Edge ai-based building equipment failure prediction management method, system, device and medium

Abstract

This application relates to the field of building equipment failure prediction and management technology. It provides a method, system, device, and medium for building equipment failure prediction and management based on edge AI. The method includes: performing reinforcement learning inference on a multi-dimensional state vector to generate joint control actions that include sensor frequency selection, model complexity selection, and upload strategy selection; executing the joint control actions to control data acquisition, model calculation, and data transmission operations during equipment operation; monitoring the equipment operating status after executing the joint control actions, calculating rewards based on the actual early warning effect and survivability, and generating a long-term reward signal; and updating the parameters of the reinforcement learning decision model and long-term reward signal stored in the local model memory to generate an optimized reinforcement learning strategy. This aims to improve the real-time performance and reliability of early warnings, reduce equipment energy consumption, and balance safety early warning with equipment survivability.

Description

Technical Field

[0001] This invention relates to the field of building equipment failure prediction and management technology, and in particular to building equipment failure prediction and management methods, systems, equipment and media based on edge AI. Background Technology

[0002] With the deep application of artificial intelligence technology in the construction industry, edge intelligent devices are playing an increasingly important role in areas such as construction safety monitoring, equipment status prediction, and energy management. Construction equipment failure prediction, as a core component for ensuring construction safety and improving operation and maintenance efficiency, has become a key technological support for the construction of modern smart construction sites. An effective failure prediction system can provide early warnings of equipment anomalies, optimize maintenance strategies, and significantly reduce the risk of safety accidents and economic losses.

[0003] Traditional technologies employ centralized cloud-based predictive models for large-scale analysis of equipment data. However, these models suffer from drawbacks such as delayed early warning responses due to network latency, data loss due to network instability, and excessive power consumption. When faced with the need for equipment fault early warning and energy consumption balance, traditional technologies use single-objective optimization strategies that focus on only a single indicator. However, this strategy can lead to systemic imbalances, resulting in a surge in energy consumption for high accuracy and an increase in false negative rates for energy saving. Furthermore, traditional technologies often rely on manual experience-driven periodic inspections for equipment maintenance, neglecting the need to detect sudden equipment anomalies. In addition, the subjective nature of human judgment can lead to inconsistent equipment risk assessment standards. Summary of the Invention

[0004] Therefore, it is necessary to provide edge AI-based methods, systems, equipment, and media for predictive management of building equipment failures to address the aforementioned technical issues, so as to improve the real-time performance and reliability of early warnings, reduce equipment energy consumption, and balance the technical effects of safety early warning and equipment survivability.

[0005] In a first aspect, this application provides a method for predictive management of building equipment failures based on edge AI, the method comprising:

[0006] Metadata compression is performed on the construction plan summary and regional weather forecast summary sent from the cloud to generate a lightweight environmental feature set; energy income prediction is performed on the lightweight environmental feature set to generate an energy income probability distribution curve;

[0007] Task criticality prediction is performed on the lightweight environment feature set to generate a time series curve of failure risk level; feature fusion is performed on the current battery state of charge, energy income probability distribution curve and failure risk level time series curve to generate a multi-dimensional state vector.

[0008] Reinforcement learning inference is performed on the multidimensional state vector to generate joint control actions that include sensor frequency selection, model complexity selection, and upload strategy selection; the joint control actions are executed to control the data acquisition, model calculation, and data transmission operations during the operation of the equipment.

[0009] The system monitors the equipment's operating status after executing joint control actions, calculates rewards based on the actual early warning effect and survival status, and generates long-term reward signals. It also updates the parameters of the reinforcement learning decision model and long-term reward signals stored in the local model memory to generate optimized reinforcement learning strategies.

[0010] In one embodiment, the device's operating status is monitored after a joint control action is executed, and a reward calculation is performed on the actual early warning effect and survival status to generate a long-term reward signal, including:

[0011] Successful early warning identification is performed on the sequence of early warning events generated after the execution of joint control actions, and a set of successful early warning events is generated.

[0012] The successful early warning event set is value-quantified and mapped to generate event value coefficients;

[0013] The sequence of fault events that occurs during the execution of joint control actions is used to identify critical fault omissions and generate a set of critical omission events.

[0014] Calculate the penalty coefficient for the key missed event set to generate the missed event penalty coefficient;

[0015] Dynamic energy consumption assessment is performed on the energy consumption data during equipment operation to generate an energy consumption penalty coefficient;

[0016] Calculate the lifecycle gain for the continuous operating time of the equipment and generate the lifecycle gain coefficient;

[0017] The event value coefficient, underreporting penalty coefficient, energy consumption penalty coefficient, and survival gain coefficient are weighted and fused to generate a long-term reward signal.

[0018] In one embodiment, a multi-objective weighted fusion of the event value coefficient, the missed detection penalty coefficient, the energy consumption penalty coefficient, and the survival gain coefficient is performed to generate a long-term reward signal, including:

[0019] The event value coefficient and the underreporting penalty coefficient are converted into a risk balance value to generate a security assessment value.

[0020] The energy consumption penalty coefficient and the survival gain coefficient are converted into an energy efficiency balance to generate a survival assessment value.

[0021] The safety assessment value and the survival assessment value are dynamically weighted and aggregated to generate an initial reward value. The expression for the initial reward value is:

[0022]

[0023] in, This represents the initial reward value. Indicates dynamic balancing weights. Indicates the safety assessment value. Indicates the survival assessment value. This represents the scaling factor for the security assessment. Represents the hyperbolic tangent function. This represents the modified linear unit function;

[0024] The initial reward value is time-discounted to generate a long-term reward signal. The expression for the long-term reward signal is:

[0025]

[0026] in, This indicates a long-term reward signal. Indicates the discount factor. Indicates dynamic balancing weights. Indicates the event value coefficient. This represents the penalty coefficient for underreporting. Represents the smoothing constant. Indicates the weight of the survival dimension. Represents the survival gain coefficient. This represents the energy consumption penalty coefficient. This represents the energy consumption sensitivity coefficient.

[0027] In one embodiment, the successful early warning event set is value-quantified and mapped to generate event value coefficients, including:

[0028] Perform a fault loss assessment on the set of successful early warning events to generate the expected loss reduction value;

[0029] Time decay compensation is applied to the expected reduction in loss to generate a time-sensitive compensation value;

[0030] The timeliness compensation value is subjected to nonlinear normalization to generate the event value coefficient. The expression for the event value coefficient is as follows:

[0031]

[0032] in, Indicates the event value coefficient. This represents the Sigmoid normalization function. Indicates the normalized scaling factor. Indicates the number of successfully alerted events. Indicates the first Weighting factors for equipment failure types Indicates the first The baseline loss value for this type of failure. Indicates the time decay coefficient. Indicates the first The lead time of an early warning event. This represents the normalized offset.

[0033] In one embodiment, feature fusion is performed on the current battery state of charge, energy income probability distribution curve, and fault risk level time series curve to generate a multi-dimensional state vector, including:

[0034] The current battery state of charge is normalized to generate a standardized charge value. The expression for the standardized charge value is:

[0035]

[0036] in, Represents standardized electricity values. Indicates the current state of battery charge. This represents the minimum value of the battery's state of charge. This indicates the maximum value of the battery's state of charge.

[0037] The energy income probability distribution curve is processed to extract time-period features, generating a time-period energy feature vector.

[0038] Key points are extracted from the time-series curve of fault risk level to generate a risk key feature vector.

[0039] The standardized energy values, time-period energy feature vectors, and risk-critical feature vectors are concatenated to generate an initial fusion vector. The expression for the initial fusion vector is as follows:

[0040]

[0041] in, Represents the initial fusion vector. This indicates a feature concatenation operation. Represents standardized electricity values. Represents the energy feature vector over a time period. Represents the key feature vector of risk;

[0042] The initial fusion vector is reduced in dimension and compressed to generate a multidimensional state vector.

[0043] In one embodiment, the task criticality prediction is performed on the lightweight environment feature set to generate a time series curve of failure risk level, including:

[0044] Semantic spatiotemporal analysis is performed on the construction plan summary with lightweight environmental feature set to generate a spatiotemporal distribution map of operation risk;

[0045] Detect risk clustering areas in the spatiotemporal distribution map of operational risks and generate key risk area markers;

[0046] The key risk areas are marked with time-dimensional propagation modeling to generate a risk propagation time series. The expression for the risk propagation time series is as follows:

[0047]

[0048] in, This represents the time series of risk transmission. Indicates the number of key risk areas. This indicates the risk intensity of the k-th critical risk area. This represents the time center point of the k-th critical risk area. This represents the time propagation radius of the k-th critical risk area. This represents a time variable, and k represents the index of the key risk area.

