Mechanical and electrical equipment management method and system based on internet of things

Abstract

The application discloses a kind of mechanical and electrical equipment management method and system based on Internet of Things, belong to mechanical and electrical equipment management technical field, including the first state vector obtained by preprocessing original data, input first state vector and working condition label, output risk score and feature vector, based on risk score and feature vector, generate first maintenance action, and construct strategy function, execute first maintenance action, after execution is completed, the current state vector of equipment is collected as second state vector, feedback score function is constructed and second state vector is input to obtain feedback score, feedback score, first maintenance action and second state vector are composed training sample, feedback score and training sample are packaged as block chain data unit, and shared access is realized by key strategy.The application realizes real-time monitoring, fault prediction, intelligent maintenance decision in equipment management by combining edge computing and intelligent decision algorithm.

Description

Technical Field

[0001] This invention belongs to the field of electromechanical equipment management technology, and in particular relates to an electromechanical equipment management method and system based on the Internet of Things. Background Technology

[0002] With the continuous development of Internet of Things (IoT) technology, electromechanical equipment management has entered a new era. IoT technology provides real-time monitoring and management capabilities for equipment, enabling remote monitoring, data collection, and analysis through the integration of various sensors and smart devices. However, existing IoT device management systems still face challenges and shortcomings, particularly in areas such as equipment fault prediction, lifecycle management, and multi-device coordination. Most existing systems rely on traditional data acquisition and centralized processing methods, transmitting large amounts of data generated by devices to the cloud for analysis. However, this centralized processing method suffers from latency issues, especially when dealing with a large number of devices requiring real-time monitoring. Network bandwidth and computing resource bottlenecks make it difficult for the device management system to guarantee efficient response. Simultaneously, the diversity of devices and the complexity of operating environments place higher demands on fault prediction and maintenance decisions. Many existing systems employ rule-based maintenance strategies, failing to dynamically optimize based on the real-time operating status of equipment, resulting in low equipment management efficiency, high maintenance costs, and long equipment downtime. Furthermore, most existing technologies lack effective fault warning mechanisms and adaptive decision-making models, making it difficult to adapt to the operational needs of different types of equipment and unable to provide effective early warning and scheduling before equipment failures occur.

[0003] Furthermore, existing systems generally suffer from issues of data credibility and transparency. Traditional equipment management systems often rely on centralized data storage and management, resulting in low data security and vulnerability to tampering or loss. This is especially problematic in multi-party environments, where ensuring the transparency and credibility of maintenance records, fault histories, and repair plans becomes a significant challenge. Because equipment management involves various stakeholders, such as equipment manufacturers, operators, and maintenance providers, existing systems often lack effective cross-departmental data sharing and transparent management. Without reliable and transparent data, equipment management decisions struggle to gain the trust of all parties and are difficult to implement promptly, severely impacting the efficiency and effectiveness of equipment management.

[0004] To address these issues, we propose an IoT-based method and system for managing electromechanical equipment. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of data latency and low efficiency in the prior art, and to propose a method and system for managing electromechanical equipment based on the Internet of Things.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for managing electromechanical equipment based on the Internet of Things (IoT), comprising:

[0008] Collect raw data, including vibration signals, current signals, equipment housing temperature, and operating condition labels; preprocess the raw data to obtain a first state vector;

[0009] Input a first state vector and a working condition label, and output a risk score and a feature vector. The risk score is obtained by remapping historical states through an affine transformation mapping function, and the feature vector is obtained by concatenating the current state and the fused state. The current state is obtained from the original data, and the fused state is obtained by dynamically activating historical states.

[0010] Based on risk scores and feature vectors, a first maintenance action is generated, and a policy function is constructed. The first maintenance action is obtained through a distribution in the action space, which is mapped to five levels of risk response actions. The policy function is implemented through a dual-branch network to realize the probability distribution of the output in the action space, and a policy smoothing term and an entropy penalty term are introduced.

[0011] Execute the first maintenance action. After the execution is completed, collect the current state vector of the device as the second state vector. Construct a feedback scoring function and input the second state vector to obtain the feedback score. Use the feedback score, the first maintenance action and the second state vector to form a training sample.

[0012] Feedback scores and training samples are encapsulated into blockchain data units, and shared access is achieved through a key policy.

[0013] Preferably, preprocessing includes data windowing, data alignment, and feature extraction.

[0014] Preferably, the first state vector includes the average value of the vibration signal, the standard deviation of the vibration signal, the average value of the current signal, the rate of change of current within the window time, the average value of temperature, and the rate of change of temperature.

[0015] Preferably, the fusion state is achieved by remapping the historical state through an affine transformation mapping function based on the current working condition label. Each working condition corresponds to a set of linear parameters. The similarity between the current state and the historical state is calculated through the working condition-related matrix, and the weight is obtained through normalization. Finally, the weighted sum of different historical states is obtained.

