CN-121997264-A - Intelligent inspection method and system based on AI

CN121997264ACN 121997264 ACN121997264 ACN 121997264ACN-121997264-A

Abstract

The invention provides an intelligent inspection method and system based on AI, comprising the steps of collecting environment data, poultry behavior data and equipment operation data, forming a multi-dimensional fusion data set containing space-time coordinates, constructing a digital twin model of a farm, updating the digital twin model and establishing a virtual-real mapping relation, inputting real-time data output by the digital twin model into a pre-trained multi-task learning AI model to obtain a comprehensive risk assessment report containing confidence coefficient and a time window, dynamically generating an optimal inspection path by adopting an ant colony algorithm in combination with a real-time risk thermodynamic diagram according to the comprehensive risk assessment report, adjusting inspection frequency and key areas in real time, feeding back new data and processing results obtained in the inspection process, updating parameters of the digital twin model and the multi-task learning AI model, and continuously optimizing an inspection strategy through a reinforcement learning algorithm. The invention can improve the accuracy of abnormality detection, greatly reduce loss and labor cost and reduce stress response of poultry.

Inventors

CHEN HONGWEI
DING NING
YANG LIUBIN
ZHOU WEN

Assignees

深圳旦品物联网科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. An intelligent inspection method based on AI is characterized by comprising the following steps: Synchronously acquiring environmental data, poultry behavior data and equipment operation data through multi-mode sensor networks deployed in different areas of a farm to form a multi-dimensional fusion data set containing space-time coordinates; Establishing a digital twin model of a farm based on the multi-dimensional fusion data set, and updating the environment state, the individual health state and the group behavior mode of the digital twin model in real time through a deep neural network to establish a virtual-real mapping relationship, wherein the digital twin model comprises four-dimensional incidence matrixes of individual identification, spatial position, physiological parameters and behavior characteristics of the birds; inputting real-time data output by the digital twin model into a pre-trained multi-task learning AI model, wherein the multi-task learning AI model adopts an attention mechanism weight to distribute the importance of different data sources; The multi-task learning AI model simultaneously executes three tasks of disease early warning, environment risk assessment and equipment fault prediction based on the real-time data, and outputs a comprehensive risk assessment report containing confidence coefficient and a time window; according to the comprehensive risk assessment report, an ant colony algorithm is adopted to dynamically generate an optimal inspection path in combination with a real-time risk thermodynamic diagram, inspection frequency and key areas are adjusted in real time, and the optimal inspection path comprehensively considers four optimization targets of risk level, geographic distance, inspection cost and bird stress response minimization; and feeding back and updating the parameters of the digital twin model and the multi-task learning AI model by new data and processing results obtained in the inspection process, and continuously optimizing the inspection strategy through a reinforcement learning algorithm to form a self-learning and self-optimizing closed-loop system.
2. The AI-based intelligent patrol method according to claim 1, wherein the step of constructing a digital twin model of a farm based on the multi-dimensional fusion dataset, updating the environmental state, the individual health state and the group behavior pattern in the digital twin model in real time through a deep neural network, and establishing a virtual-real mapping relationship comprises: Based on poultry identification information in the multi-dimensional fusion data set, distributing unique identification codes for each poultry individual, and establishing a four-dimensional incidence matrix frame containing individual identification dimensions, spatial position dimensions, physiological parameter dimensions and behavior feature dimensions; Constructing a three-dimensional virtual scene model comprising a building structure, equipment facilities and an environment area according to the actual physical layout of a farm, and creating a corresponding digital avatar for each bird in the virtual scene, wherein the digital avatar bears individual information in a four-dimensional incidence matrix to realize one-to-one mapping relation between physical birds and virtual objects; Constructing a multi-layer deep neural network special for updating the digital twin model, receiving real-time input from a multi-dimensional fusion data set by the multi-layer deep neural network, automatically identifying and analyzing environmental changes, individual health condition changes and group behavior mode changes through three core modules of a feature extraction layer, a state prediction layer and a parameter updating layer, and providing intelligent support for real-time updating of a four-dimensional incidence matrix; Establishing a double updating mechanism based on the combination of event triggering and timing polling, triggering corresponding updating of the digital twin model immediately when the poultry state is detected to be changed remarkably in the physical world, and refreshing the full data of the four-dimensional incidence matrix according to a preset time interval to ensure the high synchronism and consistency of the virtual model and physical reality; The accuracy and the deviation degree of virtual-real mapping are calculated by comparing measured data of a physical sensor with predicted data of a digital twin model, a feedback adjustment mechanism is used for continuously optimizing model parameters of a deep neural network and a data structure of a four-dimensional correlation matrix, and the representation capacity and the prediction precision of the digital twin model to the real world are improved.
