An AI intelligent inspection method and system based on deep reinforcement learning path planning

Abstract

The application relates to the technical field of intelligent community facility inspection, and discloses an AI intelligent inspection method and system based on deep reinforcement learning path planning, which converts batch group failure risk assessment reports into structured emergency event descriptors through a deterioration trigger type emergency event generator, enables an event perception meta-strategy network to identify batch group failure risks caused by internal facility deterioration, generates an emergency inspection initial strategy in real time, and generates an emergency inspection path through a deep reinforcement learning path planning model, thereby establishing a deterioration trigger type emergency switching channel between the deterioration analysis capability of preventive inspection and the immediate response capability of emergency inspection, and solving the technical problem that batch group failure risks cannot be timely triggered for path re-planning and cannot be identified and responded by an emergency inspection system.

Description

Technical Field

[0001] This invention relates to the field of smart community facility inspection technology, and more specifically, to an AI-based intelligent inspection method and system based on deep reinforcement learning path planning. Background Technology

[0002] In the AI-powered intelligent inspection scenario of smart communities, the inspection system performs two types of inspection tasks on various facilities within the community through deep reinforcement learning: In routine preventive inspections, the system identifies deterioration trends from facility operation data streams, models batch and spatial relationships between facilities based on heterogeneous graph structures, and uses deep reinforcement learning to generate inspection paths covering high-priority facilities; In emergency inspections, when the system receives an alarm for an emergency, it parses the event feature code based on the event type ontology and generates an emergency inspection path in real time through a pre-trained event-aware meta-policy network.

[0003] However, in smart communities with a large number of older facilities procured and installed in the same batch, if one of these facilities suddenly experiences a serious malfunction after long-term gradual deterioration, other facilities in the same batch face a high probability of experiencing a cluster failure due to their shared service life and batch-related defects. This necessitates inspecting all high-risk facilities within a few hours. Conventional inspection path planning uses a gradual adjustment cycle, only re-planning the path after priority changes in the next inspection cycle, which cannot meet the urgent need for troubleshooting cluster failures. The event type ontology for emergency inspection path planning only defines external emergencies such as fires and gas leaks, excluding the event type of cluster failure risk caused by internal facility deterioration. Therefore, it cannot trigger the generation of emergency inspection paths for distributed inspection of batch-related facilities.

[0004] Therefore, the lack of an emergency switching mechanism triggered by degradation between the two inspection modes results in a technical problem where the risk of batch-wide failures cannot be covered in a timely manner through the regular inspection path, nor can it be identified and responded to by the emergency inspection system. Summary of the Invention

[0005] This invention provides an AI-powered intelligent inspection method and system based on deep reinforcement learning path planning, which solves the technical problems in related technologies such as the inability to effectively cope with the risk of batch-based failures, the lack of dynamic inspection strategy generation mechanisms with event awareness, and the low efficiency of inspection path planning.

[0006] This invention provides an AI-powered intelligent inspection method based on deep reinforcement learning path planning, comprising the following steps: Perform time-series pattern matching on the community facility operation data stream. When the matching degree between the facility operation parameter sequence and the degradation characteristics of the severe fault stage in the degradation pattern library exceeds a preset threshold, output a fault confirmation signal containing the fault facility identifier and degradation type code. Based on the facility identifier in the fault confirmation signal, the set of facility nodes in the same batch is extracted along the batch association edge in the facility heterogeneity map. The remaining service life prediction distribution and degradation stage identifier of each facility in the same batch are obtained, and a batch group fault risk assessment report containing the individual risk level and spatial distribution location of each facility in the same batch is generated. Input the batch-based fault risk assessment report into the degradation-triggered emergency event generator and encode it into a structured emergency event descriptor that is compatible with the standard event coding specification of the event type ontology library; The structured emergency event descriptor is input into the event-aware meta-policy network, and the initial emergency inspection policy is output through the event type condition generation mechanism. The initial emergency inspection policy includes the set of facilities to be inspected and the inspection task type of each facility. The set of facilities to be inspected and their spatial distribution are input into a deep reinforcement learning path planning model. Combined with traffic constraints and facility accessibility conditions, an emergency inspection path covering all high-risk facilities is generated.

[0007] Furthermore, the time-series pattern matching employs a dynamic time warping algorithm. The input to the dynamic time warping algorithm is a sliding window segment of the facility's current operating parameters and typical signal sequences corresponding to each degradation type stored in the degradation pattern library. The output is a matching score. The dynamic time warping algorithm calculates the optimal time alignment path between the facility's current operating parameter sequence and the typical signal sequence to obtain the minimum cumulative distance between the two sequences. The minimum cumulative distance is normalized and used as the matching score. Each record in the degradation pattern library contains a degradation type code, a degradation stage identifier, and a typical signal sequence template for the corresponding degradation stage. The degradation stage identifier is divided into normal operation stage, early degradation stage, accelerated degradation stage, and severe failure stage according to the severity of degradation from low to high. The preset threshold is pre-set according to different facility types and degradation types and stored in the degradation pattern library and associated with the corresponding degradation type code.

[0008] Furthermore, the individual risk level is determined as follows: for each facility in the same batch, the individual risk score is calculated by multiplying the stage weight corresponding to its current deterioration stage identifier with the product of the average of its remaining service life and the reciprocal of the sum of the smallest positive numbers to prevent the denominator from being zero; among which, the severe failure stage, accelerated deterioration stage, early deterioration stage and normal operation stage correspond to decreasing stage weight values ​​respectively; based on the comparison between the individual risk score and the preset risk level threshold, each facility is divided into three levels: high risk, medium risk and low risk.

[0009] Furthermore, the processing steps of the degradation-triggered emergency event generator include: extracting the degradation type code of the faulty facility, the spatial distribution of the facility set in the same batch, and the individual risk level of each facility from the batch-based failure risk assessment report, as the original event elements to be encoded; mapping the original event elements to the field positions of the standard event coding specification in the event type ontology library, where the event type field is filled with the batch-based failure risk type identifier, the event source field is filled with the faulty facility identifier and its degradation type code, the event impact range field is filled with the spatial distribution of the facilities in the same batch, and the event severity field is assigned a value based on the proportion of high-risk facilities to the total number of facilities in the same batch; encoding the individual risk level of each facility in the same batch according to the attached object coding format in the standard event coding specification, generating a facility risk list associated with the structured emergency event descriptor, and outputting a complete structured emergency event descriptor.

[0010] Furthermore, the event-aware meta-policy network includes an event encoder, a condition generation module, and a policy output layer. The event encoder encodes the event type field, event impact range field, and event severity field in the structured emergency event descriptor into an event feature vector. The condition generation module uses the event feature vector as a gating signal and modulates the activation intensity of the facility risk list feature vector through a gating mechanism to generate an intermediate representation vector that integrates event context and facility risk information. The policy output layer decodes the intermediate representation vector into a binary vector of the set of facilities to be inspected and a probability distribution vector of the inspection task type for each facility. The inspection task types include two categories: comprehensive in-depth inspection and key parameter review. Comprehensive in-depth inspection tasks are assigned to facilities with high individual risk levels, and key parameter review tasks are assigned to facilities with medium individual risk levels. The event type condition generation mechanism refers to the event-aware meta-policy network using the type embedding vector corresponding to the event type field as a condition variable to adjust the activation weights of each policy parameter in the policy output layer, thereby generating differentiated inspection policies for different event types.

