Mechanical and electrical equipment and electrical equipment remote AI monitoring system based on management platform

Abstract

The application discloses a mechanical and electrical equipment and electrical equipment remote AI monitoring system based on a management platform, and relates to the technical field of equipment monitoring.The application comprises a data server, a display unit and a control unit, the control unit comprises an edge computing layer, a data acquisition layer and an AI analysis engine, the AI analysis engine comprises a deep learning module, a photovoltaic optimization module, an agent cluster and a fault tracing module, the output end of the data acquisition layer is connected with the input end of the edge computing layer, the port of the edge computing layer is in bidirectional communication with the port of the data server, and the output end of the edge computing layer is connected with the input end of the deep learning module and the photovoltaic optimization module respectively.The application adopts multiple agent learning technologies, can realize sharing of model parameter updating among multiple parks, protects local data privacy at the same time, constructs an industry expert rule library, cross-verify with deep learning results, and avoids misjudgment caused by the black box model.

Description

Remote AI Monitoring System for Electromechanical and Electrical Equipment Based on Management Platform Technical Field

[0001] This invention relates to the field of equipment monitoring technology, specifically to a remote AI monitoring system for electromechanical and electrical equipment based on a management platform. Background Technology

[0002] The proper functioning of public infrastructure and intelligent terminals is crucial for factories, buildings, and science parks. Monitoring systems for electromechanical and electrical equipment are particularly important for ensuring their proper operation, especially HVAC, power supply, and intelligent terminals. HVAC and power distribution, as fundamental functions of factories, buildings, and science parks, play a vital role in their operation. Photovoltaic power generation, as a clean and pollution-free energy source, has advantages such as simple installation and increasingly high conversion efficiency, leading to its growing prevalence in energy use. Intelligent terminal equipment provides convenience for offices and homes.

[0003] In previous monitoring systems for electromechanical and electrical equipment, fault handling generally relied on real-time reporting, lacking analysis and tracing of the faults. Manual troubleshooting and tracing also lacked timeliness. The technical problems this invention aims to solve are: 1. Providing a monitoring platform for HVAC, power distribution, power generation, and intelligent systems; 2. Using the latest AI technology for intelligent monitoring of related systems; 3. Overcoming the shortcomings of existing platforms that only monitor data without providing expert knowledge; 4. Traditional electromechanical equipment monitoring systems rely solely on sensor data threshold alarms, failing to perform in-depth analysis of complex faults, leading to delayed maintenance responses. Photovoltaic power generation systems lack the ability to predict power generation efficiency degradation, and fault diagnosis relies on manual experience, resulting in low efficiency. Abnormal behavior of intelligent terminals lacks multi-dimensional correlation analysis; for example, intelligent access control or environmental control equipment makes root cause localization difficult. Existing platforms mostly use single data stream processing, failing to integrate equipment operation data, environmental data, and historical maintenance records, resulting in low fault diagnosis accuracy. Therefore, developing an AI monitoring system for real-time risk warning and fault diagnosis of electromechanical and electrical equipment is a technical problem that those skilled in the art need to solve. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a remote AI monitoring system for electromechanical and electrical equipment based on a management platform, which solves the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a remote AI monitoring system for electromechanical and electrical equipment based on a management platform, comprising a data server, a display unit, and a control unit. The control unit includes an edge computing layer, a data acquisition layer, and an AI analysis engine. The AI ​​analysis engine includes a deep learning module, a photovoltaic optimization module, an intelligent agent cluster, and a fault tracing module. The output of the data acquisition layer is connected to the input of the edge computing layer. The port of the edge computing layer establishes bidirectional communication with the port of the data server. The output of the edge computing layer is connected to the inputs of the deep learning module and the photovoltaic optimization module, respectively. The outputs of the deep learning module and the photovoltaic optimization module are both connected to the input of the fault tracing module. The output of the fault tracing module is connected to the input of the intelligent agent cluster. The output of the edge computing layer is connected to the input of the intelligent agent cluster. The output of the intelligent agent cluster is connected to the input of the display unit.

[0006] The data acquisition layer is electrically connected to an external data acquisition device. The data acquisition layer acquires the operating data of electromechanical and electrical equipment through the data acquisition device. The data acquisition objects of the data acquisition device include, but are not limited to, chiller stations, air compressor stations, air conditioning terminals, power distribution equipment, chillers, air compressors, water pumps, cooling towers, air conditioning boxes, distribution cabinets, inverters, combiner boxes, irradiators, weather instruments, electricity meters, and intelligent terminals of photovoltaic power stations. The edge computing layer performs local preprocessing on the acquired operating data to obtain first data. The edge computing layer forwards the first data to the data server for storage. At the same time, the first data is output to the deep learning module and then the photovoltaic optimization module. The deep learning module outputs the remaining lifespan and failure probability of the equipment based on the first data or other data sources and forwards it to the fault tracing module. The photovoltaic optimization module uses generative adversarial networks to simulate theoretical power generation and forwards it to the fault tracing module. The fault tracing module constructs an association network and locates the fault propagation path and forwards it to the intelligent agent cluster.

[0007] When the operational data acquired by the data acquisition layer is insufficient for model training, the agent cluster acquires the first data from the edge computing layer and establishes a digital prediction model to transform the first data into a simulation training set. The agent cluster has several preset agents and an expert rule base. Different expert rules are retrieved from the expert rule base and assigned to each agent. The agents train the model based on the simulation training set. The agent cluster marks the remaining lifespan of the equipment, the failure probability, and the failure propagation path as the real verification set and performs model verification for the agents. It outputs the weight of each agent for each failure point and the handling plan of each agent for each failure point.