[0049] Risk levels are quantified and mapped to the time series of risk propagation to generate time series curves of failure risk levels.

[0050] In one embodiment, energy income prediction is performed on a lightweight environment feature set to generate an energy income probability distribution curve, including:

[0051] Multi-scale feature extraction is performed on weather forecast metadata in the lightweight environmental feature set to generate weather state feature vectors;

[0052] Probabilistic diffusion modeling is performed on weather state feature vectors to generate an energy income probability distribution;

[0053] The energy income probability distribution is aligned with the time series to generate an energy income probability distribution curve.

[0054] Secondly, this application also provides a building equipment fault prediction and management system based on edge AI, the system comprising:

[0055] The environmental feature compression module is used to compress metadata from the construction plan summary and regional weather forecast summary sent from the cloud to generate a lightweight environmental feature set; and to perform energy income prediction on the lightweight environmental feature set to generate an energy income probability distribution curve.

[0056] The risk state fusion module is used to predict the mission criticality of the lightweight environment feature set and generate a time series curve of the failure risk level; it also performs feature fusion on the current battery state of charge, energy income probability distribution curve and failure risk level time series curve to generate a multi-dimensional state vector.

[0057] The decision execution module is used to perform reinforcement learning inference on the multi-dimensional state vector to generate joint control actions that include sensor frequency selection, model complexity selection, and upload strategy selection; it executes the joint control actions to control the data acquisition, model calculation, and data transmission operations during the operation of the equipment.

[0058] The feedback learning optimization module is used to monitor the equipment's operating status after executing joint control actions, calculate rewards based on the actual early warning effect and survival status, and generate long-term reward signals; it also updates the parameters of the reinforcement learning decision model and long-term reward signals stored in the local model memory to generate optimized reinforcement learning strategies.

[0059] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0060] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0061] This application provides a method, system, device, and medium for predicting and managing building equipment failures based on edge AI. The method includes: compressing metadata from cloud-based construction plan summaries and regional weather forecast summaries to generate a lightweight environmental feature set; predicting energy income and task criticality based on this feature set; combining the current battery state of charge to complete feature fusion to obtain a multi-dimensional state vector; and then generating and executing joint control actions that include sensor frequency, model complexity, and upload strategy selection through reinforcement learning inference. This reduces redundant consumption in data transmission and computation, lowers equipment energy consumption, and reduces network latency interference with early warning response due to edge-side localized processing, thereby improving the real-time performance of early warnings.

[0062] The generation of fault risk level time series curves and the dynamic adjustment of data acquisition and model calculation by joint control actions help to capture abnormal equipment states more promptly and improve the reliability of early warnings. Subsequently, by monitoring the equipment operating status, the actual early warning effect and survival status are incorporated into the calculation of long-term reward signals and used to update the parameters of the local reinforcement learning decision model, continuously optimizing the joint control strategy. This ensures the effectiveness of safety early warnings while taking into account the continuous operation requirements of the equipment, further balancing safety early warnings and equipment survivability. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 A flowchart of a building equipment fault prediction and management method based on edge AI in one embodiment of the present invention;

[0065] Figure 2 This is a flowchart illustrating the process of predicting energy income from a lightweight environment feature set and generating an energy income probability distribution curve in one embodiment of the present invention.

[0066] Figure 3 This is a structural diagram of a building equipment fault prediction and management system based on edge AI, according to one embodiment of the present invention. Detailed Implementation

[0067] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0068] First, the application scenarios of the embodiments of this application are described. In the embodiments of this application, a method, system, equipment and medium for predicting and managing the faults of large construction equipment such as tower cranes, construction elevators and concrete pumps, applicable to but not limited to the daily operation status monitoring and fault early warning management scenarios of large construction equipment such as tower cranes, construction elevators and concrete pumps in smart construction sites, as well as the operation and maintenance optimization scenarios of various electromechanical equipment on construction sites, based on edge AI are provided.

[0069] In illustrative purposes, the method, system, equipment and medium for predicting and managing building equipment faults based on edge AI provided in this application embodiment can also be applied to application scenarios such as status control of heavy production machinery and equipment in industrial parks, operation and maintenance of special equipment for bridge and tunnel construction, and operation and maintenance management of supporting equipment for municipal engineering construction. This is only an example and does not limit the specific application scenarios.

[0070] like Figure 1 As shown, this application provides a building equipment fault prediction and management method based on edge AI, the method comprising:

[0071] S101: Compress metadata from the cloud-based construction plan summary and regional weather forecast summary to generate a lightweight environmental feature set; predict energy income from the lightweight environmental feature set to generate an energy income probability distribution curve.

[0072] For example, the system receives a summary of the construction plan and a summary of the regional weather forecast from the cloud, performs metadata filtering, retains core metadata related to the operation and fault prediction of building equipment, removes irrelevant redundant metadata, then performs structured integration of the filtered core metadata, completes metadata compression, and generates a lightweight environmental feature set.

[0073] Using a lightweight environmental feature set as the basic data source, key environmental features related to energy income are extracted, and trend analysis and probabilistic modeling of energy income are carried out to explore the potential correlation between environmental features and energy income. Through the construction and fitting of probability distributions, an energy income probability distribution curve is generated.

[0074] S102: Perform task criticality prediction on the lightweight environment feature set to generate a time series curve of fault risk level; perform feature fusion on the current battery state of charge, energy income probability distribution curve and fault risk level time series curve to generate a multi-dimensional state vector.

[0075] For example, using a lightweight environmental feature set as the basis for analysis, key environmental features directly related to the operation of building equipment are extracted. Combining the spatiotemporal attributes of task execution with the potential impact of environmental factors on equipment operation, task criticality assessment is conducted to identify risk-related factors that may induce equipment failure. By quantifying the intensity of risk factors and dynamically tracking them over time, a mapping relationship between risk level and time is established to generate a time series curve of failure risk level.

[0076] The current battery state of charge is obtained and normalized. At the same time, time-period energy correlation features are extracted from the energy income probability distribution curve and key risk change features are extracted from the fault risk level time series curve. The normalized current battery state of charge, the extracted time-period energy correlation features and key risk change features are structurally integrated. Through the orderly splicing and fusion of features, the differences and interference of different feature dimensions are eliminated to generate a multi-dimensional state vector.

[0077] S103: Perform reinforcement learning inference on the multi-dimensional state vector to generate joint control actions that include sensor frequency selection, model complexity selection, and upload strategy selection; execute the joint control actions to control the data acquisition, model calculation, and data transmission operations during equipment operation.

[0078] For example, a multi-dimensional state vector is input into a pre-built reinforcement learning decision model, and inference operations are performed by combining historical decision-making experience and real-time state characteristics. Based on the resource consumption and fault warning requirements of equipment operation, the adaptation scheme for sensor frequency selection, the matching standard for model complexity selection, and the execution rules for upload strategy selection are determined respectively. The three types of decision schemes are integrated in a coordinated manner to generate a joint control action that includes sensor frequency selection, model complexity selection, and upload strategy selection.

[0079] Based on the generated joint control actions, and in accordance with the relevant requirements for sensor frequency selection, the interval and coverage of equipment data acquisition are adjusted. Referring to the established standards for model complexity selection, an edge-side fault prediction model of corresponding complexity is matched for computation. Following the specific rules for upload strategy selection, the content scope and triggering timing of data transmission are defined, thereby achieving control over data acquisition, model calculation, and data transmission operations during equipment operation.

[0080] S104: Monitor the equipment's operating status after executing joint control actions, calculate rewards based on the actual early warning effect and survival status, and generate long-term reward signals; update the parameters of the reinforcement learning decision model and long-term reward signals stored in the local model memory to generate optimized reinforcement learning strategies.

[0081] For example, the system continuously tracks the operating status of the equipment after executing joint control actions, collects equipment fault warning trigger records, actual fault occurrences, and survival status-related data such as equipment power consumption and continuous operating time, conducts value assessment of the actual warning effect and gain analysis of the survival status, and combines the two types of analysis results with preset evaluation dimensions to complete the entire process of reward calculation and generate a long-term reward signal.