[0011] Furthermore, the deep reinforcement learning path planning model describes inspection path planning as a sequential decision-making process. The deep reinforcement learning path planning model employs a deep Q-network structure, including a state encoding layer, a policy network, and a value network. The state encoding layer encodes the current inspection position as a two-dimensional coordinate vector, and encodes the sets of visited facilities and unvisited facilities as binary vectors respectively. These are concatenated and mapped to a state vector through a fully connected layer. The policy network takes the state vector as input and outputs the probability distribution of each candidate action. The value network takes the state vector as input and outputs a state value estimate. The reward function of the deep reinforcement learning path planning model is calculated for each step... The instant reward is accumulated in each step. The instant reward for each step is a linear combination of the negative value of the distance penalty, the positive value of the task completion reward, and the negative value of the constraint violation penalty. The distance penalty is the distance between the current facility and the next facility to be visited multiplied by the distance penalty coefficient. The task completion reward is the reward value for completing the inspection task of the next facility to be visited multiplied by the task completion reward coefficient. The constraint violation penalty is the number of times the passage constraint is violated in this step multiplied by the constraint violation penalty coefficient. The task completion reward value is determined according to the individual risk level of the facility, and the reward value corresponding to high-risk facilities is greater than the reward value corresponding to medium-risk facilities.

[0012] Furthermore, after outputting the initial emergency inspection strategy and before inputting the set of facilities to be inspected into the deep reinforcement learning path planning model, the process also includes: sorting the facilities in the set of facilities to be inspected based on the variance of the predicted distribution of the remaining useful life of each batch of facilities, placing the facilities with the highest prediction uncertainty in the priority inspection position, and generating an uncertainty-driven inspection task sequence; using the uncertainty-driven inspection task sequence as the facility priority constraint input of the deep reinforcement learning path planning model, and implementing the facility priority constraint by adding an uncertainty reward term to the reward function, wherein the uncertainty reward term is the predicted variance of the remaining useful life of the facilities to be visited multiplied by the uncertainty reward coefficient.

[0013] Furthermore, after generating the emergency inspection path, the process also includes: during the execution of the emergency inspection path, collecting deep status data for each inspected facility, updating the predicted distribution of the facility's remaining service life based on the measured data; when the measured risk score of an inspected facility is lower than the batch warning threshold, marking the facility as low-risk and removing it from the remaining inspection task sequence, re-inputting the updated remaining inspection task sequence and the spatial distribution of the remaining uninspected facilities into the deep reinforcement learning path planning model to generate an updated inspection path for the remaining uninspected facilities; the batch warning threshold is a preset proportion of the average of the individual risk scores of all facilities in the same batch. The status update data of inspected facilities is backpropagated along the batch association edges in the facility heterogeneity graph to the nodes of uninspected facilities. For the deviation between the measured and predicted degradation levels of inspected facilities, the deviation is weighted according to the similarity of degradation stages of each facility in the batch, and the average remaining useful life prediction of uninspected facilities is corrected. The degradation stage similarity is determined based on the difference between the degradation stage identifiers of inspected and uninspected facilities. Based on the corrected remaining useful life prediction distribution, the individual risk level of uninspected facilities is recalculated. If the individual risk level of an uninspected facility drops to low risk, it is removed from the remaining inspection task sequence and the replanning of the remaining path is triggered.

[0014] Furthermore, after all emergency inspections are completed, the process also includes: writing the triggering degradation mode of this batch of failures into the degradation mode library. The triggering degradation mode includes the facility type that caused this batch of failures, the degradation type code, and the corresponding operating parameter signal sequence template. The operating parameter signal sequence template is obtained by extracting and summarizing the time series of operating parameters of the faulty facility before the failure occurred. The batch propagation path and actual detection results of this emergency inspection are written into the event type ontology library. The batch propagation path includes the access order, access time, and measured risk score of each facility. The actual detection results include the measured degradation stage identifier, the average measured remaining service life, and the identifier of whether a failure has occurred for each facility in the same batch. The association record between the batch of failure risk type and the corresponding inspection strategy is established using the batch of failure risk type identifier as the index key.

[0015] This invention also proposes an AI intelligent inspection system based on deep reinforcement learning path planning, comprising: The fault confirmation module is used to perform time-series pattern matching on the community facility operation data stream. When the matching degree between the facility operation parameter sequence and the degradation characteristics of the severe fault stage in the degradation pattern library exceeds a preset threshold, a fault confirmation signal containing the fault facility identifier and degradation type code is output. The batch risk assessment module is used to extract the set of facility nodes in the same batch along the batch association edge in the facility heterogeneity map based on the facility identifier in the fault confirmation signal, obtain the predicted distribution of the remaining service life and the degradation stage identifier of each facility in the same batch, and generate a batch group fault risk assessment report containing the individual risk level and spatial distribution location of each facility in the same batch. The emergency event generation module is used to input the batch-based fault risk assessment report into the degradation-triggered emergency event generator and encode it into a structured emergency event descriptor that is compatible with the standard event coding specification of the event type ontology library. The emergency strategy generation module is used to input structured emergency event descriptors into the event-aware meta-policy network and output the initial emergency inspection strategy through the event type condition generation mechanism. The initial emergency inspection strategy includes the set of facilities to be inspected and the inspection task type of each facility. The path planning module is used to input the set of facilities to be inspected and the spatial distribution of each facility into the deep reinforcement learning path planning model, and generate emergency inspection paths covering all high-risk facilities by combining traffic constraints and facility accessibility conditions.

[0016] The beneficial effects of this invention are as follows: This invention uses uncertainty-driven inspection task sequencing to enable the deep reinforcement learning path planning model to prioritize visits to facilities with the largest variance in remaining useful life prediction, thereby eliminating the uncertainty of facilities with the largest risk blind spots within a limited inspection time.

[0017] By updating the predicted distribution of remaining service life based on measured data during the inspection process, and using the shared inference of backpropagation of batch-related edges to correct the risk assessment of uninspected facilities, the inspection scope is dynamically reduced and the remaining path is replanned, thus avoiding the consumption of inspection resources caused by indiscriminately inspecting all facilities in the same batch.

[0018] After the emergency inspection is completed, the triggered degradation mode and the actual detection results are written into the degradation mode library and the event type ontology library. This enables the AI ​​intelligent inspection system to directly identify and respond when it encounters similar batch failure scenarios in the future, and gradually accumulate path planning preset knowledge for batch failure scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart of the AI ​​intelligent inspection method based on deep reinforcement learning path planning provided in an embodiment of the present invention; Figure 2 This is a time-series evolution trend diagram of the degradation signal of faulty facilities provided in an embodiment of the present invention; Figure 3 This is a distribution chart of cumulative operating time and degradation stage of facilities in the same batch provided in an embodiment of the present invention; Figure 4This is a distribution map showing the predicted remaining service life of facilities in the same batch, provided in an embodiment of the present invention. Figure 5 This is a batch mass transmission failure individual risk score assessment chart provided in an embodiment of the present invention; Figure 6 This is an uncertainty-driven inspection task priority ranking diagram provided in an embodiment of the present invention; Figure 7 This is a spatial distribution and access sequence diagram of the emergency inspection path provided in an embodiment of the present invention; Figure 8 This is an analysis chart of the cumulative travel distance of the emergency inspection route provided in an embodiment of the present invention; Figure 9 This is a statistical chart showing the allocation of inspection task types and estimated time consumption provided in an embodiment of the present invention. Detailed Implementation

[0020] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0021] At least one embodiment of the present invention discloses an AI-based intelligent inspection method based on deep reinforcement learning path planning, such as... Figure 1 As shown, it includes the following steps: Step 1: Perform time-series pattern matching on the community facility operation data stream and output a fault confirmation signal; The system acquires real-time operational data streams from various facilities within the community and performs time-series pattern matching between the current operational parameters of each facility and typical degradation signal sequences in the degradation pattern library. When the matching degree between the operational parameter sequence of a facility and the degradation characteristics of a severe fault stage in the degradation pattern library exceeds a preset threshold, a fault confirmation signal is output. The fault confirmation signal includes the identifier of the faulty facility and the degradation type code.

[0022] It should be noted that the above-mentioned operational data stream refers to the time series of operational parameters continuously reported by the sensors of each facility at a preset sampling frequency, including but not limited to parameters such as vibration amplitude, temperature, current waveform, and operational noise spectrum.