[0008] The data server is used to store the operating data acquired by the data collector and the temporary data generated during the operation of the control unit. The temporary data includes the first data, the remaining life of the equipment, the probability of failure, the theoretical power generation and the failure propagation path. The display unit establishes a visual interactive platform for operation and maintenance personnel to view, and displays a three-dimensional equipment topology map. The display unit marks the fault points and the scope of impact in the three-dimensional equipment topology map and generates a multi-dimensional operation and maintenance report.

[0009] Furthermore, the edge computing layer is equipped with a lightweight AI model based on the TinyML framework. This lightweight AI model is used to perform local preprocessing of the runtime data and monitor abnormal data in real time. The specific process is as follows:

[0010] The lightweight AI model scans the collected running data in real time and performs continuous verification in chronological order. If the lightweight AI model finds that data is missing at a certain moment, it performs a data filling process to supplement the data. The lightweight AI model calculates the average value and standard deviation of each type of data in the running data, sets the average value ± 3 times the standard deviation as the screening threshold for that type of data, removes data in the running data whose values ​​exceed the screening threshold, and performs a normalization process to normalize all the running data.

[0011] For non-numerical data in the operational data, such as water pump failure events, the lightweight AI model establishes an encoding library and encodes the non-numerical data. Each non-numerical data corresponds to an independent code. The independent code does not participate in the continuity verification, screening threshold removal and normalization processing of the operational data.

[0012] During the data filling process, the lightweight AI model obtains the values ​​of the adjacent times before and after the data loss time, and the average value of the values ​​of the adjacent times before and after the data loss time is calculated and assigned to the time of data loss.

[0013] Due to the significant differences in the dimensions of data from different devices, it is necessary to standardize the data to the range of [0,1] or [-1,1] to facilitate processing by the TinyML lightweight model. During the normalization process, the lightweight AI model uses a formula based on Z-Score standardization. Normalize all running data values ​​to obtain

[0014] The first data is x, which is the running data, x' is the first data, μ is the historical mean of this type of data, and σ is the historical standard deviation of this type of data. The historical mean μ and the historical standard deviation σ are updated synchronously based on the values ​​of the running data collected in real time.

[0015] Furthermore, the deep learning module establishes a device health prediction model based on a Transformer or LSTM temporal neural network. The model is input with initial data, environmental variables, and maintenance records. The model then outputs the remaining lifespan and failure probability of the device, as detailed below:

[0016] The equipment health prediction model aligns all the first data according to the timestamp of the collection to form a time series sample. The equipment health prediction model divides the first data into training set and validation set in a 9:1 ratio. The types of electromechanical equipment and electrical equipment corresponding to the first data are highly typical, and the model does not need to have strong generalization ability. Therefore, it does not need too much running data as validation set, nor does it need to add an extra test set. The equipment health prediction model is divided into a first prediction model based on the LSTM framework and a second prediction model based on the Transformer framework. The training set is used as the input layer of the first prediction model for model training. The intermediate layer of the first prediction model is 3-5 layers of LSTM units, which are used to analyze the variation pattern between the first data. An Adam optimizer is added to dynamically adjust the learning rate. When the difference between the training set loss and the validation set loss of the first prediction model is stable within 5%, it means that the loss function of the first prediction model has good convergence. The training process of the first prediction model is completed. The output layer of the first prediction model is a dual-branch output, one branch for life prediction and the other branch for fault prediction.

[0017] The results of the first prediction model are input into the second prediction model. The second prediction model adds a self-attention mechanism to establish a temporal correlation between the time series samples and the dual-branch output. After the correlation is completed, the second prediction model outputs the remaining lifespan and failure probability of the device.

[0018] Furthermore, the fault tracing module constructs an equipment association network based on a knowledge graph, and locates the fault propagation path based on the equipment association network, theoretical power generation, remaining equipment lifespan, and fault probability using Bayesian inference and causal analysis algorithms. The causal analysis algorithm adopts the Deep-SeeK framework, and the knowledge graph is obtained by combining and associating the collected objects of the data collector.

[0019] The process of building a device association network is as follows:

[0020] The fault tracing module classifies all collected objects based on a knowledge graph to obtain an entity list, and classifies operational data of the same unit or type into a parameter list. Based on the current transmission direction or the direction of hot and cold circulation, the entity list and parameters are associated to obtain a device association network.

[0021] The fault tracing module locates the fault propagation path.

[0022] Furthermore, the photovoltaic optimization module extracts photovoltaic-related data from the first data, specifically power distribution data and photovoltaic station inverter data, and constructs a photovoltaic dataset based on the photovoltaic-related data;

[0023] The photovoltaic optimization module constructs a GAN model, which includes a generator and a discriminator. Both are optimized through adversarial training using a photovoltaic dataset. The generator aims to generate theoretical values ​​that are close to historical normal power generation, making it indistinguishable to the discriminator. That is, the discriminator's probability of judging the generated value is close to 50%. The discriminator aims to accurately distinguish between the generated value and the real value, outputting 1 for the real value and 0 for the generated value.

[0024] When the discriminator loss stabilizes at around 50%, indicating that the discriminator cannot distinguish between true and false data, and the average error between the generated theoretical power generation and the historical normal power generation is ≤5%, training should be stopped.