3. The AI-based intelligent patrol method of claim 2, wherein the inputting the real-time data output by the digital twin model into a pre-trained multi-task learning AI model, wherein the multi-task learning AI model employs an attention mechanism to weight the importance of different data sources, comprises: receiving real-time data output by the digital twin model, wherein the real-time data comprises individual health indexes, abnormal group behavior metric values, environment comfort indexes and equipment running state parameters of birds in a four-dimensional incidence matrix, carrying out standardized processing and feature vector conversion on various data, and generating a multi-dimensional feature tensor in a unified format as standardized input of a multi-task learning AI model; And automatically calculating importance weights of all data sources in a current prediction task according to the real-time state and the history mode of the current farm by a pre-trained attention weight distribution module, wherein the weight distribution covers four main data sources of individual health data, group behavior data, environment monitoring data and equipment state data, so that the dynamic regulation and intelligent distribution of the importance of the data sources are realized.
4. The AI-based intelligent patrol method of claim 3, wherein the step of the multitask learning AI model simultaneously performs three tasks of disease pre-warning, environmental risk assessment, and equipment failure prediction based on the real-time data, and outputting a comprehensive risk assessment report including a confidence level and a time window, comprises: Performing feature separation on the real-time data subjected to attention weight modulation by the attention weight distribution module according to task correlation, and constructing three independent data processing channels of a disease early warning special channel, an environment risk assessment special channel and an equipment fault prediction special channel; The three data processing channels simultaneously start parallel computing processing, the disease early warning channel recognizes bird disease symptoms and judges the incidence probability through the deep learning network, the environment risk assessment channel analyzes the degree and trend of the environment parameter deviating from the normal range, the equipment fault prediction channel detects the equipment performance attenuation and the fault precursor characteristics, and each channel independently outputs the preliminary risk judgment result and the abnormal severity grade in the corresponding field; Calculating a corresponding confidence score for each preliminary risk judgment result based on the historical prediction accuracy, the current data quality and the model convergence state of each data processing channel, wherein the confidence score reflects the reliability degree of the prediction result, and meanwhile, the stability and the fluctuation range of the prediction result are evaluated through an uncertainty quantization algorithm to provide a credibility reference for subsequent decisions; According to the risk type and severity identified by each data processing channel, in combination with the development rule of similar conditions in historical data, predicting a time window and development trend of possible occurrence of a risk event, wherein the time window comprises three key time nodes of earliest possible occurrence time, latest necessary treatment time and optimal intervention time, and provides time dimension guidance information for routing inspection decision; And integrating the prediction results, the confidence level evaluations and the time window analyses of the three data processing channels to generate a structured comprehensive risk evaluation report containing the risk event types, the risk grades, the confidence scores, the prediction time windows, the influence range evaluations and recommended treatment measures, wherein the report organizes information according to a standardized format to ensure that the follow-up inspection path optimization algorithm can accurately analyze and use the evaluation results.
5. The AI-based intelligent inspection method according to claim 4, wherein the step of dynamically generating an optimal inspection path by adopting an ant colony algorithm in combination with a real-time risk thermodynamic diagram according to the comprehensive risk assessment report and adjusting the inspection frequency and the key area in real time, wherein the optimal inspection path comprehensively considers four optimization targets of risk level, geographic distance, inspection cost and bird stress response minimization, comprises: Based on the risk event type, the risk level and the influence range information in the comprehensive risk assessment report, constructing a real-time risk thermodynamic diagram on a plane map of the farm, enabling different risk levels to correspond to different thermodynamic values and color depths, simultaneously combining individual distribution density and activity area division of birds, and establishing a spatial mapping relation between the farm area and the risk level to provide a visual risk distribution basis for subsequent path planning; According to the current operation state and management priority of the farm, setting corresponding weight coefficients and constraint conditions for four optimization targets of risk level, geographic distance, inspection cost and bird stress response minimization, wherein the risk level targets require priority access to a high-risk area, the geographic distance targets have the shortest pursuit path total length, the inspection cost targets control manpower and time consumption, the stress response minimization targets reduce interference on normal behaviors of