[0023] It should be noted that the above time-series pattern matching employs a dynamic time warping algorithm. The input to this algorithm is a sliding window segment of the facility's current operating parameters and typical signal sequences corresponding to each degradation type stored in the degradation pattern library. The output is a matching score. Each record in the degradation pattern library contains a degradation type code, a corresponding degradation stage identifier, and a typical signal sequence template for that degradation stage. The degradation stage identifiers are categorized in ascending order of degradation severity: normal operation stage, early degradation stage, accelerated degradation stage, and severe failure stage. When the matching result corresponds to the severe failure stage, the facility is determined to have experienced a severe failure.

[0024] Furthermore, the aforementioned dynamic time warping algorithm calculates the optimal time alignment path between the current operating parameter sequence of the facility and the typical signal sequence in the degradation mode library, obtains the minimum cumulative distance between the two sequences, and normalizes the minimum cumulative distance as the matching score.

[0025] Furthermore, the aforementioned preset thresholds are pre-set according to different facility types and degradation types, and stored in the degradation pattern library associated with the corresponding degradation type codes. For elevator facilities, the matching degree threshold is set to 0.85; for distribution box facilities, the matching degree threshold is set to 0.90; and for water pump facilities, the matching degree threshold is set to 0.80. When the matching degree score output by the dynamic time warping algorithm is greater than the preset threshold for the corresponding facility type and degradation type, a fault confirmation signal is triggered.

[0026] Step 2: Extract facilities in the same batch based on the facility heterogeneity map and generate a batch-wide failure risk assessment report; Based on the facility identifier in the fault confirmation signal, the corresponding facility node is located in the facility heterogeneity graph. The set of all facility nodes belonging to the same batch as the faulty facility is extracted by traversing along the batch association edges. The predicted remaining useful life distribution and current degradation stage identifier of each facility node in the same batch are obtained. Combined with the spatial distribution of each facility, a batch-wide fault risk assessment report is generated.

[0027] The batch-wide failure risk assessment report includes the facility identification, individual risk level, spatial distribution location, and predicted remaining useful life distribution of each facility in the same batch.

[0028] It should be noted that the above-mentioned facility heterogeneity diagram is a graph structure containing multiple types of nodes and edges. Node types include facility nodes, region nodes, and batch nodes. Edge types include batch-related edges (connecting facility nodes and batch nodes within the same procurement batch), spatially related edges (connecting facility nodes and region nodes located in the same physical region), and functionally related edges (connecting pairs of facility nodes with functional dependencies). The attributes of each facility node include facility type, installation date, batch number, cumulative runtime, current degradation stage identifier, and predicted remaining useful life distribution.

[0029] Furthermore, the aforementioned remaining useful life prediction distribution is a probability distribution obtained through a pre-trained facility life prediction model. The facility life prediction model outputs the mean and variance of the facility's remaining useful life based on the facility's historical operating data, current deterioration stage identifier, and cumulative operating time, representing the expected value and prediction uncertainty of the remaining useful life.

[0030] It should be noted that the above-mentioned individual risk levels are calculated as follows: for each facility in the same batch, based on the mean of its predicted remaining useful life distribution. The stage weight corresponding to the current degradation stage identifier Calculate individual risk scores : ; in, For the index number of facilities in the same batch, For the first The projected average remaining useful life of facilities in the same batch, in hours. The current stage of deterioration of the facility is identified by the stage weight, which is a dimensionless coefficient. To prevent extremely small positive numbers with a denominator of zero, the value is set to 0.01 hours. The dimension of is the reciprocal of time, representing the risk intensity per unit time.

[0031] Furthermore, the aforementioned stage weights The value is determined based on the degradation stage identifier: the weight corresponding to the severe failure stage is 1.0, the weight corresponding to the accelerated degradation stage is 0.7, the weight corresponding to the early degradation stage is 0.4, and the weight corresponding to the normal operation stage is 0.1.

[0032] Furthermore, based on individual risk scores By comparing each facility with a preset risk level threshold, it is classified into three levels: high-risk, medium-risk, and low-risk. When, it is judged as high risk; when When, it is judged as medium risk; when When the risk level is low, the facility is classified as such. These risk level thresholds are determined based on statistical analysis of historical failure data, ensuring that high-risk facilities have a greater than 70% probability of failure within the next 24 hours, and medium-risk facilities have a greater than 50% probability of failure within the next 72 hours.

[0033] Step 3: Convert the batch mass mailing failure risk assessment report into a structured emergency event descriptor; The batch-wide failure risk assessment report is input into the degradation-triggered emergency event generator, which encodes the type of faulty facility, the spatial distribution set of facilities in the same batch, and the individual risk level of each facility into a structured emergency event descriptor. The format of the structured emergency event descriptor is compatible with the standard event coding specifications of the event type ontology library, enabling the emergency inspection route planning process, which was originally only for external emergencies, to identify and respond to the risk of batch-wide failures caused by internal facility degradation.

[0034] It should be noted that the processing procedure of the aforementioned degradation-triggered emergency event generator includes the following sub-steps: Step 301: Extract the degradation type code of the faulty facility, the spatial distribution of the facility set in the same batch, and the individual risk level of each facility from the batch failure risk assessment report as the original event elements to be coded.

[0035] Step 302: Based on the field definitions in the standard event coding specification of the event type ontology, map the above-mentioned original event elements to the respective field positions of the standard event coding. Specifically, the event type field is filled with a batch-wide failure risk type identifier, the event source field is filled with a faulty facility identifier and its degradation type code, the event impact range field is filled with the set of spatial distribution locations of facilities in the same batch, and the event severity field is assigned a value based on the proportion of high-risk facilities to the total number of facilities in the same batch.

[0036] Step 303: Encode the individual risk level of each facility in the same batch according to the auxiliary object coding format in the standard event coding specification, generate a facility risk list associated with the structured emergency event descriptor, and output the complete structured emergency event descriptor.

[0037] It should be noted that the aforementioned standard event coding specification is a predefined event description format standard in the event type ontology library, including event type field, event source field, event impact scope field, event severity field, and associated object coding field. By encoding the risk of batch-based failures into structured descriptors conforming to the standard event coding specification, the event-aware meta-policy network can process structured emergency event descriptors as valid input without retraining.

[0038] Furthermore, the above-mentioned event severity field is assigned a value by calculating the proportion of high-risk facilities to the total number of facilities in the same batch. ,when When the severity of the event is assessed, it is assigned the emergency level; when When, the severity of the event is assigned an importance level; when At that time, the severity of the event is assigned the "normal" level.

[0039] Step 4: Generate an initial emergency inspection strategy based on structured emergency event descriptors; The structured emergency event descriptor is input into the event-aware meta-policy network, and an event type condition generation mechanism is used to instantly output an initial emergency inspection policy covering all high-risk facilities in the same batch. The initial emergency inspection policy includes the set of facilities to be inspected and the inspection task type for each facility.

[0040] It should be noted that the aforementioned event-aware meta-policy network is a pre-trained conditional policy generation network, whose components include an event encoder, a conditional generation module, and a policy output layer. The event encoder receives a structured emergency event descriptor and encodes the event type, event impact scope, and event severity fields into an event feature vector. The conditional generation module uses the event feature vector as conditional input and, combined with the individual risk level of each facility in the facility risk list, generates an intermediate representation for inspection task allocation. The policy output layer decodes the intermediate representation into a specific set of facilities to be inspected and the corresponding inspection task type for each facility. The inspection task types include two categories: comprehensive in-depth inspection and key parameter review. For facilities with a high individual risk level, a comprehensive in-depth inspection task is assigned; for facilities with a medium individual risk level, a key parameter review task is assigned.

[0041] Furthermore, the aforementioned event encoder employs a multilayer perceptron structure, converting each field in the structured emergency event descriptor into a fixed-dimensional vector representation through an embedding layer, and then fusing them into a unified event feature vector through a fully connected layer. The condition generation module uses a gating mechanism, employing the event feature vector as a gating signal to modulate the activation intensity of the facility risk list feature vector, generating an intermediate representation vector that fuses the event context and facility risk information. The policy output layer employs a multilayer perceptron structure, mapping the intermediate representation vector through a fully connected layer into a binary vector of the set of facilities to be inspected and a probability distribution vector of the inspection task type for each facility.