[0082] The reinforcement learning decision model stored in the local model memory is retrieved. The long-term reward signal is used as the core basis for updating the model parameters. The decision weights and correlation coefficients inside the model are adjusted according to the parameter iteration logic of reinforcement learning. Through multiple rounds of parameter calibration and logic optimization, the parameter update process of the reinforcement learning decision model is completed, and the optimized reinforcement learning policy is generated.

[0083] An embodiment of this application also provides a building equipment fault prediction and management method based on edge AI, which includes: compressing metadata of construction plan summaries and regional weather forecast summaries sent from the cloud to generate a lightweight environmental feature set; performing energy income prediction and task criticality prediction based on the feature set; combining the current battery state of charge to complete feature fusion to obtain a multi-dimensional state vector; and then generating and executing a joint control action that includes sensing frequency, model complexity, and upload strategy selection through reinforcement learning inference. This reduces redundant consumption in the data transmission and calculation process, lowers equipment energy consumption, and reduces network latency interference on early warning response due to edge-side localized processing, thereby improving the real-time performance of early warning.

[0084] The generation of fault risk level time series curves and the dynamic adjustment of data acquisition and model calculation by joint control actions help to capture abnormal equipment states more promptly and improve the reliability of early warnings. Subsequently, by monitoring the equipment operating status, the actual early warning effect and survival status are incorporated into the calculation of long-term reward signals and used to update the parameters of the local reinforcement learning decision model, continuously optimizing the joint control strategy. This ensures the effectiveness of safety early warnings while taking into account the continuous operation requirements of the equipment, further balancing safety early warnings and equipment survivability.

[0085] In one embodiment, the device's operating status is monitored after a joint control action is executed, and a reward calculation is performed on the actual early warning effect and survival status to generate a long-term reward signal, including:

[0086] (1) Successful early warning identification is performed on the sequence of early warning events generated after the joint control action is executed, and a set of successful early warning events is generated.

[0087] For example, all early warning event sequences generated by the equipment during operation after the execution of joint control actions are collected. The time range and recording dimensions of the early warning event sequences are clearly defined to ensure coverage of early warning output information under various operating scenarios of the equipment. The equipment operating period, associated fault type, early warning triggering conditions, and early warning level corresponding to each early warning event are analyzed to establish a detailed ledger of early warning events, providing complete data support for subsequent identification operations. Simultaneously, the actual fault situations that occur within the same operating period are collected, including the fault occurrence time, fault type, scope of impact, and fault handling results, forming an actual fault occurrence record.

[0088] Each early warning event in the detailed early warning event ledger is compared one by one with the actual fault occurrence record to clarify the comparison dimensions between the early warning event and the actual fault, including time correlation, fault type consistency, and impact range matching degree. For each early warning event, it is verified whether its early warning trigger time is earlier than the corresponding actual fault occurrence time, whether the fault type associated with the early warning is completely consistent with the actual fault type, and whether the impact range indicated by the early warning matches the impact range caused by the actual fault. Early warning events that meet the requirements of time correlation, type consistency, and scope matching degree are selected, categorized and integrated to complete the successful early warning identification operation and generate a successful early warning event set.

[0089] Among them, the early warning event sequence includes all early warning-related records triggered by the equipment during data acquisition, model calculation, status monitoring and other stages during the execution of joint control actions. It is a complete data set output by the equipment's early warning function.

[0090] (2) Quantify and map the successful early warning event set to generate event value coefficients.

[0091] For example, for the generated set of successful early warning events, the actual role of each successful early warning event in the equipment operation and maintenance process is analyzed one by one, clarifying the specific contribution of each successful early warning event to avoiding serious equipment damage, reducing construction interruption time, and reducing the risk of safety accidents. Combining industry impact standards for various faults in the construction engineering field, the value weight levels of different types of faults are divided according to factors such as the severity of the fault, the scope of impact, and the difficulty of handling, thus establishing a fault value assessment system.

[0092] Based on the aforementioned evaluation system, the fault value weight corresponding to each successful early warning event is determined. Combining the degree to which the early warning duration mitigates potential losses, a value quantification mapping rule is established. Each successful early warning event is transformed into a quantifiable value indicator according to the established rules. The quantified value indicators of all successful early warning events are then aggregated to complete the value quantification mapping operation and generate an event value coefficient.

[0093] Among them, the event value coefficient is a core indicator that comprehensively quantifies the operation and maintenance value, safety value and economic value generated by a successful early warning event, and intuitively reflects the degree of positive impact of a successful early warning on equipment operation and construction progress.

[0094] (3) Identify key failures in the sequence of failure events that occur during the execution of joint control actions and generate a set of key failure events.

[0095] For example, during the execution of joint control actions, a sequence of fault events occurring in the equipment under all operating scenarios is collected. The statistical time period and recording elements of the fault event sequence are clearly defined to ensure that all types of faults, such as minor faults, general faults, and serious faults, are included. The fault level, scope of impact, time of occurrence, cause of fault, and resulting losses of each fault event are analyzed to establish a detailed fault event file, fully presenting the overall picture of equipment fault occurrence during that time period. At the same time, all early warning records of the equipment during the corresponding operating period are retrieved, including triggered early warning events and blank records of non-triggered early warnings, to form an early warning output comparison file.

[0096] By comparing the warning output with the reference file, each fault event in the detailed fault event file is checked one by one to determine whether a corresponding warning record exists. For fault events for which no corresponding warning record is found, the fault level is further assessed to determine whether it meets the preset critical fault criteria, and the scope of its impact is analyzed to determine whether it involves construction safety, core equipment operation, or significant economic losses. Fault events that were not captured by the warning system but meet the critical fault judgment criteria are screened out, classified and organized, and the critical fault missed reporting identification operation is completed to generate a critical missed reporting event set.

[0097] Among them, the fault event sequence includes all fault-related records that actually occur in the equipment during the execution of joint control actions due to various reasons such as mechanical wear, circuit failure, and software anomaly, which is a complete presentation of the abnormal operating state of the equipment.

[0098] (4) Calculate the penalty coefficient for the set of key missed events and generate the missed event penalty coefficient.

[0099] For example, based on the generated set of critical missed events, the potential chain reactions triggered by each critical missed event are analyzed one by one, and their potential risks to the lifespan of core equipment components, construction progress, and on-site personnel safety are assessed. In accordance with the safety management specifications and loss assessment standards for building equipment operation and maintenance, quantitative rules for critical missed event penalties are established, clarifying the basic penalty scores corresponding to critical missed events of different severity levels. Simultaneously, penalty coefficient adjustment items are set based on factors such as the duration and scope of the impact of critical missed events.

[0100] For each critical missed event, a base penalty score is determined based on its severity, and then the final penalty score is obtained by adjusting the score. The final penalty scores of all critical missed events are summed to complete the penalty coefficient calculation operation and generate the missed event penalty coefficient.

[0101] The critical missed event set is a set of failure events that were not identified by the early warning system during the execution of joint control actions and that have a significant or potentially significant negative impact on equipment operation safety, construction progress, and personnel safety.

[0102] (5) Perform dynamic energy consumption assessment on the energy consumption data during equipment operation and generate energy consumption penalty coefficient.

[0103] For example, energy consumption data generated during the execution of joint control actions by the equipment is collected in real time at each operational stage. The data collection nodes are clearly defined, covering the data acquisition unit, model calculation unit, data transmission unit, and energy consumption records in the equipment's standby state. The time-period distribution characteristics of the energy consumption data are analyzed, and the changing patterns of energy consumption under different operating states are investigated to establish an energy consumption time-period distribution ledger. Based on the equipment's design energy consumption standards, industry energy consumption levels of similar equipment, and energy-saving operation requirements of edge devices, a dynamic energy consumption assessment benchmark is established to clarify the reasonable energy consumption range under different operating scenarios.

[0104] The collected energy consumption data is compared with the dynamic energy consumption assessment benchmark to determine whether the energy consumption in each time period and each stage exceeds the reasonable range, and to analyze the specific reasons and influencing factors for energy consumption exceeding the standard. Based on the preset energy consumption penalty quantification standard, the corresponding energy consumption penalty score is determined according to the magnitude of energy consumption exceeding the standard, the duration, and the degree of impact on the equipment's endurance, thus completing the dynamic energy consumption assessment operation and generating the energy consumption penalty coefficient.