[0025] Furthermore, the data acquisition unit and control unit collect, process, archive, and store data from these devices, and upload it to the data server via GPRS, Ethernet, WIFI, etc. Maintenance personnel can view the relevant data through the Internet or the display unit, which facilitates the viewing and management of the operating data of electromechanical and electrical equipment. The operating data includes, but is not limited to, chiller data, water pump data, cooling tower data, power distribution data, photovoltaic station inverter data, sensor data, and intelligent terminal data.

[0026] The chiller data includes, but is not limited to, chiller voltage, chiller current, chiller power, chiller current percentage, chilled water supply temperature, chilled water return temperature, cooling water outlet temperature, cooling water return temperature, chiller operating time, and chiller start-stop count. The water pump data includes, but is not limited to, water pump voltage, water pump current, water pump frequency, water pump fault events, water pump operating time, and water pump start-stop count. The cooling tower data includes, but is not limited to, cooling tower voltage, cooling tower current, cooling tower frequency, cooling tower fault events, cooling tower operating time, and cooling tower start-stop count. The power distribution data includes, but is not limited to, power distribution voltage, power distribution current, and power distribution. The photovoltaic station inverter data includes, but is not limited to, inverter voltage, inverter current, inverter fault events, inverter operating time, and inverter start-stop count. The sensor data includes, but is not limited to, sensor temperature, sensor pressure, sensor differential pressure, flow meter data, irradiance meter data, and meteorological data. The intelligent terminal data includes, but is not limited to, terminal temperature, terminal humidity, and terminal logs.

[0027] The control unit preprocesses the operating data, performs AI modeling, deep learning, archives the data, and establishes expert rules to form an intelligent agent cluster. During the operation of the control unit, the expert rules and the intelligent agent cluster are iterated to form a reasonable expert rule base.

[0028] By integrating DeepSeek into the platform's development environment, and leveraging its advanced data analytics capabilities, the platform can deeply mine data from the database, extract valuable information, and provide strong support for operational and maintenance decisions. DeepSeek's intelligent algorithms, combined with the platform's business logic, enable automated optimization of processes. By utilizing DeepSeek's natural language processing capabilities and combining expertise in HVAC, power distribution, and photovoltaic power generation, the platform can suggest more reasonable fault rule bases.

[0029] Furthermore, the process by which the display unit marks the fault points and their impact range in the 3D device topology diagram is as follows:

[0030] If a device reports a fault event, the data collector acquires the fault event and transmits it to the data acquisition layer. The data acquisition layer then transmits the fault event to the display unit. The display unit obtains the device association network and fault propagation path from the fault tracing module through the intelligent agent cluster. The display unit converts the device association network into a three-dimensional device topology map and matches the fault propagation path according to the fault point corresponding to the fault event, marking the fault propagation path as the fault point and the scope of influence.

[0031] The process by which the display unit generates a multi-dimensional operation and maintenance report is as follows:

[0032] The display unit records the occurrence time, fault location, and affected area for each reported fault event. At the same time, the display unit obtains the corresponding handling plan for each fault location from the intelligent agent cluster. The display unit summarizes the occurrence time, fault location, affected area, and handling plan to obtain an operation and maintenance report.

[0033] Furthermore, the transformation process of the simulated training set specifically includes the following steps:

[0034] The intelligent agent cluster is constructed with n differentiated intelligent agents. Each intelligent agent is equipped with a basic algorithm framework, but the initial parameters are randomized to ensure differentiated parameter tuning. The expert rule base includes the physical laws of equipment, the causal relationship of faults and operation and maintenance experience. m expert rules are randomly selected for each intelligent agent, and repetition is allowed.

[0035] The digital prediction model divides the first data into time periods of different lengths according to time sequence. The weights of n agents are initialized with a value of 1 / n. Each agent makes predictions for different time periods, focusing on only one type of time period. The prediction results are validated against the first data points immediately preceding and following each time period. If the error between the agent's prediction and the first data points of the preceding and following time periods is less than or equal to 5%, it indicates that the agent's prediction is relatively accurate. In this case, the agent's weight is increased by 0.05 / n, and the learning rate is increased by 1%. The weights of the remaining agents are decreased. Conversely, if the error between the agent's prediction result and the first data in the adjacent time period is greater than 5%, it means that the numerical deviation of the agent's prediction result is large. The agent's weight remains unchanged, and the learning rate is reduced by 0.5%.

[0036] The digital prediction model aggregates the prediction results of the agent with the highest weight for each time period of different duration to obtain a simulation training set.

[0037] Furthermore, the intelligent agent's contingency plan for handling fault points specifically includes the following steps:

[0038] After completing training on the simulated training set, the agent acquires first data from random timestamps and random devices and analyzes it to obtain prediction results. The agent compares its prediction results with the results output by the fault tracing module under the same device and timestamp. If the error between the agent's prediction result and the fault tracing module's output result is less than or equal to 3%, it means that the agent's prediction result is relatively accurate. In this case, the agent's weight increases by 0.02 / n, and the learning rate increases by 0.3%. The weights of all other agents decrease. Conversely, if the error between the agent's prediction result and the result output by the fault tracing module is greater than 3%, it means that the numerical deviation of the agent's prediction result is large. The agent's weight remains unchanged, the learning rate is reduced by 0.15%, and the result output by the fault tracing module is the fault point, fault probability, and fault propagation path. The fault point and fault propagation path can be calculated using similarity to convert the fault point and fault propagation path into numerical vectors for comparison.