[0042] Furthermore, each dimension of the binary vector representing the set of facilities to be inspected corresponds to one facility, with a value of 1 indicating that the facility needs inspection and a value of 0 indicating that it does not need inspection. The probability distribution vector of inspection task types is output through a softmax activation function to determine the probability distribution of inspection task types for each facility, and the task type with the highest probability is selected as the inspection task type for that facility. The output dimension of the fully connected layer matches the total number of facilities and the number of inspection task types in the community.

[0043] It should be noted that the aforementioned event type condition generation mechanism refers to the event-aware meta-policy network using the type embedding vector corresponding to the event type field as a condition variable to adjust the activation weights of each policy parameter in the policy output layer, thereby generating differentiated inspection policies for different event types. Because the structured emergency event descriptor format is compatible with the standard event coding specifications of the event type ontology, the event-aware meta-policy network can process batch-based fault risk type identifiers as a new event type condition, instantly generating the corresponding initial emergency inspection policy.

[0044] Step 5: Generate emergency inspection paths based on a deep reinforcement learning path planning model; The set of facilities to be inspected and their spatial distribution in the initial emergency inspection strategy are input into a deep reinforcement learning path planning model. Combined with access constraints and facility accessibility conditions from the community map, an emergency inspection path covering all high-risk facilities is generated. This emergency inspection path includes the order in which facilities are visited and the routes between adjacent facilities.

[0045] It should be noted that the deep reinforcement learning path planning model described above represents inspection path planning as a sequential decision-making process. The state space of the deep reinforcement learning path planning model is defined as a combination of the current inspection position, the set of visited facilities, and the set of unvisited facilities. The action space is defined as selecting the next facility to be visited from the current position. The reward function comprehensively considers the negative value of the path travel distance, the positive reward for completing each facility inspection task, and the penalty value for violating passage constraints. Based on the above state space, action space, and reward function, the deep reinforcement learning path planning model outputs the facility visit sequence and corresponding travel route that maximizes the cumulative reward.

[0046] Furthermore, the aforementioned deep reinforcement learning path planning model employs a deep Q-network structure, including a state encoding layer, a policy network, and a value network. The state encoding layer encodes the current inspection location, the set of visited facilities, and the set of unvisited facilities into a state vector. The policy network uses a multilayer perceptron structure, taking the state vector as input and outputting the probability distribution of each candidate action. The value network also uses a multilayer perceptron structure, taking the state vector as input and outputting a state value estimate. The output layer of the policy network is a fully connected layer, with an output dimension matching the number of candidate facilities in the action space. It outputs the probability distribution of each candidate action through a softmax activation function, selecting the action with the highest probability as the next facility to be visited. The output layer of the value network is also a fully connected layer with an output dimension of 1, outputting a scalar estimate of the current state's value.

[0047] Furthermore, the aforementioned state encoding layer encodes the current inspection location as a two-dimensional coordinate vector, and encodes the set of visited facilities and the set of unvisited facilities as binary vectors respectively. Each dimension corresponds to a facility, with a value of 1 indicating that the facility has been visited or not, and a value of 0 indicating the opposite state. These three vectors are concatenated and mapped to a fixed-dimensional state vector through a fully connected layer.

[0048] It should be noted that the aforementioned traffic constraints include one-way traffic restrictions on roads within the community, area traffic bans during specific time periods, and size restrictions for inspection equipment. Facility accessibility conditions include the inspectable time window for each facility and the access permissions required to enter the area where the facility is located.

[0049] Furthermore, the specific form of the above reward function is: when the inspection personnel leave the facility Move to facility The instant reward for this step The calculation is as follows: ; in, This is the index number of the current facility. Assign an index number to the next facility to be visited. This is the index number for the inspection steps. For facilities With facilities The distance traveled between them, in meters. To complete the facilities The rewards for completing inspection tasks are dimensionless values. This represents the number of times the passage constraint was violated in this step, and is a dimensionless count value. This is the distance penalty coefficient, with a value of 0.1. This is the task completion reward coefficient, with a value of 1.0. To constrain the penalty coefficient for violations, a value of 5.0 is used, and all coefficients are dimensionless. Task completion reward. According to the facilities Individual risk levels are determined as follows: the reward for completing a task is 10 for high-risk facilities, 5 for medium-risk facilities, and 2 for low-risk facilities.

[0050] Furthermore, the above reward function incorporates a distance penalty term. This reflects the constraints of the time dimension, because of the distance traveled. The distance penalty term is directly proportional to the travel time of the inspection personnel between facilities. The longer the travel distance, the longer the travel time. This distance penalty term prompts the deep reinforcement learning path planning model to prioritize the closer facilities within a limited time window in order to improve the inspection coverage efficiency per unit time.

[0051] Furthermore, the aforementioned goal of maximizing cumulative rewards is achieved by applying real-time rewards at each step of the entire inspection path execution process. The cumulative realization, cumulative rewards The calculation is as follows: ; in, This is the index number for the inspection steps. This determines the total number of steps required to complete all facility visits to be inspected. A deep reinforcement learning path planning model trains a policy network and a value network to generate cumulative rewards corresponding to the facility visit sequences. The maximum value is reached, thereby generating an emergency inspection route with the shortest travel distance, priority coverage of high-risk facilities, and the fewest constraint violations, while meeting the conditions of traffic constraints and facility accessibility.

[0052] In this embodiment of the application, in order to prioritize the coverage of facilities with the largest risk blind spots by the emergency inspection path, after outputting the initial emergency inspection strategy in step 4 and before inputting the set of facilities to be inspected into the deep reinforcement learning path planning model, the following step is also included: predicting the variance of the distribution based on the remaining useful life of facilities in different batches. The facilities to be inspected in the initial emergency inspection strategy are ranked, and those with the highest predictive uncertainty are prioritized for inspection, generating an uncertainty-driven inspection task sequence. For the index number of facilities in the same batch, For the first The variance of the predicted remaining useful life of facilities in the same batch is expressed in squared hours. A larger variance indicates a higher uncertainty in the actual deterioration state of the facility, necessitating priority for obtaining in-depth status data through on-site inspections to eliminate risk blind spots. An uncertainty-driven inspection task sequence is used as the facility priority constraint input to the deep reinforcement learning path planning model, enabling the model to prioritize visits to facilities with high uncertainty during path optimization.

[0053] Furthermore, the aforementioned facility priority constraint is implemented by modifying the reward function of the deep reinforcement learning path planning model. An uncertain reward term is added to the original reward function. ,in, The index number of the facility to be accessed. Facilities to be visited The variance of the remaining useful life prediction, in the square of hours. The uncertainty reward coefficient is set to 2.0 and is dimensionless. The modified reward function is as follows: ; in, This is the index number for the inspection steps. This is the index number of the current facility. Assign an index number to the next facility to be visited. For facilities With facilities The distance traveled between them To complete the facilities The rewards earned from completing inspection missions. This represents the number of times the passage constraint was violated during this step. This is the distance penalty coefficient. The reward coefficient for task completion. To constrain the penalty coefficient for violations, For uncertain reward coefficients, Facilities to be visited The remaining useful life prediction variance is used to optimize the deep reinforcement learning path planning model. By adding an uncertainty reward term, the model is made to prioritize facilities with larger remaining useful life prediction variances when selecting the next facility to be visited, thus achieving priority coverage of facilities with the largest risk blind spots.

[0054] In this embodiment, to dynamically reduce the inspection scope and replan the remaining path based on measured data during the inspection process, after generating the emergency inspection path in step 5, the following steps are also included: During the execution of the emergency inspection path, depth status data is collected for each inspected facility, and the predicted distribution of the facility's remaining service life is updated based on the measured data. When the measured risk score of an inspected facility is significantly lower than the batch warning threshold, the facility is marked as low-risk and removed from the remaining inspection task sequence. The updated remaining inspection task sequence and the spatial distribution of the remaining uninspected facilities are re-input into the deep reinforcement learning path planning model to generate an updated inspection path for the remaining uninspected facilities. Here, the measured risk score is an individual risk score recalculated based on the measured depth status data. The batch warning threshold is a threshold set based on the overall risk level of facilities in the same batch when triggering this emergency inspection.