[0105] Among them, energy consumption data includes records of electrical energy consumption generated by the equipment in all operating states such as data acquisition, model calculation, data transmission, and standby during the execution of joint control actions, which is a comprehensive reflection of the equipment's energy usage.

[0106] (6) Calculate the life cycle gain of the equipment during continuous operation and generate the life cycle gain coefficient.

[0107] For example, the continuous operating time of the equipment from the start of executing joint control actions to the current moment is recorded, and the statistical starting point and rules for continuous operating time are clearly defined to ensure the accuracy and continuity of timing. The matching relationship between this continuous operating time and the equipment's design service life, rated continuous operating time, and routine maintenance cycle is analyzed to assess the impact of continuous operating status on equipment wear and tear. Simultaneously, the stability of the equipment's status during continuous operation is monitored, including indicators such as the range of fluctuations in operating parameters, the frequency of failures, and the degree of performance degradation, to establish a continuous operating status assessment log for the equipment.

[0108] By combining construction progress requirements with equipment operation and maintenance efficiency targets, a lifecycle gain assessment system is established to clarify the quantitative standards for the gains made by continuous equipment operation in ensuring construction continuity, improving operation and maintenance response efficiency, and reducing equipment start-up and shutdown losses. Based on this assessment system, the comprehensive benefits brought by continuous equipment operation are quantified by considering continuous operating time, state stability, and construction and operation and maintenance needs, and the lifecycle gain calculation operation is completed to generate a lifecycle gain coefficient.

[0109] Among them, the life cycle gain coefficient is an indicator that comprehensively quantifies the construction guarantee value, operation and maintenance efficiency value and equipment loss control value generated by the continuous and stable operation of equipment, reflecting the supporting role of the equipment's life status in the overall construction and operation and maintenance work.

[0110] (7) Perform multi-objective weighted fusion of event value coefficient, underreporting penalty coefficient, energy consumption penalty coefficient and survival gain coefficient to generate long-term reward signal.

[0111] For example, based on the core requirements of building equipment failure prediction management, the weighting principles for event value coefficient, missed reporting penalty coefficient, energy consumption penalty coefficient, and survival gain coefficient in the calculation of long-term reward signals are clarified. These principles must balance the priority of safety warnings with the energy-saving survival needs of equipment. The specific weighting of each coefficient is dynamically determined according to the dynamic changes in the construction scenario, the different characteristics of the equipment operation phase, and the adjustment direction of maintenance goals, ensuring that the weighting allocation is adapted to actual application needs.

[0112] Based on the set weighting ratios, the event value coefficient, missed detection penalty coefficient, energy consumption penalty coefficient, and survival gain coefficient are substituted into the multi-objective fusion calculation process. The event value coefficient serves as a positive incentive indicator, the missed detection penalty coefficient and energy consumption penalty coefficient serve as negative constraint indicators, and the survival gain coefficient serves as a positive support indicator. The four types of coefficients are integrated through weighted summation to balance the two core objectives of safety early warning effect and equipment survival status, eliminating the impact of differences in different indicator dimensions, completing the multi-objective weighted fusion operation, and generating a long-term reward signal.

[0113] Among them, multi-objective weighted fusion is a technical process that comprehensively considers two core objectives: optimizing the effect of safety early warning and ensuring the survival status of equipment. It systematically integrates and calculates the corresponding quantitative coefficients through dynamic weight allocation.

[0114] In one embodiment, a multi-objective weighted fusion of the event value coefficient, the missed detection penalty coefficient, the energy consumption penalty coefficient, and the survival gain coefficient is performed to generate a long-term reward signal, including:

[0115] (1) Perform risk balance conversion between the event value coefficient and the underreporting penalty coefficient to generate a safety assessment value.

[0116] For example, the event value coefficient and the missed detection penalty coefficient are obtained. The event value coefficient is the result of value quantification mapping of the successful early warning event set, and the missed detection penalty coefficient is the result of penalty coefficient calculation for the critical missed detection event set. The core logic of risk balance transformation is clarified: based on the actual effect of security early warning, the positive value corresponding to the event value coefficient and the negative impact corresponding to the missed detection penalty coefficient are offset and adapted. Simultaneously, the quantification scales of the two types of coefficients are uniformly calibrated to eliminate differences caused by different evaluation dimensions. Then, the adapted and calibrated results are merged to complete the risk balance transformation operation and generate a security assessment value.

[0117] Among them, the safety assessment value is a quantitative indicator that can uniformly measure the effectiveness of safety early warning, formed by combining the positive safety benefits of the comprehensive event value coefficient and the negative safety risks of the underreporting penalty coefficient after risk balancing transformation.

[0118] (2) Perform energy efficiency balance conversion on the energy consumption penalty coefficient and the survival gain coefficient to generate survival assessment value.

[0119] For example, the energy consumption penalty coefficient and the survival gain coefficient are obtained. The energy consumption penalty coefficient is the result of dynamic energy consumption assessment of energy consumption data during equipment operation, and the survival gain coefficient is the result of survival cycle gain calculation of continuous equipment operation time. The core logic of energy efficiency balance conversion is clarified, namely, based on the survival status benefit of the equipment, the negative energy consumption constraint corresponding to the energy consumption penalty coefficient and the positive operating benefit corresponding to the survival gain coefficient are adapted and integrated. At the same time, the quantification scale of the two types of coefficients is uniformly calibrated to eliminate the differences caused by different assessment dimensions. Then, the adapted and calibrated results are fused to complete the energy efficiency balance conversion operation and generate a survival assessment value.

[0120] Among them, the survival assessment value is a quantitative indicator that can uniformly measure the survival status of equipment, formed by combining the negative energy consumption impact of the comprehensive energy consumption penalty coefficient and the positive operating benefits of the survival gain coefficient after energy efficiency balance conversion.

[0121] (3) The safety assessment value and the survival assessment value are dynamically weighted and aggregated to generate an initial reward value. The expression for the initial reward value is:

[0122]

[0123] in, This represents the initial reward value. Indicates dynamic balancing weights. Indicates the safety assessment value. Indicates the survival assessment value. This represents the scaling factor for the security assessment. Represents the hyperbolic tangent function. This represents the modified linear unit function.

[0124] For example, safety assessment values ​​and survival assessment values ​​are obtained, and dynamic balancing weights are determined. The setting of these dynamic balancing weights needs to consider the safety priority requirements and equipment endurance guarantee requirements of the current construction scenario, clarifying the weight ratio under different requirements. The safety assessment values ​​are processed using a hyperbolic tangent function to compress the extreme fluctuation range of the safety assessment values, making the contribution of the safety assessment values ​​to the rewards more stable.

[0125] The survival assessment values ​​are processed by modifying the linear unit function, retaining the positive benefit contributions corresponding to the survival assessment values ​​and filtering out invalid negative interference terms. According to the determined dynamic balance weights, the safety assessment values ​​processed by the function are weighted and integrated with the survival assessment values ​​to complete the dynamic weighted aggregation process and generate the initial reward value.

[0126] The initial reward value is a quantitative reward indicator that initially reflects the comprehensive benefits of safety early warning and equipment survival status by combining dynamic balancing weights and weighted aggregation of safety assessment value and survival assessment value after function processing.

[0127] (4) Apply time discounting to the initial reward value to generate a long-term reward signal. The expression for the long-term reward signal is:

[0128]

[0129] in, This indicates a long-term reward signal. Indicates the discount factor. Indicates dynamic balancing weights. Indicates the event value coefficient. This represents the penalty coefficient for underreporting. Represents the smoothing constant. Indicates the weight of the survival dimension. Represents the survival gain coefficient. This represents the energy consumption penalty coefficient. This represents the energy consumption sensitivity coefficient.

[0130] For example, an initial reward value is obtained, and a discount factor is determined. The setting of this discount factor needs to be based on the timeliness characteristics of the reward signal, clarifying the difference in the weight of recent and long-term evaluation results on decision optimization. Simultaneously, the event value coefficient, underreporting penalty coefficient, survival gain coefficient, and energy consumption penalty coefficient are correlated. According to the time discount rule, the weight of the initial reward value in the time dimension is adjusted, weakening the proportion of long-term evaluation results in the reward signal and strengthening the actual impact of recent evaluation results on the reward signal. This completes the time discount processing operation and generates a long-term reward signal.