[0039] After the agent completes training on the simulation verification set, the output result of the agent with the largest weight is marked as the processing plan.

[0040] Furthermore, the fault propagation path localization process is as follows:

[0041] The fault tracing module constructs a Bayesian network probability model. This model predicts the fault point for each device in the entity list based on the remaining lifespan and fault probability of the equipment. The fault point is specific to a particular parameter, mechanical component, or electrical component of the device. A causal analysis algorithm is used to calculate several causes for each fault point. The fault tracing module completes the association path for each fault cause based on the device association network. Then, the module sorts the operational data deviations for each fault cause from largest to smallest, marking the top-ranked association path as the fault propagation path for that fault point. The operational data deviation is the difference between the current operational data value and the historical average μ. Excessive deviations between the theoretical power generation of the photovoltaic optimization module and the actual operational data are also considered fault points for the photovoltaic equipment.

[0042] The present invention has the following beneficial effects:

[0043] 1. Building upon traditional monitoring platforms, AI technology has been incorporated, utilizing the DeepSeek framework or intelligent agents to better analyze and locate fault sources. Leveraging the computing power of AI models provides more rational operational decisions, significantly benefiting HVAC and PV operation and maintenance.

[0044] 2. Based on multimodal data fusion, equipment operation data, video surveillance, and maintenance work order text information are jointly modeled to improve fault identification accuracy and diagnostic accuracy. Dynamic model optimization and the use of multiple intelligent agent learning technology can realize the sharing of model parameter updates across multiple parks, while protecting local data privacy. An industry expert rule base is built and cross-validated with deep learning results to avoid misjudgment caused by black box models.

[0045] 3. By constructing multiple intelligent agents, a training set can be simulated when there is insufficient running data in the early stages of device operation. The results obtained from subsequent analysis based on actual data can then be used for verification. Different expert rules among the intelligent agents can make different predictions for the results, and the best result is output according to the weights, thereby improving the accuracy of model prediction.

[0046] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 is a block diagram of the remote AI monitoring system for electromechanical and electrical equipment based on the management platform of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please refer to Figure 1. This invention provides a technical solution: a remote AI monitoring system for electromechanical and electrical equipment based on a management platform, including a data server, a display unit, and a control unit. The control unit includes an edge computing layer, a data acquisition layer, and an AI analysis engine. The AI ​​analysis engine includes a deep learning module, a photovoltaic optimization module, an intelligent agent cluster, and a fault tracing module. The output of the data acquisition layer is connected to the input of the edge computing layer. The port of the edge computing layer establishes bidirectional communication with the port of the data server. The output of the edge computing layer is connected to the input of the deep learning module and the photovoltaic optimization module, respectively. The outputs of the deep learning module and the photovoltaic optimization module are both connected to the input of the fault tracing module. The output of the fault tracing module is connected to the input of the intelligent agent cluster. The output of the edge computing layer is connected to the input of the intelligent agent cluster. The output of the intelligent agent cluster is connected to the input of the display unit.

[0051] The data acquisition layer is electrically connected to an external data acquisition device. The data acquisition layer acquires the operating data of electromechanical and electrical equipment through the data acquisition device. The data acquisition objects of the data acquisition device include, but are not limited to, chiller stations, air compressor stations, air conditioning terminals, power distribution equipment, chillers, air compressors, water pumps, cooling towers, air conditioning boxes, distribution cabinets, inverters, combiner boxes, irradiators, weather instruments, electricity meters, and intelligent terminals of photovoltaic power stations. The edge computing layer performs local preprocessing on the acquired operating data to obtain the first data. The edge computing layer forwards the first data to the data server for storage. At the same time, the first data is output to the deep learning module and then the photovoltaic optimization module. The deep learning module outputs the remaining lifespan and failure probability of the equipment based on the first data or other data sources and forwards it to the fault tracing module. The photovoltaic optimization module uses generative adversarial networks to simulate the theoretical power generation and forwards it to the fault tracing module. The fault tracing module constructs an association network and locates the fault propagation path and forwards it to the intelligent agent cluster.

[0052] When the operational data acquired by the data acquisition layer is insufficient for model training, the agent cluster acquires the first data from the edge computing layer and establishes a digital prediction model to transform the first data into a simulation training set. The agent cluster has several pre-set agents and an expert rule base. Different expert rules are retrieved from the expert rule base and assigned to each agent. The agents train the model based on the simulation training set. The agent cluster marks the remaining lifespan of the equipment, the failure probability, and the failure propagation path as the real verification set and performs model verification for the agents. It outputs the weight of each agent for each failure point and the handling plan of each agent for each failure point.

[0053] The data server is used to store the operating data acquired by the data acquisition unit and the temporary data generated during the calculation process of the control unit. The temporary data includes the first data, the remaining life of the equipment, the probability of failure, the theoretical power generation and the failure propagation path. The display unit establishes a visual interactive platform for operation and maintenance personnel to view, displaying a three-dimensional equipment topology diagram. The display unit marks the fault points and the scope of impact in the three-dimensional equipment topology diagram and generates a multi-dimensional operation and maintenance report. For example, the display unit displays the content that the efficiency of photovoltaic array #3 has decreased by 12%, and it is recommended to clean and check the string wiring.

