine the optimal processing device among multiple processing devices corresponding to multiple sub-regions within the abnormal region, based on the risk coefficients corresponding to multiple sub-regions within the abnormal region. The meta-agent is used to update the initial parameter group according to the anomaly type group to obtain the parameter group of the neural network model when the anomaly type corresponding to the anomaly region does not exist in the preset anomaly type library.

Images