[0131] Among them, the long-term reward signal is a quantitative reward signal that, after being processed with time discount, can comprehensively consider the safety warning effect and equipment survival status under different time dimensions, and provide an optimization basis for the reinforcement learning decision model.

[0132] In one embodiment, the successful early warning event set is value-quantified and mapped to generate event value coefficients, including:

[0133] (1) Perform a fault loss assessment on the set of successful early warning events and generate the expected loss reduction value.

[0134] For example, a set of successful early warning events is obtained, and the equipment failure type corresponding to each successful early warning event in the set is extracted one by one. Based on the failure loss standards in the field of building equipment operation and maintenance, the baseline loss range corresponding to different equipment failure types is defined. The intervention node of each successful early warning event on the corresponding equipment failure is analyzed, and the equipment failure loss that the intervention node can avoid is evaluated. The avoided losses corresponding to each successful early warning event are summarized to complete the failure loss assessment operation and generate the expected loss reduction value.

[0135] Among them, the successful early warning event set is a collection of all valid early warning events obtained after the successful early warning identification operation; the expected loss reduction value is a quantitative result of the equipment failure loss avoided due to successful early warning, determined through failure loss assessment.

[0136] (2) Time decay compensation is applied to the expected loss reduction value to generate a time-sensitive compensation value.

[0137] For example, the expected loss reduction value is obtained, and the warning lead time corresponding to each successful warning event is extracted. The correlation logic between the warning lead time and the equipment failure loss avoidance effect is clarified. According to the preset time decay rule, the expected loss reduction value corresponding to different warning lead times is adjusted differently. For successful warning events with short warning lead times, the corresponding loss quantification value is appropriately reduced. For successful warning events with reasonable warning lead times, the corresponding loss quantification value is retained. The adjusted loss quantification values ​​of all successful warning events are integrated to complete the time decay compensation operation and generate the timeliness compensation value.

[0138] Among them, time decay compensation is a process of adjusting the expected loss reduction value based on the early warning time of successful early warning events; the timeliness compensation value is a quantitative indicator that reflects the combined effect of early warning timeliness and equipment failure loss avoidance value after time decay compensation.

[0139] (3) The timeliness compensation value is subjected to nonlinear normalization to generate the event value coefficient. The expression for the event value coefficient is as follows:

[0140]

[0141] in, Indicates the event value coefficient. This represents the Sigmoid normalization function. Indicates the normalized scaling factor. Indicates the number of successfully alerted events. Indicates the first Weighting factors for equipment failure types Indicates the first The baseline loss value for this type of failure. Indicates the time decay coefficient. Indicates the first The lead time of an early warning event. This represents the normalized offset.

[0142] For example, the timeliness compensation value is obtained, the appropriate nonlinear normalization function is determined, and the corresponding normalization scaling factor and normalization offset are set. The timeliness compensation value is input into the nonlinear normalization function, and the numerical fluctuation range of the timeliness compensation value is compressed through function operation. This eliminates the numerical scale difference between the timeliness compensation values ​​corresponding to different successful early warning events, so that the processed result can match the indicator requirements of subsequent reward calculation. This completes the nonlinear normalization processing operation and generates the event value coefficient.

[0143] Among them, nonlinear normalization is a process of adjusting the numerical range of the timeliness compensation value using a specific function; the event value coefficient is a core indicator that can quantify the value of a successful early warning event after nonlinear normalization.

[0144] In one embodiment, feature fusion is performed on the current battery state of charge, energy income probability distribution curve, and fault risk level time series curve to generate a multi-dimensional state vector, including:

[0145] (1) Normalize the current battery state of charge to generate a standardized charge value. The expression for the standardized charge value is:

[0146]

[0147] in, Represents standardized electricity values. Indicates the current state of battery charge. This represents the minimum value of the battery's state of charge. This represents the maximum value of the battery's state of charge.

[0148] For example, the current battery state of charge (SOC) is acquired. This SOC is collected from the device's battery real-time monitoring unit. The time accuracy and data validity standards for the acquisition are clearly defined, and SOC data that meets the requirements is selected. The minimum and maximum SOC values ​​are determined based on the device's battery design specifications, serving as a normalization baseline. According to preset normalization rules, the selected SOC data is mapped to this baseline range, eliminating absolute scale differences caused by variations in device battery models. This completes the normalization process and generates a standardized charge value.

[0149] Among them, the current battery state of charge is the stored state of power that meets the data validity standards and is collected from the device's real-time battery monitoring module; the standardized power value is a uniform scale power quantification index that is within the reference range of the device's battery design parameters after normalization processing.

[0150] (2) Extract time period features from the energy income probability distribution curve to generate time period energy feature vectors.

[0151] For example, an energy income probability distribution curve is obtained, which is an energy income probability change curve generated based on a lightweight environment feature set. Combining the construction plan operation periods of the building equipment, time segmentation rules are determined, dividing the energy income probability distribution curve into several time periods corresponding to the operation periods. For each time period, core features of the energy income probability distribution are extracted, including the peak position, distribution range, and probability concentration. These core features are then organized into an ordered feature set according to the time sequence of the corresponding time period, completing the time period feature extraction processing operation and generating a time period energy feature vector.

[0152] Among them, the energy income probability distribution curve is a curve that reflects the change of energy income probability at different time points; the time period energy feature vector is an ordered feature set containing the core features of energy income distribution in each work period after time period feature extraction.

[0153] (3) Extract key points from the time series curve of fault risk level and generate risk key feature vector.

[0154] For example, a time-series curve of fault risk level is obtained, which is a curve of equipment fault risk level change generated based on a lightweight environment feature set. A preset criterion for judging changes in fault risk level is established, clearly defining nodes where the risk level crosses a preset threshold or the magnitude of the risk level change exceeds a preset range as significant change nodes. Time nodes in the fault risk level time-series curve that meet this criterion are identified, and the corresponding risk level value, risk level change magnitude, and corresponding time point information for each node are extracted. This information is then organized into an ordered feature set according to the chronological order of the time nodes, completing the key point extraction processing operation and generating a risk key feature vector.

[0155] Among them, the fault risk level time series curve is a curve that reflects the change of equipment fault risk level at different time points; the risk key feature vector is an ordered feature set containing relevant information of nodes with significant changes in fault risk level after key points are extracted.

[0156] (4) Perform feature concatenation processing on the standardized electricity value, time period energy feature vector, and risk key feature vector to generate an initial fusion vector. The expression of the initial fusion vector is:

[0157]

[0158] in, Represents the initial fusion vector. This indicates a feature concatenation operation. Represents standardized electricity values. Represents the energy feature vector over a time period. This represents the key feature vector of risk.

[0159] For example, standardized electricity values, time-period energy feature vectors, and risk-critical feature vectors are obtained. Based on the correlation between various features and the operating status of building equipment, the order of feature concatenation is determined, typically in the order of standardized electricity values, time-period energy feature vectors, and risk-critical feature vectors. Following this order, the feature data corresponding to the standardized electricity values, the ordered feature set of the time-period energy feature vectors, and the ordered feature set of the risk-critical feature vectors are sequentially connected to form a continuous feature sequence. This eliminates the independent storage states between different features, completes the feature concatenation process, and generates an initial fusion vector.

[0160] Among them, the feature splicing operation is an operation that connects features from different sources into a unified sequence in a preset order based on the degree of correlation between features and the device's operating status; the initial fusion vector is a continuous feature sequence that integrates multiple types of features after feature splicing.

[0161] (5) The initial fusion vector is reduced in dimension and compressed to generate a multidimensional state vector.

[0162] For example, an initial fusion vector is obtained, and an appropriate dimensionality reduction and compression method is determined based on the input dimensionality requirements of the subsequent reinforcement learning decision model. This method calculates the contribution of each feature in the initial fusion vector, sets a contribution threshold, retains core feature information with contributions higher than the threshold, and removes redundant features with contributions lower than the threshold or highly correlated repetitive feature content, thereby reducing the number of feature dimensions, completing the dimensionality reduction and compression operation, and generating a multidimensional state vector.

[0163] The initial fusion vector is a continuous feature sequence that integrates features from multiple classes; the multidimensional state vector is a low-dimensional feature vector that retains core features and meets the input dimension requirements of the reinforcement learning decision model after dimensionality reduction and compression.