[0055] Furthermore, the aforementioned deep-state data refers to detailed operating parameters of the facility collected through on-site precision testing equipment, including but not limited to parameters such as high-frequency vibration signals, thermal imaging temperature distribution, electrical insulation impedance, and mechanical wear. When updating the remaining useful life prediction distribution based on the deep-state data, the measured parameters are input into a pre-trained facility life prediction model to obtain the mean and variance of the remaining useful life prediction based on the measured data, replacing the original remaining useful life prediction distribution based on historical data.

[0056] Furthermore, the aforementioned batch warning threshold is determined as follows: when generating the batch-wide failure risk assessment report in step 2, the individual risk scores of all facilities in the same batch are calculated. average ,Will 50% is used as the batch warning threshold, i.e., the batch warning threshold. .in, For the index number of facilities in the same batch, For the first Individual risk scores for facilities in the same batch in Taiwan. This is the average of the individual risk scores for all facilities in the same batch. This is the batch warning threshold. It is based on the measured risk score of the inspected facility. If the measured risk of a facility is determined to be significantly lower than the batch warning threshold, the facility will be marked as low risk and removed from the remaining inspection task sequence.

[0057] In this embodiment, to further utilize the measured data of inspected facilities to correct the risk assessment accuracy of uninspected facilities, in addition to the above-mentioned dynamic reduction of the inspection scope, the following steps are also included: The status update data of inspected facilities is backpropagated along the batch association edges in the facility heterogeneity graph to the uninspected facility nodes; shared inference is performed using the degradation homogeneity of facilities within the batch; and the predicted distribution of the remaining useful life of uninspected facilities is corrected. Specifically, for the deviation between the measured degradation degree and the predicted degradation degree in inspected facilities, the deviation value is weighted according to the similarity of the degradation stages of each facility within the batch, and the average predicted remaining useful life of uninspected facilities is adjusted. and variance Output the verified and corrected residual risk distribution. The similarity of degradation stages is calculated as follows: for inspected facilities... and uninspected facilities If both degradation stages have the same current degradation stage identifier, the degradation stage similarity value is 1; if their degradation stage identifiers are adjacent (e.g., one is in the accelerated degradation stage and the other is in the severe failure stage), the degradation stage similarity value is 0.5; if their degradation stage identifiers are separated by two or more stages, the degradation stage similarity value is 0. Uninspected facilities. The adjustment to the predicted mean of remaining useful life is the weighted average of the products of the deviation values ​​of all inspected facilities and the similarity of their corresponding deterioration stages. Based on the validated and corrected remaining risk distribution, the individual risk level of uninspected facilities is recalculated. If the individual risk level of some uninspected facilities decreases from high risk to low risk, they are removed from the remaining inspection task sequence, and the deep reinforcement learning path planning model is triggered to replan the remaining paths.

[0058] Furthermore, the aforementioned degradation homology refers to the similarity in degradation patterns exhibited by facilities in the same batch due to shared manufacturing processes, material batches, and service environments. When making shared inferences based on degradation homology, it is assumed that the measured degradation deviation of inspected facilities reflects, to some extent, the systematic impact of batch-wide common defects on degradation prediction. Therefore, the measured degradation deviation can be weighted according to the similarity of degradation stages and propagated to uninspected facilities to correct their remaining service life prediction distribution.

[0059] Furthermore, the deviation between the measured degree of degradation and the predicted degree of degradation is defined as: for the inspected facilities Its predicted average remaining useful life is The average remaining service life, recalculated based on measured depth data, is Then the deviation value The unit is hours. For the index number of the inspected facilities, For inspected facilities The predicted average remaining useful life, For inspected facilities The average remaining service life recalculated based on measured data For inspected facilities Deviation value. Uninspected facilities. The adjustment to the predicted mean of remaining useful life is calculated as follows: ; in, For the index number of the uninspected facility, For the index number of the inspected facilities, For the collection of inspected facilities, For inspected facilities Uninspected facilities The similarity of the degradation stages between them is represented by a dimensionless coefficient. For inspected facilities The deviation value, in hours. Uninspected facilities The adjustment to the predicted mean remaining useful life, in hours. Uninspected facilities. The corrected mean remaining useful life prediction is The unit is hours. Uninspected facilities Mean of remaining useful life forecast before correction Uninspected facilities Corrected mean remaining useful life forecast.

[0060] In this embodiment, to expand the AI ​​intelligent inspection system's pre-defined knowledge of path planning for batch-wide failure scenarios, after all emergency inspections are completed, the following steps are also included: The triggering degradation mode of this batch failure is written into a degradation mode library. The triggering degradation mode includes the facility type that caused the batch-wide failure, the degradation type code, and the corresponding operating parameter signal sequence template. The batch propagation path and actual detection results of this emergency inspection are written into an event type ontology library, including the event description template corresponding to the batch-wide failure risk type identifier, the actual detected failure type of each facility in the same batch, and the actual risk level distribution. The above writing operations enable the degradation mode library and the event type ontology library to directly match and identify similar batch-wide failure scenarios in subsequent inspections, reducing the reliance on the encoding conversion of degradation-triggered emergency event generators.

[0061] Furthermore, the aforementioned operating parameter signal sequence template for triggering the degradation mode is obtained by extracting and summarizing the time series of operating parameters of the facility before the failure. The operating parameter signal sequence template contains the typical changing trends and characteristic peak positions of key parameters, serving as a matching benchmark for subsequent similar failures.

[0062] Furthermore, the aforementioned batch propagation path refers to the actual facility access sequence executed during this emergency inspection, including the access order, access time, and measured risk score of each facility. The actual detection results include the measured degradation stage identifier, the average measured remaining service life, and an identifier indicating whether a fault has occurred for each facility in the same batch. When writing the batch propagation path and actual detection results into the event type ontology, the batch-wide fault risk type identifier is used as the index key to establish an association record between the batch-wide fault risk type and the corresponding inspection strategy. This allows the event-aware meta-policy network to directly retrieve historical inspection strategies for reference when encountering similar events subsequently.

[0063] This implementation introduces a degradation-triggered emergency event generator in step 3, automatically converting the batch-wide failure risk assessment report identified by the routine inspection degradation analysis capability in step 2 into a structured emergency event descriptor that conforms to the standard event coding specifications of the event type ontology. Because the format of the structured emergency event descriptor is consistent with the existing external emergency event coding format in the event type ontology, the event-aware meta-policy network in step 4 can identify and process batch-wide failure risks as legitimate event types, generating an initial emergency inspection policy immediately without waiting for the next routine inspection scheduling cycle. Therefore, this implementation establishes a degradation-triggered emergency switching path between the degradation analysis capability of preventive inspection path planning and the immediate response capability of emergency inspection path planning, overcoming the dilemma that batch-wide failure risks cannot trigger timely inspection path replanning under progressive scheduling cycles and cannot be identified and responded to by the emergency inspection system.

[0064] Furthermore, by prioritizing inspection tasks based on uncertainty, the deep reinforcement learning path planning model prioritizes visits to facilities with the largest variance in remaining service life predictions when generating emergency inspection paths. This eliminates uncertainty in facilities with the largest risk blind spots within the limited inspection time. Updates to the remaining service life prediction distribution using measured data during inspection execution, and corrections to the risk assessment of uninspected facilities through shared inference using backpropagation of batch-related edges, dynamically reduce the inspection scope and trigger replanning of remaining paths. This avoids the resource consumption associated with indiscriminately inspecting all facilities in the same batch. After emergency inspections are completed, the triggered degradation modes and actual detection results are written into the degradation mode library and event type ontology library. This enables the AI ​​intelligent inspection system to directly identify and respond to similar batch-wide failure scenarios in the future, gradually accumulating pre-defined path planning knowledge for batch-wide failure scenarios.