[0054] Based on multimodal data fusion, equipment operation data, video surveillance, and maintenance work order text information are jointly modeled to improve fault identification accuracy and diagnostic accuracy. Dynamic model optimization and the use of multiple intelligent agent learning technology can enable the sharing of model parameter updates across multiple parks while protecting local data privacy. An industry expert rule base is built and cross-validated with deep learning results to avoid misjudgments caused by black box models. Video surveillance is based on open source YOLOv8 for object recognition and data extraction, and the extracted data is also operation data.

[0055] The edge computing layer is equipped with a lightweight AI model based on the TinyML framework. This lightweight AI model is used to preprocess the runtime data locally and monitor abnormal data in real time. The specific process is as follows:

[0056] The lightweight AI model scans the collected running data in real time and performs continuous verification in chronological order. If the lightweight AI model finds that data is missing at a certain moment, it performs a data filling process to supplement the data. The lightweight AI model calculates the average value and standard deviation of each type of data in the running data, sets the average value ± 3 times the standard deviation as the screening threshold for that type of data, removes data in the running data whose values ​​exceed the screening threshold, and performs a normalization process to normalize all the running data.

[0057] For non-numerical data in the operational data, such as water pump failure events, the lightweight AI model establishes an encoding library and encodes the non-numerical data. Each non-numerical data corresponds to an independent code. The independent code does not participate in the continuity verification, screening threshold removal and normalization processing of the operational data.

[0058] During the data filling process, the lightweight AI model obtains the values ​​of the adjacent times before and after the data loss time, and the average value of the values ​​of the adjacent times before and after the data loss time is calculated and assigned to the time of data loss.

[0059] Because the units of data from different devices vary greatly—for example, voltage is measured in volts (V) and temperature in degrees Celsius (°C)—it needs to be standardized to a range of [0,1] or [-1,1] to facilitate processing by the TinyML lightweight model. During the normalization process, the lightweight AI model uses a formula based on Z-Score standardization. All operational data values ​​are normalized to obtain the first data, x is the operational data, x' is the first data, μ is the historical mean of this type of data, and σ is the historical standard deviation of this type of data. The historical mean μ and the historical standard deviation σ are updated synchronously based on the values ​​of the operational data collected in real time.

[0060] The deep learning module establishes an equipment health prediction model based on Transformer or LSTM temporal neural networks. Initial data, environmental variables, and maintenance records are input into the model, which then outputs the remaining lifespan and failure probability of the equipment. The specific details are as follows:

[0061] The equipment health prediction model aligns all the first data according to the timestamp of the collection to form a time series sample. The first data is divided into training set and validation set in a 9:1 ratio. The types of electromechanical equipment and electrical equipment corresponding to the first data are highly typical, and the model does not need to have strong generalization ability. Therefore, it does not need too much running data as validation set, nor does it need to add an extra test set. The equipment health prediction model is divided into a first prediction model based on the LSTM framework and a second prediction model based on the Transformer framework. The training set is used as the input layer of the first prediction model for model training. The intermediate layer of the first prediction model is 3-5 layers of LSTM units, which are used to analyze the variation pattern between the first data. An Adam optimizer is added to dynamically adjust the learning rate. When the difference between the training set loss and the validation set loss of the first prediction model is stable within 5%, it means that the loss function of the first prediction model has good convergence. The training process of the first prediction model is completed. The output layer of the first prediction model is a two-branch output, one for life prediction and the other for fault prediction.

[0062] The results of the first prediction model are input into the second prediction model. The second prediction model adds a self-attention mechanism to establish a temporal correlation between the time series samples and the dual-branch output, such as the impact of seasonal changes corresponding to the timestamp on the lifespan of photovoltaic equipment. After the correlation is completed, the second prediction model outputs the remaining lifespan and failure probability of the equipment.

[0063] Among them, the fault tracing module constructs an equipment association network based on a knowledge graph. Based on Bayesian inference and causal analysis algorithms, it locates the fault propagation path according to the equipment association network, theoretical power generation, remaining equipment lifespan, and fault probability. The causal analysis algorithm adopts the Deep-SeeK framework. For example, when an abnormal supply air temperature of the HVAC system is detected, the system automatically associates refrigerant pressure, compressor current, and recent maintenance records to infer that the cause of the fault is evaporator frosting or refrigerant leakage. The knowledge graph is obtained by combining and associating the collected objects of the data collector.

[0064] The process of building a device association network is as follows:

[0065] The fault tracing module classifies all collected objects based on a knowledge graph to obtain an entity list. At the same time, it classifies operational data of the same unit or type into parameter lists. For example, chiller current and cooling tower current are of the same unit, and inverter fault events and cooling tower fault events are of the same type. The entity list and parameters are associated according to the current transmission direction or the direction of hot and cold circulation to obtain the device association network.

[0066] The fault tracing module locates the fault propagation path.

[0067] Among them, the photovoltaic optimization module extracts photovoltaic-related data from the first data, which specifically includes power distribution data and photovoltaic station inverter data, and constructs a photovoltaic dataset based on the photovoltaic-related data.

[0068] The photovoltaic optimization module constructs a GAN model, which includes a generator and a discriminator. Both are optimized through adversarial training using a photovoltaic dataset. The generator aims to generate theoretical values ​​that are close to historical normal power generation, making it indistinguishable to the discriminator. That is, the discriminator's probability of judging the generated value is close to 50%. The discriminator aims to accurately distinguish between the generated value and the real value, outputting 1 for the real value and 0 for the generated value.