[0164] In one embodiment, the task criticality prediction is performed on the lightweight environment feature set to generate a time series curve of failure risk level, including:

[0165] (1) Perform semantic spatiotemporal analysis on the construction plan summary of lightweight environmental feature set to generate spatiotemporal distribution of operation risk.

[0166] For example, a lightweight environment feature set is obtained, from which a construction plan summary is extracted. This summary clarifies the types of tasks covered, the specific geographical scope of the tasks, and the time intervals during which the tasks are performed. Combining this with the inherent risk attributes of different tasks in the construction equipment operation field, the construction plan summary is semantically decomposed to identify risk-related content such as equipment load and operating environment. Simultaneously, a corresponding spatiotemporal framework is constructed using the operational area as the spatial dimension and the operational time period as the temporal dimension. The decomposed risk elements are then matched one by one to the corresponding nodes within this spatiotemporal framework, completing the semantic spatiotemporal parsing operation and generating a spatiotemporal distribution map of operational risks.

[0167] Among them, the lightweight environmental feature set is a feature set containing construction plan summary and regional weather forecast summary obtained after metadata compression; the construction plan summary is the core information content related to construction equipment operation in the construction plan; the spatiotemporal distribution map of operation risk is a risk distribution presentation framework that integrates risk elements obtained from semantic parsing and corresponds to spatiotemporal framework nodes.

[0168] (2) Detect risk clustering areas on the spatiotemporal distribution map of operational risks and generate key risk area markers.

[0169] For example, a spatiotemporal distribution map of operational risks is obtained. Combined with the risk management requirements for construction equipment operations, a criterion for identifying risk clusters is determined. This criterion must be set based on the spatial density level and the concentration of risk levels. According to this criterion, the spatiotemporal framework nodes of the operational risk distribution map are traversed, and a set of nodes with risk density higher than a preset level and risk levels within the same range is selected. The spatial area range and temporal coverage interval corresponding to the above node set are clarified. These areas are then uniquely identified, and their spatiotemporal attributes are recorded. This completes the risk cluster detection operation and generates key risk area markers.

[0170] Among them, the spatiotemporal distribution map of operational risks is a risk distribution presentation framework that integrates risk elements and corresponding spatiotemporal framework nodes; the key risk area marker is a record of the spatial range and time interval of the risk cluster area in the spatiotemporal distribution map of operational risks.

[0171] (3) Perform time-dimensional propagation modeling on the key risk area markers to generate a risk propagation time series. The expression for the risk propagation time series is:

[0172]

[0173] in, This represents the time series of risk transmission. Indicates the number of key risk areas. This indicates the risk intensity of the k-th critical risk area. This represents the time center point of the k-th critical risk area. This represents the time propagation radius of the k-th critical risk area. This represents the time variable, and k represents the index of the key risk area.

[0174] For example, key risk area markers are obtained, and core attribute information such as risk intensity, time center point, and time propagation radius corresponding to each key risk area marker is extracted one by one. With time dimension as the axis, an adapted risk propagation model is constructed. Based on the risk intensity of each key risk area, a diffusion starting point is set. With the time center point as the benchmark and the time propagation radius as the diffusion range boundary, the diffusion process of risk in the key risk area in the time dimension is simulated. The degree of risk diffusion and coverage corresponding to each time point are recorded. The risk status data of all time points are integrated to complete the time dimension propagation modeling processing operation and generate a risk propagation time series.

[0175] Among them, the key risk area marker is the recorded information that identifies the spatial range and time interval of the risk cluster area; the risk propagation time series is the sequence data that records the degree and coverage of the risk spread over time within the key risk area.

[0176] (4) Perform risk level quantification mapping on the risk propagation time series to generate a fault risk level time series curve.

[0177] For example, a risk propagation time series is obtained. Referring to the risk management specifications for building equipment failure early warning, a quantitative mapping rule for risk levels is determined. This rule needs to correlate and match the degree of risk diffusion and coverage corresponding to each time point in the risk propagation time series with a preset failure risk level standard. Based on this rule, the risk status at each time point in the risk propagation time series is converted into a corresponding failure risk level. The failure risk level results of all time points are then systematically integrated in chronological order to complete the risk level quantitative mapping operation and generate a failure risk level time series curve.

[0178] Among them, the risk propagation time series is a sequence of data that records the degree and coverage of risk diffusion over time; the failure risk level time series curve is a curve that presents the failure risk level of building equipment at different points in time in chronological order.

[0179] like Figure 2 As shown, energy income prediction is performed on the lightweight environment feature set, generating an energy income probability distribution curve, including:

[0180] S201: Perform multi-scale feature extraction on weather forecast metadata in the lightweight environmental feature set to generate a weather state feature vector.

[0181] For example, a lightweight environmental feature set is obtained, and weather forecast metadata is extracted from the lightweight environmental feature set. The types of meteorological elements covered by the weather forecast metadata are identified, and a multi-scale division standard is determined. The weather forecast metadata is then divided into time periods with different time granularities according to this standard. For the weather forecast metadata of each time period, the core content such as the changing trend and fluctuation characteristics of the corresponding meteorological elements is extracted. The meteorological element features extracted at different scales are integrated into an ordered set according to a preset order to complete the multi-scale feature extraction operation and generate a weather state feature vector.

[0182] Among them, the lightweight environmental feature set is a feature set containing construction plan summary and regional weather forecast summary after metadata compression; the weather forecast metadata is the core information content related to regional weather in the lightweight environmental feature set; and the weather state feature vector is an ordered set that integrates meteorological element features at different scales.

[0183] S202: Perform probability diffusion modeling on the weather state feature vector to generate an energy income probability distribution.

[0184] For example, a weather state feature vector is obtained, and the correlation logic between the meteorological elements in the weather state feature vector and energy income is clarified. Based on this correlation logic, a probability diffusion model is constructed. The weather state feature vector is input into the probability diffusion model to simulate the possible range of energy income variation under different meteorological conditions, determine the probability level corresponding to each possible energy income value, integrate the correlation between the possible energy income values ​​and the corresponding probability levels, complete the probability diffusion modeling process, and generate an energy income probability distribution.

[0185] Among them, the weather state feature vector is an ordered set that integrates the features of meteorological elements at different scales; the energy income probability distribution is a set of correlations between the possible values ​​of energy income and the corresponding probability levels.

[0186] S203: Perform time series alignment on the energy income probability distribution to generate an energy income probability distribution curve.

[0187] For example, an energy income probability distribution is obtained, and an alignment benchmark for the time series is determined, with the operating periods of building equipment as the time axis. The probability level corresponding to each possible energy income value in the energy income probability distribution is matched to the corresponding time period of the time axis benchmark. The energy income probability information of different time periods is integrated in chronological order of the time axis, and the changing trend of this information over time is presented in the form of a curve, thus completing the time series alignment operation and generating an energy income probability distribution curve.

[0188] Among them, the energy income probability distribution is the set of associations between possible energy income values ​​and corresponding probability levels; the energy income probability distribution curve is a curve showing the trend of energy income probability changes according to the operating period of building equipment.

[0189] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0190] In one embodiment, such as Figure 3 As shown, this application also provides a building equipment fault prediction and management system 300 based on edge AI, the system 300 including:

[0191] The environmental feature compression module 301 is used to compress metadata of the construction plan summary and regional weather forecast summary sent from the cloud to generate a lightweight environmental feature set; and to perform energy income prediction on the lightweight environmental feature set to generate an energy income probability distribution curve.

[0192] The risk state fusion module 302 is used to predict the mission criticality of the lightweight environment feature set and generate a time series curve of the fault risk level; it also performs feature fusion on the current battery state of charge, energy income probability distribution curve and fault risk level time series curve to generate a multi-dimensional state vector.

[0193] The decision execution module 303 is used to perform reinforcement learning inference on the multi-dimensional state vector to generate joint control actions that include sensor frequency selection, model complexity selection, and upload strategy selection; and executes the joint control actions to control the data acquisition, model calculation, and data transmission operations during the operation of the equipment.

[0194] The feedback learning optimization module 304 is used to monitor the equipment operating status after executing joint control actions, calculate rewards based on the actual early warning effect and survival status, and generate long-term reward signals; it also performs parameter update processing on the reinforcement learning decision model and long-term reward signals stored in the local model memory to generate optimized reinforcement learning strategies.