[0065] In one embodiment of the present invention, an AI intelligent inspection system based on deep reinforcement learning path planning is proposed, comprising the following modules: The fault confirmation module is used to perform time-series pattern matching on the community facility operation data stream. When the matching degree between the facility operation parameter sequence and the degradation characteristics of the severe fault stage in the degradation pattern library exceeds a preset threshold, a fault confirmation signal containing the fault facility identifier and degradation type code is output. The batch risk assessment module is used to extract the set of facility nodes in the same batch along the batch association edge in the facility heterogeneity map based on the facility identifier in the fault confirmation signal, obtain the remaining service life prediction distribution and degradation stage identifier of each facility in the same batch, and generate a batch group fault risk assessment report containing the individual risk level and spatial distribution location of each facility in the same batch. The emergency event generation module is used to input the batch-based fault risk assessment report into the degradation-triggered emergency event generator and encode it into a structured emergency event descriptor that is compatible with the standard event coding specification of the event type ontology library. An emergency strategy generation module is used to input the structured emergency event descriptor into the event-aware meta-policy network and output an initial emergency inspection strategy through an event type condition generation mechanism. The initial emergency inspection strategy includes a set of facilities to be inspected and the inspection task type of each facility. The path planning module is used to input the set of facilities to be inspected and the spatial distribution of each facility into the deep reinforcement learning path planning model, and generate an emergency inspection path covering all high-risk facilities by combining traffic constraints and facility accessibility conditions.

[0066] like Figures 2-9 As shown, the steps of the AI ​​intelligent inspection method based on deep reinforcement learning path planning are applied to the following application scenarios: A large residential community was completed in 20XX, comprising 15 residential buildings with 120 elevators, 80 electrical distribution boxes, and 45 water pumps. The community management deployed an AI-powered intelligent inspection system, and all facilities were equipped with operational parameter acquisition terminals to collect real-time data on equipment vibration, temperature, current, and other operating parameters. At 9:32 AM on June 15, 20XX, elevator No. 6 in Building A experienced a sudden brake failure, resulting in people being trapped. The system received a fault alarm signal. This elevator was part of the 24 elevators procured and installed in the third batch of 20XX, distributed across buildings A, B, and C, with a cumulative operating time of 52,000 hours. The inspection system needed to immediately activate the emergency inspection procedure for batch-wide fault risks, conducting an urgent inspection of all elevators in the same batch.

[0067] Step 1 is implemented, including the identification of degradation signals and confirmation of faults in the facilities: The system acquires real-time data streams of operating parameters for elevator No. 6 in Building A for the 72 hours prior to the malfunction, extracting three key parameters: vibration amplitude, brake current, and operating noise spectrum. The extracted parameter sequences are then dynamically time-warped and matched against typical degradation signal sequences corresponding to elevator brake malfunctions in the degradation pattern library. The elevator brake malfunction degradation signal sequence templates stored in the degradation pattern library contain typical signal characteristics of four stages: normal operation, early degradation, accelerated degradation, and severe malfunction.

[0068] Table 1 Time-series data fragments of operating parameters of the faulty facility The system uses a dynamic time warping algorithm to calculate the matching degree between the operating parameter sequence of elevator No. 6 in Building A and the signal sequence template of the severe fault stage in the deterioration mode library. The minimum cumulative distance output by the algorithm is 18.7, and the matching degree score after normalization is 0.91. Since the matching degree threshold for elevator facilities is set at 0.85, the current matching degree of 0.91 exceeds the preset threshold, and the system triggers the output of a fault confirmation signal.

[0069] Table 2 Fault Confirmation Signal Contents The degradation type code DT-ELE-BRAKE-F in the fault confirmation signal indicates the elevator brake fault type. This code serves as a key input parameter for risk assessment of related facilities in subsequent batches.

[0070] Step 2 implementation: Identification of facilities in the same batch and risk assessment of batch-wide failures: Based on the facility identifier ELE-A-006 in the fault confirmation signal, the system locates the corresponding facility node in the facility heterogeneity graph. The facility heterogeneity graph contains 120 elevator facility nodes, 80 electrical distribution box facility nodes, 45 water pump facility nodes, 15 batch nodes, and 15 area nodes. The system traverses along the batch association edges to extract all facility nodes that are connected to the same batch node BATCH-20XX-03 as ELE-A-006.

[0071] Table 3 Basic Information of Facilities in the Same Batch The system obtains the predicted distribution of remaining useful life for each facility node in the same batch. The predicted distribution is output by a pre-trained facility useful life prediction model, which generates the mean and variance of the remaining useful life based on the facility's historical operating data, current degradation stage identifier, and cumulative operating time.

[0072] Table 4. Predicted distribution of remaining useful life of facilities in the same batch The system calculates individual risk scores based on the predicted average remaining useful life of each facility and the stage weight corresponding to the current deterioration stage. Taking facility ELE-A-012 as an example, its predicted average remaining useful life is 580 hours, and its current deterioration stage is identified as accelerated deterioration, with a corresponding stage weight of 0.7. The individual risk score is calculated as follows: ; Similarly, calculate the individual risk score for all facilities in the same batch, and determine the risk level by comparing the risk score with the preset threshold.

[0073] Table 5 Batch Mass Sending Failure Risk Assessment Report Note: Due to the extremely low average remaining useful life (120 hours) and severe failure stage of the faulty facility ELE-A-006 (weight 1.0), its individual risk score reaches 0.833, far exceeding the high-risk threshold of 0.5. Although other facilities are in the accelerated deterioration or early deterioration stage, their remaining useful life is relatively long, and their risk scores are all below 0.2, which are initially judged as low risk.

[0074] The batch failure risk assessment report contains detailed risk information for 8 facilities from the same batch, including 1 high-risk facility and 7 low-risk facilities. This report serves as input data for a degradation-triggered emergency event generator.

[0075] Step 3 is implemented, and structured emergency event descriptors are generated: The system inputs the batch-based fault risk assessment report into the degradation-triggered emergency event generator. The generator first extracts the degradation type code DT-ELE-BRAKE-F of the faulty facility, the spatial distribution set of facilities in the same batch (east side of Building A, west side of Building A, east side of Building B, middle of Building B, west side of Building B, east side of Building C, middle of Building C, west side of Building C), and the individual risk level of each facility.

[0076] The generator maps fields according to the standard event coding specifications of the event type ontology. The event type field is filled with "batch-wide failure risk", the event source field is filled with "ELE-A-006|DT-ELE-BRAKE-F", the event impact range field is filled with the set of spatial distribution locations of facilities in the same batch, and the event severity field is calculated based on the proportion of high-risk facilities.

[0077] The proportion of high-risk facilities to the total number of facilities in the same batch is: ; because The severity of the event is assigned a "normal" level.

[0078] The generator encodes the individual risk level of each facility in the same batch according to the associated object coding format in the standard event coding specification, and generates a facility risk list.

[0079] Table 6 Structured Emergency Event Descriptors The structured emergency event descriptor format is fully compatible with the standard event coding specifications of the event type ontology library, enabling the event-aware meta-policy network to process it as a legitimate event input and trigger the emergency inspection path planning process.

[0080] Step 4 is implemented, and the initial emergency inspection strategy is generated: The system inputs structured emergency event descriptors into the event-aware meta-policy network. The network's event encoder receives the structured emergency event descriptors and encodes the event type fields "batch-wide failure risk," "event impact scope field," and "event severity field," as well as the general level, into a 128-dimensional event feature vector.

[0081] The condition generation module uses the event feature vector as input and combines it with the individual risk level of each facility in the associated object encoding to generate an intermediate representation vector for inspection task allocation. The strategy output layer decodes the intermediate representation vector into a set of facilities to be inspected and the corresponding inspection task type for each facility.