birds, and a parameter configuration framework of multi-target optimization is established; Initializing an ant colony algorithm on the basis of real-time risk thermodynamic diagram, deploying a plurality of virtual ants at the entrance position of a farm as path search agents, wherein each virtual ant carries four target optimization parameters and current risk thermodynamic diagram information, presetting higher pheromone concentration in a high-risk area through an pheromone initialization mechanism, and guiding the ants to preferentially explore a dangerous area needing key inspection; Starting a path searching process of the virtual ants, comprehensively calculating the optimal direction of the next movement by each ant according to the risk thermodynamic value of the current position, the distance cost for reaching each candidate area, the estimated inspection time and the poultry stress intensity, gradually converging to generate a candidate inspection path set considering four optimization targets through an iterative searching and pheromone updating mechanism, and selecting a path with the highest comprehensive score from the candidate inspection path set as the current optimal inspection path; According to the change conditions of the optimal routing inspection path and the real-time risk thermodynamic diagram, the routing inspection frequency and the residence time of each area are dynamically adjusted, the routing inspection frequency is increased and the inspection time is prolonged for the high-risk area, the routing inspection frequency is properly reduced for the low-risk area, meanwhile, a path real-time updating mechanism is established, when a new high-risk event or the original risk state is detected to be remarkably changed, the routing inspection path is immediately recalculated and adjusted, and the dynamic matching of the routing inspection strategy and the actual risk distribution is ensured.
6. The AI-based intelligent patrol method according to claim 5, wherein the step of feeding back new data and processing results obtained in the patrol process to update parameters of the digital twin model and the multitask learning AI model, continuously optimizing the patrol strategy by the reinforcement learning algorithm, and forming a self-learning self-optimizing closed-loop system comprises: In the process of executing the inspection path, collecting behavior track data of an inspector, actual discovery results of detection points, execution conditions of treatment measures and feedback of treatment effects in real time, recording stress reaction degree, environmental state change and equipment response conditions of birds in the process of inspecting, comparing and analyzing a prediction result before inspection with the actual discovered problems, and generating a comprehensive execution effect evaluation report comprising prediction accuracy, response timeliness and treatment effectiveness; Classifying and sorting feedback data in the comprehensive execution effect evaluation report, classifying virtual-real mapping deviation data, poultry state change data and environmental parameter correction data related to a digital twin model into a twin model update data set, classifying prediction error cases, newly discovered abnormal modes and risk evaluation deviation data related to a multi-task learning AI model into an AI model training data set, and providing standardized data input for subsequent model parameter update and optimization; Establishing a reinforcement learning rewarding mechanism based on a patrol effect, setting a successful prediction and timely discovery risk event as positive rewarding, setting missed detection, false alarm and resource waste as negative rewarding, quantitatively evaluating the goodness of a current patrol strategy through a rewarding function, constructing a strategy value evaluation system, comprehensively grading the rationality of a path selection decision, a frequency adjustment decision and a resource allocation decision, and providing a clear optimization direction and an objective function for the reinforcement learning algorithm; Based on classified feedback data, carrying out parameter updating on a digital twin model and a multi-task learning AI model by adopting an incremental learning method, wherein the digital twin model is used for mainly updating the mapping relation of a four-dimensional incidence matrix and the accuracy of a virtual-real synchronization mechanism, the multi-task learning AI model is used for mainly adjusting the weight parameters of three data processing channels and the distribution strategy of an attention mechanism, and a knowledge fusion technology is used for organically combining newly learned experience with historical knowledge so as to avoid catastrophic forgetting and improve the generalization capability of the model; Based on a reward mechanism and a strategy evaluation result, a reinforcement learning algorithm is used for continuously optimizing a patrol path generation strategy, a risk prediction strategy and a resource scheduling strategy, and through a strategy gradient updating and experience playback mechanism, the system can learn improvement from each patrol experience, gradually improve prediction accuracy, path optimization effect and emergency response capability, form an intelligent closed-loop system with self-learning, self-optimization and self-evolution, and realize continuous improvement of patrol performance and continuous enhancement of system adaptability.