[0195] Specifically, the environmental feature compression module 301 receives the construction plan summary and regional weather forecast summary sent from the cloud, clarifies the core information categories related to the prediction of building equipment failure in the above two summaries, filters out the metadata content within the corresponding categories, removes irrelevant redundant metadata, and integrates the filtered metadata in a structured manner to complete the metadata compression operation and generate a lightweight environmental feature set.

[0196] Based on the lightweight environmental feature set, environmental information related to energy revenue is extracted, the impact of the above environmental information on equipment energy revenue is analyzed, trend and probability analysis of energy revenue is performed, the correspondence between environmental information and energy revenue is constructed, energy revenue prediction is completed, and an energy revenue probability distribution curve is generated.

[0197] Among them, the construction plan summary is the core information related to the construction equipment operation tasks distributed from the cloud; the regional weather forecast summary is the core information related to the weather in the operation area distributed from the cloud; the lightweight environmental feature set is a feature set that integrates core metadata after metadata compression; and the energy income probability distribution curve is a curve that shows the changing trend of energy income probability at different time periods.

[0198] The risk status fusion module 302 extracts information related to construction equipment operation tasks based on the lightweight environmental feature set, analyzes the impact of operation tasks on equipment operation, identifies fault risk factors associated with tasks, tracks the changing trends of the above risk factors over time, completes task criticality prediction operations, and generates fault risk level time series curves.

[0199] Obtain the current battery state of charge, extract the time-period features corresponding to the energy income probability distribution curve and the key features corresponding to the fault risk level time series curve, integrate the current battery state of charge with the above features in an orderly manner, eliminate the dimensional differences of different features, complete the feature fusion operation, and generate a multi-dimensional state vector.

[0200] Among them, the fault risk level time series curve is a curve that presents the change of equipment fault risk level in chronological order. The current battery state of charge is the current energy storage state of the device's battery; the multidimensional state vector is a low-dimensional feature vector that integrates multiple types of features.

[0201] The decision execution module 303 acquires a multi-dimensional state vector, inputs the multi-dimensional state vector into the reinforcement learning decision model, and performs inference operations by combining the historical decision logic and real-time state characteristics in the model to determine the adaptation scheme of the sensing frequency, the matching standard of the model complexity, and the execution rules of the upload strategy. It integrates the above schemes, standards and rules to complete the reinforcement learning inference operation and generate a joint control action that includes the selection of sensing frequency, model complexity and upload strategy.

[0202] Based on the joint control actions including sensor frequency selection, model complexity selection, and upload strategy selection, the interval and coverage of equipment data acquisition are adjusted, the edge-side fault prediction model with corresponding complexity is matched to carry out calculation processing, the content range and triggering time of data transmission are clarified, and the joint control actions are executed to control the data acquisition, model calculation and data transmission operations during equipment operation.

[0203] Among these, the joint control action is a set of operational instructions that integrates sensing frequency, model complexity, and uploaded strategy decision content. Data acquisition, model calculation, and data transmission operations are the core resource-consuming links in the equipment operation process.

[0204] The feedback learning optimization module 304 continuously tracks the equipment's operating status after executing joint control actions, collects the equipment's early warning trigger records, actual fault occurrences, energy consumption data, and continuous operating time, analyzes the value of the actual early warning effect and the benefit level corresponding to the survival status, and completes the reward calculation operation in combination with preset evaluation rules to generate a long-term reward signal.

[0205] The reinforcement learning decision model stored in the local model memory is retrieved. The long-term reward signal is used as the core basis for parameter adjustment. The decision weights and correlation logic inside the model are adjusted. The decision adaptability of the model is optimized through multiple rounds of calibration, the parameter update processing operation is completed, and the optimized reinforcement learning policy is generated.

[0206] Among them, the long-term reward signal is a quantitative reward indicator that comprehensively reflects the effectiveness of safety early warning and the survival status of equipment. The reinforcement learning decision model is a decision reasoning model stored in the local model memory; the optimized reinforcement learning strategy is a set of decision rules with stronger adaptability after parameter updates.

[0207] The feedback learning optimization module 304 is also used for:

[0208] Successful early warning identification is performed on the sequence of early warning events generated after the execution of joint control actions, and a set of successful early warning events is generated.

[0209] The successful early warning event set is value-quantified and mapped to generate event value coefficients;

[0210] The sequence of fault events that occurs during the execution of joint control actions is used to identify critical fault omissions and generate a set of critical omission events.

[0211] Calculate the penalty coefficient for the key missed event set to generate the missed event penalty coefficient;

[0212] Dynamic energy consumption assessment is performed on the energy consumption data during equipment operation to generate an energy consumption penalty coefficient;

[0213] Calculate the lifecycle gain for the continuous operating time of the equipment and generate the lifecycle gain coefficient;

[0214] The event value coefficient, underreporting penalty coefficient, energy consumption penalty coefficient, and survival gain coefficient are weighted and fused to generate a long-term reward signal.

[0215] The feedback learning optimization module 304 is also used for:

[0216] The event value coefficient and the underreporting penalty coefficient are converted into a risk balance value to generate a security assessment value.

[0217] The energy consumption penalty coefficient and the survival gain coefficient are converted into an energy efficiency balance to generate a survival assessment value.

[0218] The safety assessment value and the survival assessment value are dynamically weighted and aggregated to generate an initial reward value. The expression for the initial reward value is:

[0219]

[0220] in, This represents the initial reward value. Indicates dynamic balancing weights. Indicates the safety assessment value. Indicates the survival assessment value. This represents the scaling factor for the security assessment. Represents the hyperbolic tangent function. This represents the modified linear unit function;

[0221] The initial reward value is time-discounted to generate a long-term reward signal. The expression for the long-term reward signal is:

[0222]

[0223] in, This indicates a long-term reward signal. Indicates the discount factor. Indicates dynamic balancing weights. Indicates the event value coefficient. This represents the penalty coefficient for underreporting. Represents the smoothing constant. Indicates the weight of the survival dimension. Represents the survival gain coefficient. This represents the energy consumption penalty coefficient. This represents the energy consumption sensitivity coefficient.

[0224] The feedback learning optimization module 304 is also used for:

[0225] Perform a fault loss assessment on the set of successful early warning events to generate the expected loss reduction value;

[0226] Time decay compensation is applied to the expected reduction in loss to generate a time-sensitive compensation value;

[0227] The timeliness compensation value is subjected to nonlinear normalization to generate the event value coefficient. The expression for the event value coefficient is as follows:

[0228]

[0229] in, Indicates the event value coefficient. This represents the Sigmoid normalization function. Indicates the normalized scaling factor. Indicates the number of successfully alerted events. Indicates the first Weighting factors for equipment failure types Indicates the first The baseline loss value for this type of failure. Indicates the time decay coefficient. Indicates the first The lead time of an early warning event. This represents the normalized offset.

[0230] Risk status fusion module 302 is also used for:

[0231] The current battery state of charge is normalized to generate a standardized charge value. The expression for the standardized charge value is:

[0232]

[0233] in, Represents standardized electricity values. Indicates the current state of battery charge. This represents the minimum value of the battery's state of charge. This indicates the maximum value of the battery's state of charge.

[0234] The energy income probability distribution curve is processed to extract time-period features, generating a time-period energy feature vector.

[0235] Key points are extracted from the time-series curve of fault risk level to generate a risk key feature vector.

[0236] The standardized energy values, time-period energy feature vectors, and risk-critical feature vectors are concatenated to generate an initial fusion vector. The expression for the initial fusion vector is as follows:

[0237]

[0238] in, Represents the initial fusion vector. This indicates a feature concatenation operation. Represents standardized electricity values. Represents the energy feature vector over a time period. Represents the key feature vector of risk;

[0239] The initial fusion vector is reduced in dimension and compressed to generate a multidimensional state vector.

[0240] Risk status fusion module 302 is also used for:

[0241] Semantic spatiotemporal analysis is performed on the construction plan summary with lightweight environmental feature set to generate a spatiotemporal distribution map of operation risk;

[0242] Detect risk clustering areas in the spatiotemporal distribution map of operational risks and generate key risk area markers;

[0243] The key risk areas are marked with time-dimensional propagation modeling to generate a risk propagation time series. The expression for the risk propagation time series is as follows:

[0244]

[0245] in, This represents the time series of risk transmission. Indicates the number of key risk areas. This indicates the risk intensity of the k-th critical risk area. This represents the time center point of the k-th critical risk area. This represents the time propagation radius of the k-th critical risk area. This represents a time variable, and k represents the index of the key risk area.