[0082] Since only facility ELE-A-006 has a high risk level, the strategy output layer assigns it a comprehensive and in-depth inspection task. For the other 7 low-risk facilities, considering batch correlation and the common origin of degradation, the strategy output layer assigns key parameter review tasks to the 5 facilities (ELE-A-012, ELE-B-003, ELE-B-015, ELE-C-005, and ELE-C-010) that are in the accelerated degradation stage, and does not assign inspection tasks to the 2 facilities (ELE-B-008 and ELE-C-018) that are in the early degradation stage.

[0083] Table 7 Initial Strategy for Emergency Inspection The initial strategy for emergency inspection identifies a set of facilities to be inspected, including 6 elevators, with a total estimated inspection time of 145 minutes. This strategy serves as input to a deep reinforcement learning path planning model.

[0084] Step 5 is implemented, and the emergency inspection route is generated: Before inputting the deep reinforcement learning path planning model, the system ranks the facilities to be inspected based on the predicted variance of their remaining useful life. The predicted variance of remaining useful life reflects the uncertainty of the prediction; the larger the variance, the higher the uncertainty of the actual deterioration state of the facility, and the more priority it needs to be inspected.

[0085] Table 8. Inspection task sequence driven by uncertainty The system inputs an uncertainty-driven inspection task sequence and the spatial distribution of facilities to be inspected into a deep reinforcement learning path planning model. The model employs a modified reward function, adding an uncertainty reward term to the original distance penalty, task completion reward, and constraint violation penalty. The model's state space is defined as a combination of the current inspection position, the set of visited facilities, and the set of unvisited facilities. The action space is defined as selecting the next facility to be visited from the current position.

[0086] Assume the initial location of the inspection personnel is at the community management center, with coordinates (0,0). The spatial coordinates of each facility are as follows: ELE-A-006 is located at (50,20), ELE-A-012 is located at (50,80), ELE-B-003 is located at (150,30), ELE-B-015 is located at (150,90), ELE-C-005 is located at (250,25), and ELE-C-010 is located at (250,60).

[0087] The model performs path planning and generates a facility visit sequence. Taking the first step of decision-making as an example, starting from the initial position (0,0), the candidate facilities include 6 elevators to be inspected.

[0088] Calculate the instant rewards for each candidate facility: For facility ELE-C-010: Distance traveled: rice; Task completion reward: (Review of key parameters for facilities at medium risk). Number of constraint violations: ; Uncertainty-based rewards: ; Instant rewards: ; For facility ELE-A-006: Distance traveled: rice; Task completion reward: (High-risk facilities require comprehensive and in-depth testing.) Uncertainty-based rewards: ; Instant rewards: ; Because the uncertainty reward term for ELE-C-010 is significantly higher than that for other facilities, the model selects ELE-C-010 as the first facility to be visited. The model then proceeds to perform sequential decision-making, ultimately generating a complete emergency inspection route.

[0089] Table 9 Emergency Inspection Route Execution Plan The total distance traveled on the emergency inspection route is 642 meters, and the estimated total time is approximately 2 hours and 20 minutes (including 145 minutes of inspection operation time and approximately 15 minutes of travel time). The route prioritizes covering the facility with the highest predicted uncertainty, ELE-C-010, and then visits other facilities in descending order of uncertainty, finally visiting the confirmed faulty facility, ELE-A-006, for a comprehensive and in-depth inspection.

[0090] During the inspection process, inspectors visit each facility sequentially according to the planned route, collecting deep status data and uploading it to the system in real time. After completing the inspection of the first three facilities (ELE-C-010, ELE-C-005, and ELE-B-015), the system updates the predicted remaining service life distribution of these facilities based on the measured data and uses batch association edges to backpropagate to uninspected facilities for shared inference. If the measured results show that the risk score of some uninspected facilities is significantly lower than the batch warning threshold, the system removes them from the remaining inspection task sequence and triggers the deep reinforcement learning path planning model to replan the remaining path, thereby dynamically reducing the inspection scope and optimizing the path.

[0091] The data flow process for the entire emergency inspection procedure is as follows: The system first identifies the degradation signal characteristics of the severe fault stage from the operation data stream of the faulty facility ELE-A-006, and outputs a fault confirmation signal through dynamic time warping matching. Based on the facility identifier in the fault confirmation signal, the system extracts 8 elevator facilities in the same batch along the batch association edge in the facility heterogeneity map, and calculates the individual risk score and generates a batch group fault risk assessment report by combining the predicted distribution of the remaining service life of each facility and the current deterioration stage identifier. The degradation-triggered emergency event generator transforms batch-based failure risk assessment reports into structured emergency event descriptors that conform to the standard event coding specifications of the event type ontology library, enabling emergency inspection systems that were originally only designed for external emergencies to identify the risk of batch-based failures caused by the degradation of internal facilities. The event-aware meta-policy network receives structured emergency event descriptors and instantly generates an initial emergency inspection policy containing six facilities to be inspected. It assigns comprehensive and in-depth inspection tasks to high-risk facilities and key parameter verification tasks to low-risk facilities in the accelerated deterioration stage. The system constructs an uncertainty-driven inspection task sequence based on the predicted variance of the remaining useful life of each facility, and uses it as a priority constraint input to the deep reinforcement learning path planning model. The deep reinforcement learning path planning model uses a modified reward function, which comprehensively considers the distance traveled, the reward for completing the task, the penalty for violating constraints, and the reward for uncertainty, to generate an emergency inspection path with a total travel distance of 642 meters, prioritizing the coverage of facilities with the highest prediction uncertainty. During the inspection process, the system dynamically updates the facility risk assessment based on measured data, uses batch-related edges for backpropagation to perform shared inference, dynamically reduces the inspection scope, and triggers the replanning of the remaining paths.

[0092] The entire data flow process achieves a seamless switch from routine degradation monitoring to emergency inspection response, overcoming the dilemma that batch-wide failure risks cannot be responded to in a timely manner under the traditional inspection mode.

[0093] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An AI-powered intelligent inspection method based on deep reinforcement learning path planning, characterized in that, Includes the following steps: Perform time-series pattern matching on the community facility operation data stream. When the matching degree between the facility operation parameter sequence and the degradation characteristics of the severe fault stage in the degradation pattern library exceeds a preset threshold, output a fault confirmation signal containing the fault facility identifier and degradation type code. Based on the facility identifier in the fault confirmation signal, extract the set of facility nodes in the same batch along the batch association edge in the facility heterogeneity graph, obtain the predicted distribution of remaining service life and degradation stage identifier of each facility in the same batch, and generate a batch group fault risk assessment report containing the individual risk level and spatial distribution location of each facility in the same batch. The batch-based fault risk assessment report is input into the degradation-triggered emergency event generator and encoded into a structured emergency event descriptor that is compatible with the standard event coding specification of the event type ontology library. The structured emergency event descriptor is input into the event-aware meta-policy network, and the initial emergency inspection policy is output through the event type condition generation mechanism. The initial emergency inspection policy includes the set of facilities to be inspected and the inspection task type of each facility. The set of facilities to be inspected and the spatial distribution of each facility are input into a deep reinforcement learning path planning model. Combined with traffic constraints and facility accessibility conditions, an emergency inspection path covering all high-risk facilities is generated.

2. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, The timing pattern matching employs a dynamic time warping algorithm. The input to this algorithm is a sliding window segment of the facility's current operating parameters and typical signal sequences corresponding to each degradation type stored in the degradation pattern library. The output is a matching score. The algorithm calculates the optimal time alignment path between the facility's current operating parameter sequence and the typical signal sequence to obtain the minimum cumulative distance between the two sequences. This minimum cumulative distance is then normalized and used as the matching score. Each record in the degradation pattern library contains a degradation type code, a degradation stage identifier, and a typical signal sequence template corresponding to that degradation stage. The degradation stage identifier is categorized from low to high degradation severity into normal operation stage, early degradation stage, accelerated degradation stage, and severe failure stage. The preset threshold is pre-set according to different facility types and degradation types and stored in the degradation pattern library, associated with the corresponding degradation type code.

3. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, The individual risk level is determined as follows: For each facility in the same batch, the individual risk score is calculated by multiplying the stage weight corresponding to its current deterioration stage identifier with the product of the average of its remaining service life and the reciprocal of the sum of the smallest positive numbers to prevent the denominator from being zero; wherein, the severe failure stage, accelerated deterioration stage, early deterioration stage and normal operation stage correspond to decreasing stage weight values ​​respectively; based on the comparison of the individual risk score with the preset risk level threshold, each facility is divided into three levels: high risk, medium risk and low risk.

4. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, The processing steps of the degradation-triggered emergency event generator include: extracting the degradation type code of the faulty facility, the spatial distribution location of the facility set in the same batch, and the individual risk level of each facility from the batch-based fault risk assessment report, as the original event elements to be encoded; mapping the original event elements to the field positions of the standard event coding specification according to the field definitions in the event type ontology library, wherein the event type field is filled with the batch-based fault risk type identifier, the event source field is filled with the faulty facility identifier and its degradation type code, the event impact range field is filled with the spatial distribution location set of the facilities in the same batch, and the event severity field is assigned a value based on the proportion of high-risk facilities to the total number of facilities in the same batch; encoding the individual risk level of each facility in the same batch according to the attached object coding format in the standard event coding specification, generating a facility risk list associated with the structured emergency event descriptor, and outputting a complete structured emergency event descriptor.

5. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, The event-aware meta-policy network includes an event encoder, a condition generation module, and a policy output layer. The event encoder encodes the event type field, event impact range field, and event severity field in the structured emergency event descriptor into an event feature vector. The condition generation module uses the event feature vector as a gating signal and modulates the activation intensity of the facility risk list feature vector through a gating mechanism to generate an intermediate representation vector that integrates event context and facility risk information. The policy output layer decodes the intermediate representation vector into a binary vector of the set of facilities to be inspected and a probability distribution vector of the inspection task type for each facility. The inspection task types include two categories: comprehensive in-depth inspection and key parameter review. Comprehensive in-depth inspection tasks are assigned to facilities with high individual risk levels, and key parameter review tasks are assigned to facilities with medium individual risk levels. The event type condition generation mechanism refers to the event-aware meta-policy network using the type embedding vector corresponding to the event type field as a condition variable to adjust the activation weight of each policy parameter in the policy output layer, thereby generating differentiated inspection policies for different event types.

6. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, The deep reinforcement learning path planning model describes inspection path planning as a sequential decision-making process. This model employs a deep Q-network structure, including a state encoding layer, a policy network, and a value network. The state encoding layer encodes the current inspection position as a two-dimensional coordinate vector, and encodes the sets of visited facilities and unvisited facilities as binary vectors respectively. These are concatenated and mapped to a state vector through a fully connected layer. The policy network takes the state vector as input and outputs the probability distribution of each candidate action. The value network takes the state vector as input and outputs a state value estimate. The reward function of the deep reinforcement learning path planning model... The number is the sum of the immediate rewards for each step. The immediate reward for each step is a linear combination of the negative value of the distance penalty, the positive value of the task completion reward, and the negative value of the constraint violation penalty. The distance penalty is the distance traveled between the current facility and the next facility to be visited multiplied by the distance penalty coefficient. The task completion reward is the reward value for completing the inspection task of the next facility to be visited multiplied by the task completion reward coefficient. The constraint violation penalty is the number of times the passage constraint is violated in this step multiplied by the constraint violation penalty coefficient. The task completion reward value is determined according to the individual risk level of the facility, and the reward value corresponding to high-risk facilities is greater than the reward value corresponding to medium-risk facilities.

7. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 6, characterized in that, After outputting the initial emergency inspection strategy and before inputting the set of facilities to be inspected into the deep reinforcement learning path planning model, the method further includes: sorting the facilities in the set of facilities to be inspected based on the variance of the predicted distribution of the remaining useful life of each batch of facilities, arranging the facilities with the highest prediction uncertainty in the priority inspection position, and generating an uncertainty-driven inspection task sequence; using the uncertainty-driven inspection task sequence as the facility priority constraint input of the deep reinforcement learning path planning model, and implementing the facility priority constraint by adding an uncertainty reward term to the reward function, wherein the uncertainty reward term is the predicted variance of the remaining useful life of the facilities to be visited multiplied by the uncertainty reward coefficient.

8. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, After generating the emergency inspection path, the method further includes: during the execution of the emergency inspection path, collecting deep status data for each inspected facility, updating the predicted distribution of the facility's remaining service life based on the measured data; when the measured risk score of an inspected facility is lower than the batch warning threshold, marking the facility as low-risk and removing it from the remaining inspection task sequence, re-inputting the updated remaining inspection task sequence and the spatial distribution of the remaining uninspected facilities into the deep reinforcement learning path planning model to generate an updated inspection path for the remaining uninspected facilities; the batch warning threshold is a preset proportion of the average of the individual risk scores of all facilities in the same batch; The status update data of the inspected facilities is propagated back along the batch association edges in the facility heterogeneity graph to the nodes of the uninspected facilities. For the deviation between the measured and predicted degradation levels of the inspected facilities, the deviation is weighted according to the similarity of degradation stages of each facility in the batch, and the average remaining service life prediction of the uninspected facilities is corrected. The degradation stage similarity is determined based on the difference between the degradation stage identifiers of the inspected and uninspected facilities. Based on the corrected remaining service life prediction distribution, the individual risk level of the uninspected facilities is recalculated. If the individual risk level of the uninspected facility drops to low risk, it is removed from the remaining inspection task sequence and the replanning of the remaining path is triggered.

9. The AI ​​intelligent inspection method based on deep reinforcement learning path planning according to claim 1, characterized in that, After all emergency inspections are completed, the process also includes: writing the triggering degradation mode of this batch of failures into a degradation mode library. The triggering degradation mode includes the facility type that caused this batch of failures, the degradation type code, and the corresponding operating parameter signal sequence template. The operating parameter signal sequence template is obtained by extracting and summarizing the time series of operating parameters of the faulty facility before the failure occurred. The batch propagation path and actual detection results of this emergency inspection are written into an event type ontology library. The batch propagation path includes the access order, access time, and measured risk score of each facility. The actual detection results include the measured degradation stage identifier, the average measured remaining service life, and the identifier of whether a failure has occurred for each facility in the same batch. The association record between the batch of failure risk type and the corresponding inspection strategy is established using the batch failure risk type identifier as the index key.

10. An AI intelligent inspection system based on deep reinforcement learning path planning, used to execute the steps of the AI ​​intelligent inspection method based on deep reinforcement learning path planning as described in any one of claims 1 to 9, characterized in that, include: The fault confirmation module is used to perform time-series pattern matching on the community facility operation data stream. When the matching degree between the facility operation parameter sequence and the degradation characteristics of the severe fault stage in the degradation pattern library exceeds a preset threshold, a fault confirmation signal containing the fault facility identifier and degradation type code is output. The batch risk assessment module is used to extract the set of facility nodes in the same batch along the batch association edge in the facility heterogeneity map based on the facility identifier in the fault confirmation signal, obtain the remaining service life prediction distribution and degradation stage identifier of each facility in the same batch, and generate a batch group fault risk assessment report containing the individual risk level and spatial distribution location of each facility in the same batch. The emergency event generation module is used to input the batch-based fault risk assessment report into the degradation-triggered emergency event generator and encode it into a structured emergency event descriptor that is compatible with the standard event coding specification of the event type ontology library. An emergency strategy generation module is used to input the structured emergency event descriptor into the event-aware meta-policy network and output an initial emergency inspection strategy through an event type condition generation mechanism. The initial emergency inspection strategy includes a set of facilities to be inspected and the inspection task type of each facility. The path planning module is used to input the set of facilities to be inspected and the spatial distribution of each facility into the deep reinforcement learning path planning model, and generate an emergency inspection path covering all high-risk facilities by combining traffic constraints and facility accessibility conditions.

Images