7. The AI-based intelligent patrol method of claim 6, wherein the step of synchronously collecting environmental data, poultry behavior data, and equipment operation data through a multi-modal sensor network deployed in different areas of a farm to form a multi-dimensional fusion dataset comprising space-time coordinates comprises: Disposing multi-mode sensor nodes in a farm according to a gridding layout principle, configuring unique space coordinate identifications and equipment IDs for each sensor node, realizing unification of time references of all the sensor nodes through a wireless clock synchronization protocol, and ensuring time consistency of data acquisition; the sensor nodes synchronously acquire three basic data according to preset acquisition frequency, wherein the environment data comprise temperature, humidity, illumination intensity and air quality parameters, the poultry behavior data comprise individual positions, movement tracks, voiceprint characteristics and body temperature distribution, the equipment operation data comprise feeding system states, ventilation system parameters and lighting system working conditions, and each type of data carries an acquisition time stamp and sensor position coordinates; noise filtering, abnormal value detection and missing value compensation processing are carried out on the collected original data, invalid data are removed through a data integrity check sum sensor fault diagnosis algorithm, and the quality of the data entering fusion processing is ensured to meet the requirement of subsequent analysis; Adding a three-dimensional space-time coordinate mark for each effective data record, wherein the three-dimensional space-time coordinate mark comprises a two-dimensional space coordinate and a one-dimensional time coordinate, and establishing an accurate mapping relation between data and a physical space position and time node to form structured data with space positioning capability; And packaging and integrating the environmental data, the poultry behavior data and the equipment operation data which are marked by the space-time coordinate association according to a unified data format to generate a standardized multi-dimensional fusion data set containing data type identification, space-time coordinates, numerical content and quality level for subsequent digital twin modeling.
8. The AI-based intelligent patrol method according to claim 7, wherein in the step of synchronously collecting environmental data, poultry behavior data and equipment operation data through a multi-mode sensor network deployed in different areas of a farm to form a multi-dimensional fusion data set containing space-time coordinates, the multi-dimensional data fusion adopts an adaptive weight fusion algorithm, and specifically, the dynamic fusion weight of each sensor data is calculated through the following formula: Wherein, the Dynamic fusion weights for the mth sensor; The value range [0.1, 2.0] is the reliability influence factor of the mth sensor; Calculating the real-time reliability index of the mth sensor based on the historical accuracy and the current signal-to-noise ratio; the coverage area influence factor of the mth sensor is a value range [0.1, 1.5]; A space coverage efficiency index for the mth sensor; Is the total number of sensors; The coefficients are modified for the type characteristics of the mth sensor.
9. The AI-based intelligent inspection method of claim 8, wherein the multi-objective ant colony algorithm in the optimal inspection path is dynamically generated by adopting the ant colony algorithm in combination with the real-time risk thermodynamic diagram according to the comprehensive risk assessment report, and the state transition probability of ants among nodes is updated by the following formula: Wherein, the Is the probability of an ant transitioning from node u to node v; is the pheromone concentration on the sides (u, v); The visibility heuristic for nodes u through v is defined as ; Euclidean distance for nodes u to v; The risk of the node v is a risk attracting factor, and the higher the risk is, the larger the value is; The stress response intensity factor of the poultry at the node v; A set of allowed access nodes reachable from node u; 、、、 The relative importance parameters of pheromone, visibility, risk attraction and stress avoidance, respectively.
10. An AI-based intelligent patrol system for executing the AI-based intelligent patrol method according to any one of claims 1 to 9, comprising a multi-modal sensor network and a server deployed in different areas of a farm; the server is configured to: acquiring environment data, poultry behavior data and equipment operation data synchronously acquired by the multi-modal sensor network, and forming a multi-dimensional fusion data set containing space-time coordinates; Establishing a digital twin model of a farm based on the multi-dimensional fusion data set, and updating the environment state, the individual health state and the group behavior mode of the digital twin model in real time through a deep neural network to establish a virtual-real mapping relationship, wherein the digital twin model comprises four-dimensional incidence matrixes of individual identification, spatial position, physiological parameters and behavior characteristics of the birds; inputting real-time data output by the digital twin model into a pre-trained multi-task learning AI model, wherein the multi-task learning AI model adopts an attention mechanism weight to distribute the importance of different data sources; The multi-task learning AI model simultaneously executes three tasks of disease early warning, environment risk assessment and equipment fault prediction based on the real-time data, and outputs a comprehensive risk assessment report containing confidence coefficient and a time window; according to the comprehensive risk assessment report, an ant colony algorithm is adopted to dynamically generate an optimal inspection path in combination with a real-time risk thermodynamic diagram, inspection frequency and key areas are adjusted in real time, and the optimal inspection path comprehensively considers four optimization targets of risk level, geographic distance, inspection cost and bird stress response minimization; and feeding back and updating the parameters of the digital twin model and the multi-task learning AI model by new data and processing results obtained in the inspection process, and continuously optimizing the inspection strategy through a reinforcement learning algorithm to form a self-learning and self-optimizing closed-loop system.