[0069] When the discriminator loss stabilizes at around 50%, indicating that the discriminator cannot distinguish between true and false data, and the average error between the generated theoretical power generation and the historical normal power generation is ≤5%, training should be stopped.

[0070] The data acquisition unit and control unit collect, process, archive, and store data from these devices, and upload it to the data server via GPRS, Ethernet, WIFI, etc. Maintenance personnel can view the relevant data through the Internet or the display unit, which facilitates the viewing and management of the operating data of electromechanical and electrical equipment. The operating data includes, but is not limited to, chiller data, water pump data, cooling tower data, power distribution data, photovoltaic station inverter data, sensor data, and intelligent terminal data.

[0071] Chiller data includes, but is not limited to, chiller voltage, chiller current, chiller power, chiller current percentage, chilled water supply temperature, chilled water return temperature, cooling water outlet temperature, cooling water return temperature, chiller operating time, and chiller start-stop count. Pump data includes, but is not limited to, pump voltage, pump current, pump frequency, pump failure events, pump operating time, and pump start-stop count. Cooling tower data includes, but is not limited to, cooling tower voltage, cooling tower current, cooling tower frequency, cooling tower failure events, cooling tower operating time, and cooling tower start-stop count. Power distribution data includes, but is not limited to, power distribution voltage, power distribution current, and power distribution. Photovoltaic station inverter data includes, but is not limited to, inverter voltage, inverter current, inverter failure events, inverter operating time, and inverter start-stop count. Sensor data includes, but is not limited to, sensor temperature, sensor pressure, sensor differential pressure, flow meter data, irradiance meter data, and meteorological data. Intelligent terminal data includes, but is not limited to, terminal temperature, terminal humidity, and terminal logs.

[0072] The control unit preprocesses the operational data, performs AI modeling, deep learning, archives the data, and establishes expert rules to form an intelligent agent cluster. During the operation of the control unit, the expert rules and the intelligent agent cluster are iterated to form a reasonable expert rule base.

[0073] By integrating DeepSeek into the platform's development environment, and leveraging its advanced data analytics capabilities, the platform can deeply mine data from the database, extract valuable information, and provide strong support for operational and maintenance decisions. DeepSeek's intelligent algorithms, combined with the platform's business logic, enable automated optimization of processes, such as intelligent prediction of equipment failures. Utilizing DeepSeek's natural language processing capabilities, and combining expertise in HVAC, power distribution, and photovoltaic power generation, the platform can suggest more reasonable fault rule bases.

[0074] The process by which the display unit marks fault points and their impact range in the 3D device topology diagram is as follows:

[0075] If a device reports a fault event, the data collector acquires the fault event and transmits it to the data acquisition layer. The data acquisition layer then transmits the fault event to the display unit. The display unit obtains the device-related network and fault propagation path from the fault tracing module through the intelligent agent cluster. The display unit converts the device-related network into a 3D device topology map and matches the fault propagation path according to the fault point corresponding to the fault event, marking the fault propagation path as the fault point and its scope of influence.

[0076] The process of generating multi-dimensional operation and maintenance reports by the display unit is as follows:

[0077] The display unit records the occurrence time, fault location, and affected area for each reported fault event. At the same time, the display unit obtains the corresponding handling plan for each fault location from the intelligent agent cluster. The display unit summarizes the occurrence time, fault location, affected area, and handling plan to obtain an operation and maintenance report.

[0078] The transformation process of the simulated training set specifically includes the following steps:

[0079] The agent cluster constructs n differentiated agents, each equipped with a basic algorithm framework, such as a simplified version of Transformer or LSTM, but with randomized initial parameters to ensure differentiated parameter tuning. The expert rule base includes equipment physical laws, fault causal relationships, and operation and maintenance experience. Equipment physical laws include overvoltage protection triggered when the chiller voltage exceeds 240V for 10 minutes. Fault causal relationships include increased fan current due to filter blockage, which in turn leads to abnormal air supply temperature. Operation and maintenance experience includes a 15% ± 5% increase in power generation after cleaning photovoltaic panels. m expert rules are randomly selected for each agent, allowing repetition to ensure diversity.

[0080] The digital prediction model divides the initial data into time segments of varying lengths according to chronological order. The weights of n agents are initialized with a value of 1 / n. Each agent makes predictions for a given time segment, and each agent focuses on only one segment. The prediction results are validated against the first data points immediately preceding and following each segment. If the error between the agent's prediction and the first data points of the preceding and following segments is less than or equal to 5%, the prediction is considered relatively accurate. In this case, the agent's weight is increased by 0.05 / n, and the learning rate is increased by 1%. The weights of the remaining agents are decreased. Conversely, if the error between the agent's prediction result and the first data in the adjacent time period is greater than 5%, it means that the numerical deviation of the agent's prediction result is large. The agent's weight remains unchanged, and the learning rate is reduced by 0.5%.

[0081] The digital prediction model aggregates the prediction results of the agent with the highest weight for each time period of different duration to obtain the simulation training set.