[0016] Preferred Level 5 risk response actions include: continue operation; slightly reduce load during operation; dispatch manual inspection; automatically enter maintenance preparation state; and immediately shut down.

[0017] Preferably, the dual-branch network includes a state branch and a risk branch. The state branch extracts action preferences through two linear layers plus ReLU, and the risk branch performs risk dynamic recalibration on the action preference weights through a Sigmoid modulation layer.

[0018] Preferably, the strategy smoothing term represents the KL divergence between the current strategy distribution and the previous time step distribution, which is used to keep the strategy stable under stable operating conditions.

[0019] Preferably, a feedback confidence weighting factor is introduced into the training samples. The feedback confidence weighting factor is calculated based on the risk score and is used for sampling during subsequent training.

[0020] An IoT-based electromechanical equipment management system includes:

[0021] The data acquisition module is used to acquire various physical signals from the device and preprocess the various physical signals into a first state vector;

[0022] The risk assessment module generates a risk score and a feature vector using a first state vector and a working condition label. The risk score is obtained by remapping historical states through an affine transformation mapping function, and the feature vector is obtained by concatenating the current state and the fused state. The current state is obtained from the original data, and the fused state is obtained by dynamically activating historical states.

[0023] A strategy generation module is used to generate a first maintenance action and construct a strategy function to obtain the maintenance strategy distribution at the current time.

[0024] The execution feedback module performs a first maintenance action and collects a second state vector after execution. It constructs a feedback score and inputs the second state vector to obtain the feedback score. The feedback score, the first maintenance action, and the second state vector are used to form a training sample.

[0025] The information encapsulation and sharing module encapsulates feedback scores and training samples into blockchain data units and enables shared access through a key strategy.

[0026] In summary, the technical effects and advantages of this invention are as follows: This IoT-based electromechanical equipment management method and system, through data preprocessing and preliminary analysis at edge computing nodes, can significantly reduce response latency in equipment management, ensuring real-time monitoring and fault early warning during actual operation. The introduction of edge computing not only solves the network bandwidth problem in existing systems but also enables dynamic optimization based on real-time equipment data and environmental changes, ensuring the efficiency and real-time performance of the equipment management system. Regarding equipment fault prediction, this invention employs an intelligent decision-making algorithm, integrating historical data, real-time data, and factors such as equipment type and operating environment to provide more accurate fault warnings and maintenance suggestions. Furthermore, the introduction of decentralized blockchain technology ensures high transparency and credibility in data storage and sharing within the equipment management system. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure in this invention;

[0028] Figure 2 This is a schematic diagram of the structure in this invention. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] like Figure 1 As shown, an IoT-based electromechanical equipment management method includes:

[0031] Collect raw data, including vibration signals, current signals, equipment housing temperature, and operating condition labels; preprocess the raw data to obtain a first state vector;

[0032] Input a first state vector and a working condition label, and output a risk score and a feature vector. The risk score is obtained by remapping historical states through an affine transformation mapping function, and the feature vector is obtained by concatenating the current state and the fused state. The current state is obtained from the original data, and the fused state is obtained by dynamically activating historical states.

[0033] Based on risk scores and feature vectors, a first maintenance action is generated, and a policy function is constructed. The first maintenance action is obtained through a distribution in the action space, which is mapped to five levels of risk response actions. The policy function is implemented through a dual-branch network to realize the probability distribution of the output in the action space, and a policy smoothing term and an entropy penalty term are introduced.

[0034] Execute the first maintenance action. After the execution is completed, collect the current state vector of the device as the second state vector. Construct a feedback scoring function and input the second state vector to obtain the feedback score. Use the feedback score, the first maintenance action and the second state vector to form a training sample.

[0035] Feedback scores and training samples are encapsulated into blockchain data units, and shared access is achieved through a key policy.

[0036] The specific steps are as follows:

[0037] Step 1: Data Acquisition and State Representation Construction

[0038] The goal of this step is to construct a unified, low-dimensional, downstream-adaptable device state representation vector S from the various physical signals collected during the operation of electromechanical equipment. t This is the root input node of the entire system, determining the accuracy and stability of downstream risk identification and maintenance strategy generation. Because the operating status of electromechanical equipment is affected by multiple factors (mechanical condition, electrical characteristics, environmental factors, operating condition switching, etc.), traditional static statistical processing of multi-channel data suffers from insufficient expressive power, time-series distortion, and operating condition mismatch.

[0039] Therefore, this step innovatively proposes a fusion:

[0040] Asynchronous sampling signal alignment mechanism;

[0041] Sparse activation regularization structure based on dynamic selection of operating conditions;

[0042] Disturbance suppression term (used to reduce false abnormal responses of equipment during operating condition switching);

[0043] This enables the construction of a more stable, generalized, and adaptable device status representation that can be adapted to various operating environments.

[0044] Assume the observation window at time t is the first 30 seconds, and the sampling step size is 10 seconds. The input raw dataset is:

[0045] The triaxial vibration signal (1kHz, ADXL355 accelerometer) was synthesized using Euclidean algorithm.