Description

Intelligent inspection method and system based on AI Technical Field The invention relates to the technical field of safety, in particular to an intelligent inspection method and system based on AI. Background With the rapid development of scientific technologies such as the Internet of things, big data, artificial intelligence and the like, digitization and intellectualization have become the necessary trend of the development of the aquaculture. In modern farms, efficient inspection is critical to ensuring health of birds and livestock and improving the level of cultivation management. The traditional manual inspection mode has the problems of low efficiency, low accuracy, high cost, omission and the like. Disclosure of Invention The intelligent inspection method and system based on the AI are provided based on the problems, the abnormal detection accuracy is improved through the combination of multi-mode data fusion and digital twin technology, early warning of diseases and equipment faults is achieved through an AI multi-task learning model, loss is greatly reduced, inspection efficiency is improved through a self-adaptive path optimization algorithm, labor cost and bird stress response are reduced, a closed loop feedback mechanism enables the system to adapt to changes of different seasons and different cultivation stages, autonomous learning capacity is achieved, and through accurate monitoring and prediction, drug use is reduced, and ecological balance of a cultivation farm is maintained. In view of this, an aspect of the present invention proposes an AI-based intelligent patrol method, including: Synchronously acquiring environmental data, poultry behavior data and equipment operation data through multi-mode sensor networks deployed in different areas of a farm to form a multi-dimensional fusion data set containing space-time coordinates; Establishing a digital twin model of a farm based on the multi-dimensional fusion data set, and updating the environment state, the individual health state and the group behavior mode of the digital twin model in real time through a deep neural network to establish a virtual-real mapping relationship, wherein the digital twin model comprises four-dimensional incidence matrixes of individual identification, spatial position, physiological parameters and behavior characteristics of the birds; inputting real-time data output by the digital twin model into a pre-trained multi-task learning AI model, wherein the multi-task learning AI model adopts an attention mechanism weight to distribute the importance of different data sources; The multi-task learning AI model simultaneously executes three tasks of disease early warning, environment risk assessment and equipment fault prediction based on the real-time data, and outputs a comprehensive risk assessment report containing confidence coefficient and a time window; according to the comprehensive risk assessment report, an ant colony algorithm is adopted to dynamically generate an optimal inspection path in combination with a real-time risk thermodynamic diagram, inspection frequency and key areas are adjusted in real time, and the optimal inspection path comprehensively considers four optimization targets of risk level, geographic distance, inspection cost and bird stress response minimization; and feeding back and updating the parameters of the digital twin model and the multi-task learning AI model by new data and processing results obtained in the inspection process, and continuously optimizing the inspection strategy through a reinforcement learning algorithm to form a self-learning and self-optimizing closed-loop system. Optionally, the step of constructing a digital twin model of the farm based on the multi-dimensional fusion dataset, updating the environmental state, the individual health state and the group behavior mode in the digital twin model in real time through a deep neural network, and establishing the virtual-real mapping relationship comprises the following steps: Based on poultry identification information in the multi-dimensional fusion data set, distributing unique identification codes for each poultry individual, and establishing a four-dimensional incidence matrix frame containing individual identification dimensions, spatial position dimensions, physiological parameter dimensions and behavior feature dimensions; Constructing a three-dimensional virtual scene model comprising a building structure, equipment facilities and an environment area according to the actual physical layout of a farm, and creating a corresponding digital avatar for each bird in the virtual scene, wherein the digital avatar bears individual information in a four-dimensional incidence matrix to realize one-to-one mapping relation between physical birds and virtual objects; Constructing a multi-layer deep neural network special for updating the digital twin model, receiving real-time input from a multi-dimensional fusion data se