[0082] The intelligent agent's contingency plan for handling fault points specifically includes the following steps:

[0083] After completing training on the simulated training set, the agent acquires first data from random timestamps and random devices and analyzes it to obtain prediction results. The agent compares its prediction results with the results output by the fault tracing module under the same device and timestamp. If the error between the agent's prediction result and the fault tracing module's output result is less than or equal to 3%, it means that the agent's prediction result is relatively accurate. In this case, the agent's weight increases by 0.02 / n, and the learning rate increases by 0.3%. The weights of all other agents decrease. Conversely, if the error between the agent's prediction result and the result output by the fault tracing module is greater than 3%, it means that the numerical deviation of the agent's prediction result is large. The agent's weight remains unchanged, the learning rate is reduced by 0.15%, and the result output by the fault tracing module is the fault point, fault probability, and fault propagation path. The fault point and fault propagation path can be calculated using similarity to convert the fault point and fault propagation path into numerical vectors for comparison.

[0084] After the agent completes training on the simulation validation set, the output of the agent with the largest weight is marked as the processing plan.

[0085] The fault propagation path localization process is as follows:

[0086] The fault tracing module constructs a Bayesian network probability model. Based on the remaining lifespan of the equipment and the probability of failure, the Bayesian network probability model predicts the fault point of each device in the entity list. The fault point is specific to a certain parameter of the equipment or a certain mechanical or electrical component. The causal analysis algorithm is used to calculate several causes of failure for each fault point. The fault tracing module completes the association path for each fault cause based on the equipment association network. The fault tracing module then sorts the operating data deviation corresponding to each fault cause from largest to smallest, and marks the association path at the top of the sort as the fault propagation path of the fault point. The degree of operating data deviation is the difference between the current operating data value and the historical average μ. When the theoretical power generation of the photovoltaic optimization module deviates too much from the actual operating data, it is also regarded as a fault point of the photovoltaic equipment.

[0087] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A remote AI monitoring system for electromechanical and electrical equipment based on a management platform, comprising a data server, a display unit, and a control unit, wherein the control unit includes an edge computing layer, a data acquisition layer, and an AI analysis engine, and the AI ​​analysis engine includes a deep learning module, a photovoltaic optimization module, an intelligent agent cluster, and a fault tracing module; characterized in that: The data acquisition layer acquires operational data of electromechanical and electrical equipment through a data collector. The edge computing layer preprocesses the operational data to obtain first data, which is then forwarded to a data server for storage. The first data is then output to a deep learning module and subsequently to a photovoltaic optimization module. The deep learning module outputs the remaining lifespan and failure probability of the equipment and forwards it to a fault tracing module. The photovoltaic optimization module simulates theoretical power generation and forwards it to the fault tracing module. The fault tracing module locates the fault propagation path and forwards it to the intelligent agent cluster. When the operational data acquired by the data acquisition layer is insufficient for model training, the intelligent agent cluster obtains the first data from the edge computing layer. The system first transforms the first data into a simulated training set to build a digital prediction model. The intelligent agent cluster has several pre-set intelligent agents and an expert rule base. Different expert rules are retrieved from the expert rule base and assigned to each intelligent agent. The intelligent agents train their models based on the simulated training set. The intelligent agent cluster marks the remaining lifespan of the equipment, the probability of failure, and the failure propagation path as a real-world validation set and performs model validation on the intelligent agents. It outputs the weight of each intelligent agent for each failure point and the handling plan for each failure point. The photovoltaic optimization module extracts photovoltaic-related data from the first data and constructs a photovoltaic dataset based on this data. The photovoltaic optimization module then constructs a GAN model. The GAN model includes a generator and a discriminator, both optimized through adversarial training using a photovoltaic dataset. The generator aims to generate theoretical values ​​that closely approximate historical normal power generation. The discriminator's probability of judging the generated values ​​is close to 50%, and its goal is to accurately distinguish between the generated and real values, outputting 1 for real values ​​and 0 for generated values. Training stops when the discriminator's loss stabilizes at around 50% and the average error between the generated theoretical power generation and historical normal power generation is ≤5%. The transformation process of the simulated training set specifically includes the following steps: the agent cluster constructs n differentiated agents, each equipped with a basic algorithm framework, but with randomized initial parameters. The expert rule base includes equipment and object... Based on the principles of logic, causal relationships of faults, and operational experience, m expert rules are randomly selected for each agent, allowing repetition. The digital prediction model divides the first data into time periods of different lengths according to chronological order, initializes the weights of n agents with an initial value of 1 / n, and the n agents perform time predictions for different time periods. Each agent only targets one type of time period. The prediction result of an agent is verified by the first data adjacent to the time period. If the error between the agent's prediction result and the first data of the adjacent time period is less than or equal to 5%, the weight of that agent is increased by 0.05 / n, the learning rate is increased by 1%, and the weights of the other agents are decreased. Conversely, if the error between the agent's prediction result and the first data in the adjacent time periods is greater than 5%, the agent's weight remains unchanged, and the learning rate decreases by 0.5%. The digital prediction model summarizes the prediction results of the agent with the largest weight corresponding to each time period of different duration to obtain a simulation training set. The agent's handling plan for fault points specifically includes the following steps: the agent obtains the first data of random timestamps and random devices and analyzes it to obtain prediction results. The agent compares the prediction results with the results output by the fault tracing module under the same device and the same timestamp. If the error between the agent's prediction result and the results output by the fault tracing module is less than or equal to 3%, the agent's weight increases by 0.02 / n, the learning rate increases by 0.3%, and the weights of the remaining agents decrease. Conversely, if the error between the prediction result of the agent and the output result of the fault tracing module is greater than 3%, the weight of the agent remains unchanged and the learning rate is reduced by 0.15%. After the agent completes the training of the simulation verification set, the output result of the agent with the largest weight is marked as the processing plan.