[0246] Risk levels are quantified and mapped to the time series of risk propagation to generate time series curves of failure risk levels.

[0247] The environmental feature compression module 301 is also used for:

[0248] Multi-scale feature extraction is performed on weather forecast metadata in the lightweight environmental feature set to generate weather state feature vectors;

[0249] Probabilistic diffusion modeling is performed on weather state feature vectors to generate an energy income probability distribution;

[0250] The energy income probability distribution is aligned with the time series to generate an energy income probability distribution curve.

[0251] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0252] In one embodiment, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0253] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0254] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predictive management of building equipment failures based on edge AI, characterized in that, The method includes: Metadata compression is performed on the construction plan summary and regional weather forecast summary sent from the cloud to generate a lightweight environmental feature set; energy income prediction is performed on the lightweight environmental feature set to generate an energy income probability distribution curve; The task criticality of the lightweight environment feature set is predicted to generate a time series curve of the fault risk level; the current battery state of charge, the energy income probability distribution curve and the time series curve of the fault risk level are fused to generate a multi-dimensional state vector. The multidimensional state vector is subjected to reinforcement learning inference to generate a joint control action that includes sensor frequency selection, model complexity selection, and upload strategy selection; the joint control action is executed to control the data acquisition, model calculation, and data transmission operations during the operation of the device. The system monitors the equipment's operating status after executing the joint control action, calculates rewards based on the actual early warning effect and survival status, and generates a long-term reward signal. It then updates the parameters of the reinforcement learning decision model stored in the local model memory and the long-term reward signal to generate an optimized reinforcement learning strategy. 2.The edge-AI-based building equipment failure prediction management method of claim 1, wherein The monitoring of the equipment's operating status after the execution of the joint control action, the calculation of rewards based on the actual early warning effect and survival status, and the generation of long-term reward signals include: Successful early warning identification is performed on the sequence of early warning events generated after the execution of the joint control action, and a set of successful early warning events is generated. The successful early warning event set is value-quantified and mapped to generate event value coefficients; The sequence of fault events occurring during the execution of the joint control action is used to identify critical fault omissions and generate a set of critical omission events. The penalty coefficient is calculated for the set of key missed events to generate a missed event penalty coefficient; Dynamically assess energy consumption data during equipment operation and generate energy consumption penalty coefficients; Calculate the lifecycle gain for the continuous operating time of the equipment and generate the lifecycle gain coefficient; The event value coefficient, the missed detection penalty coefficient, the energy consumption penalty coefficient, and the survival gain coefficient are weighted and fused by multiple objectives to generate the long-term reward signal. 3.The edge-AI-based building equipment failure prediction management method of claim 2, wherein, The process of performing multi-objective weighted fusion of the event value coefficient, the missed detection penalty coefficient, the energy consumption penalty coefficient, and the survival gain coefficient to generate the long-term reward signal includes: The event value coefficient and the false alarm penalty coefficient are converted into a risk balance value to generate a security assessment value. The energy consumption penalty coefficient and the survival gain coefficient are converted into an energy efficiency balance value to generate a survival assessment value. The safety assessment value and the survival assessment value are dynamically weighted and aggregated to generate an initial reward value, the expression of which is: in, This represents the initial reward value. Indicates dynamic balancing weights. Indicates the safety assessment value. Indicates the survival assessment value. This represents the scaling factor for the security assessment. Represents the hyperbolic tangent function. This represents the modified linear unit function; The initial reward value is time-discounted to generate the long-term reward signal, the expression of which is: in, This indicates a long-term reward signal. Indicates the discount factor. Indicates dynamic balancing weights. Indicates the event value coefficient. This represents the penalty coefficient for underreporting. Represents the smoothing constant. Indicates the weight of the survival dimension. Represents the survival gain coefficient. This represents the energy consumption penalty coefficient. This represents the energy consumption sensitivity coefficient. 4.The edge-AI based building equipment failure prediction management method of claim 2, wherein, The step of quantifying and mapping the successful early warning event set to generate event value coefficients includes: The set of successful early warning events is used to assess the failure loss and generate the expected loss reduction value; Time decay compensation is applied to the expected loss reduction value to generate a timeliness compensation value; The timeliness compensation value is subjected to nonlinear normalization to generate the event value coefficient, the expression of which is: in, Indicates the event value coefficient. This represents the Sigmoid normalization function. Indicates the normalized scaling factor. Indicates the number of successfully alerted events. Indicates the first Weighting factors for equipment failure types Indicates the first The baseline loss value for this type of failure. Indicates the time decay coefficient. Indicates the first The lead time of an early warning event. This represents the normalized offset.

5. The building equipment fault prediction and management method based on edge AI according to claim 1, characterized in that, The feature fusion of the current battery state of charge, the energy income probability distribution curve, and the fault risk level time series curve generates a multi-dimensional state vector, including: The current battery state of charge is normalized to generate a standardized charge value, the expression of which is: in, Represents standardized electricity values. Indicates the current state of battery charge. This represents the minimum value of the battery's state of charge. This indicates the maximum value of the battery's state of charge. The energy income probability distribution curve is subjected to time period feature extraction processing to generate a time period energy feature vector; The time series curve of the fault risk level is processed by extracting key points to generate a risk key feature vector; The standardized energy value, the time period energy feature vector, and the risk key feature vector are subjected to feature concatenation processing to generate an initial fusion vector. The expression of the initial fusion vector is as follows: wherein, represents an initial fusion vector, represents a feature concatenation operation, represents a normalized power value, represents a time period energy feature vector, represents a risk key feature vector; The initial fusion vector is reduced in dimension and compressed to generate the multidimensional state vector. 6.The edge-AI-based building equipment failure prediction management method of claim 1, wherein, The step of predicting the task criticality of the lightweight environment feature set and generating a time-series curve of the failure risk level includes: Semantic spatiotemporal parsing is performed on the construction plan summary of the lightweight environment feature set to generate a spatiotemporal distribution map of operational risks; The spatiotemporal distribution map of the operational risks is used to detect risk clustering areas and generate key risk area markers; The key risk area markers are subjected to time-dimensional propagation modeling to generate a risk propagation time series. The expression of the risk propagation time series is as follows: in, This represents the time series of risk transmission. Indicates the number of key risk areas. This indicates the risk intensity of the k-th critical risk area. This represents the time center point of the k-th critical risk area. This represents the time propagation radius of the k-th critical risk area. This represents a time variable, and k represents the index of the key risk area. The risk propagation time series is subjected to risk level quantification mapping to generate the fault risk level time series curve.

7. The building equipment fault prediction and management method based on edge AI according to claim 1, characterized in that, The step of predicting energy income from the lightweight environment feature set and generating an energy income probability distribution curve includes: Multi-scale feature extraction is performed on the weather forecast metadata in the lightweight environmental feature set to generate a weather state feature vector; The weather state feature vector is subjected to probability diffusion modeling to generate an energy income probability distribution; The energy income probability distribution is time-series aligned to generate the energy income probability distribution curve.

8. The edge AI-based building equipment failure prediction management system, characterized by, The system includes: The environmental feature compression module is used to compress metadata of the construction plan summary and regional weather forecast summary sent from the cloud to generate a lightweight environmental feature set; and to perform energy income prediction on the lightweight environmental feature set to generate an energy income probability distribution curve. The risk state fusion module is used to predict the mission criticality of the lightweight environment feature set and generate a time series curve of the fault risk level; and to fuse the current battery state of charge, the energy income probability distribution curve and the time series curve of the fault risk level to generate a multi-dimensional state vector. The decision execution module is used to perform reinforcement learning inference on the multi-dimensional state vector to generate a joint control action that includes sensor frequency selection, model complexity selection, and upload strategy selection; and executes the joint control action to control the data acquisition, model calculation, and data transmission operations during the operation of the device. The feedback learning optimization module is used to monitor the equipment's operating status after executing the joint control action, calculate rewards based on the actual early warning effect and survival status, and generate a long-term reward signal; it also performs parameter update processing on the reinforcement learning decision model stored in the local model memory and the long-term reward signal to generate an optimized reinforcement learning strategy. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8. When the processor executes the computer program, it implements the steps of the building equipment fault prediction and management method based on edge AI as described in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the building equipment fault prediction and management method based on edge AI as described in any one of claims 1 to 7.

Images