2. The remote AI monitoring system for electromechanical and electrical equipment based on a management platform according to claim 1, characterized in that, The output of the data acquisition layer is connected to the input of the edge computing layer. The port of the edge computing layer establishes bidirectional communication with the port of the data server. The output of the edge computing layer is connected to the inputs of the deep learning module and the photovoltaic optimization module, respectively. The outputs of both the deep learning module and the photovoltaic optimization module are connected to the input of the fault tracing module. The output of the fault tracing module is connected to the input of the intelligent agent cluster. The output of the edge computing layer is connected to the input of the intelligent agent cluster, and the output of the intelligent agent cluster is connected to the input of the display unit. The edge computing layer carries a lightweight AI model, which performs local preprocessing of the running data and monitors abnormal data in real time. The specific process is as follows: The lightweight AI model scans the running data in real time and performs continuity verification. If data loss is detected at a certain moment, a data filling process is executed to replenish it. The lightweight AI model calculates the mean and standard deviation of each type of data in the running data, and sets the mean ± 3 times the standard deviation as the screening threshold for that type of data. The lightweight AI model removes data whose values ​​exceed the screening threshold. The lightweight AI model then performs a normalization process to normalize all running data. For non-numerical data in the running data, the lightweight AI model establishes an encoding library and encodes it. Each non-numerical data corresponds to an independent code. The independent code does not participate in the continuity verification, screening threshold removal, and normalization processing of the running data. When the data filling process is executed, the values ​​of the adjacent moments before and after the data loss moment are obtained, and the average value of the values ​​of the adjacent moments before and after is calculated and assigned to the moment when the data was lost. During the normalization process, the formula is used based on Z-Score standardization. Normalize all the running data values ​​to obtain the first data, where x is the running data and x' is the first data. This is the historical average of this type of data. The historical standard deviation and historical mean of this type of data. Compared with historical standard deviation The data is updated synchronously based on the real-time collected operational data.

3. The remote AI monitoring system for electromechanical and electrical equipment based on a management platform according to claim 1, characterized in that, The deep learning module establishes an equipment health prediction model based on a temporal neural network. First data, environmental variables, and maintenance records are input into the model, which then outputs the remaining lifespan and failure probability of the equipment. Specifically: The model aligns all first data according to the timestamps at the time of collection to form time-series samples. It then divides the first data into training and validation sets in a 9:1 ratio. The model consists of a first prediction model and a second prediction model. The training set is used as the input layer for training the first prediction model, which has 3-5 intermediate layers. An optimizer is added to dynamically adjust the learning rate. Training is complete when the difference between the training set loss and the validation set loss is within 5%. The first prediction model has a dual-branch output: one for lifespan prediction and the other for failure prediction. The results from the first prediction model are input into the second prediction model, which incorporates a self-attention mechanism to establish a temporal correlation between the time-series samples and the dual-branch output. After the correlation is established, the second prediction model outputs the remaining lifespan and failure probability of the equipment.

4. The remote AI monitoring system for electromechanical and electrical equipment based on a management platform according to claim 1, characterized in that, The fault tracing module constructs an equipment association network based on a knowledge graph. Using Bayesian inference and causal analysis algorithms, it locates fault propagation paths based on the equipment association network, theoretical power generation, remaining equipment lifespan, and fault probability. The knowledge graph is obtained by combining and associating objects collected by the data collector. The equipment association network construction process is as follows: The fault tracing module classifies all collected objects based on the knowledge graph to obtain an entity list, classifies operational data of the same unit or type into parameter lists, and associates the entity lists and parameters according to the current transmission direction or the direction of hot and cold cycles to obtain the equipment association network. The fault tracing module locates the fault propagation path.

5. The remote AI monitoring system for electromechanical and electrical equipment based on a management platform according to claim 1, characterized in that, The operational data includes chiller data, water pump data, cooling tower data, power distribution data, photovoltaic station inverter data, sensor data, and intelligent terminal data.

6. The remote AI monitoring system for electromechanical and electrical equipment based on a management platform according to claim 1, characterized in that, The process by which the display unit marks fault points and their impact range in the 3D device topology map is as follows: If a device reports a fault event, the data collector acquires the fault event and transmits it to the data acquisition layer. The data acquisition layer transmits the fault event to the display unit. The display unit obtains the device-related network and fault propagation path from the fault tracing module through the intelligent agent cluster. The display unit converts the device-related network into a 3D device topology map and matches the fault propagation path according to the fault event, marking the fault propagation path as the fault point and its impact range. The process by which the display unit generates a multi-dimensional operation and maintenance report is as follows: The display unit records the occurrence time, fault point, and impact range for each reported fault event. The display unit obtains the corresponding handling plan for each fault point from the intelligent agent cluster. The display unit summarizes the occurrence time, fault point, impact range, and handling plan to obtain the operation and maintenance report.

7. The remote AI monitoring system for electromechanical and electrical equipment based on a management platform according to claim 4, characterized in that, The fault propagation path localization process is as follows: The fault tracing module constructs a Bayesian network probability model. The Bayesian network probability model predicts the fault point of each device in the entity list based on the remaining lifespan of the device and the fault probability. It uses a causal analysis algorithm to calculate several fault causes for each fault point. The fault tracing module completes the association path for each fault cause based on the device association network. The fault tracing module then sorts the associated paths for each fault cause from largest to smallest according to the degree of deviation of the operating data. The association path at the top of the sort is marked as the fault propagation path.

Images