[0046] Current signal (1Hz, Hall effect transformer);

[0047] Equipment casing temperature (0.2Hz, thermocouple);

[0048] C t Operating status label (obtained from the PLC interface, indicating the current speed range and load range).

[0049] All data is locally acquired via edge nodes (such as Jetson Nano) and stored as a sliding window D. [t-30,t] This serves as the foundation for state construction.

[0050] Step details

[0051] (1) Signal preprocessing and statistics within the window

[0052] Alignment and feature extraction are performed on all input signals within the window [t-30,t].

[0053] Calculate the envelope mean μ of the acceleration signal a(t). a,t Standard deviation σ a,t ;

[0054] Extract the mean μ from the current signal I(t). I,t Compared with the most recent 10s increment ΔI t ;

[0055] Temperature signal T(t), extract the 30s mean μ T,t With the current instantaneous slope ΔT t .

[0056] Construct the initial feature representation vector:

[0057]

[0058] in:

[0059] μ a,t : The average value of the vibration signal within the window;

[0060] σ a,t Standard deviation of vibration signal;

[0061] μ I,t : Average value of current signal;

[0062] Rate of change of current over the past 10 seconds;

[0063] μ T,t Average temperature;

[0064] Rate of temperature change.

[0065] (2) Design of state mapping function

[0066] To improve the adaptability of state representation to actual device behavior, a condition-sensitive structure mapping mechanism is introduced. The final state representation construction function is defined as follows:

[0067]

[0068] in:

[0069] The final state vector to be solved;

[0070] Initial state feature vector;

[0071] λ1: Sparse regularization coefficient (e.g., 0.01);

[0072] λ2: Disturbance suppression regularization coefficient (e.g., 0.001);

[0073] β i (C t ): Regarding the current working condition C t The sparsity intensity of the i-th state component (which can be defined by looking up a table);

[0074] A set of indexes for a specific dimension, representing state dimensions that are sensitive to changes in operating conditions;

[0075] This represents the gradient of the state dimension j in response to changes in the working condition label.

[0076] This optimization problem is solved on edge devices using a lightweight optimizer, which has real-time processing capabilities.

[0077] Output

[0078] S t The final state representation vector (6-dimensional) provides input for the subsequent risk assessment module.

[0079] C t : Current equipment operating condition label, used for subsequent fusion model auxiliary judgment and personalized strategy input.

[0080] Step 2: Risk Assessment and Dynamic Feature Fusion

[0081] This step, "Risk Assessment and Dynamic Feature Fusion," serves as a crucial link in the entire system, building upon the state vector S constructed in step 1. t and operating condition label C t Output a real-time, continuous risk score R. t And generate a feature vector F that integrates context and working condition factors. t This is used by the subsequent maintenance strategy generation module. Considering the highly heterogeneous state distribution of electromechanical equipment under various operating conditions, directly using the current state for classification will lead to frequent false alarms. Therefore, we designed a lightweight model structure that combines a temporal attention mechanism, condition-adaptive affine transformation, and rare anomaly constraints, so that the risk score is not only based on the current state S. tFurthermore, it can integrate historical context and form a stronger ability to identify abnormal samples in special working conditions, thereby systematically improving the stability and practical usability of risk scoring.

[0082] enter

[0083] This step uses all the output from step 1:

[0084] Structured state vectors, including features such as vibration, current, and temperature;

[0085] The current operating status label of the equipment is uploaded by the PLC system at a fixed frequency;

[0086] The state cache for the past k=3 time steps is stored in the memory of the edge device.

[0087] First, to improve the model's ability to utilize historical information without increasing computational costs excessively, we designed a weighted fusion mechanism based on operating condition gating. By modeling the structural correlation between the current state and historical states, we achieve dynamic activation of historical state information.

[0088]

[0089] in:

[0090] This refers to the historical state after the integration;

[0091] Based on the current operating condition C t The affine transformation mapping function, with a set of linear parameters corresponding to each type of working condition.

[0092] An attention matching matrix relevant to the working conditions;

[0093] γ i The activation weight for each historical segment;

[0094] All parameters are obtained from model training, and the preloaded model structure is invoked based on the operating conditions.

[0095] The innovation of this structure lies in... By realizing condition-aware historical state remapping, historical states can participate in scoring and judgment from different "perspectives" under different operating conditions, thereby improving the system's ability to express complex dynamic operating conditions.

[0096] Next, the current state and the merged state are concatenated to form an intermediate representation. The input is fed into the risk scoring submodule. To address the weak representation of high-risk rare events in the sample distribution, we designed a sparse activation regularization term in the scoring function to prevent the model from overfitting to high-frequency, low-risk samples.

[0097]

[0098] in:

[0099] Weights for fully connected networks;

[0100] σ(·) is the Sigmoid function, and its output is R. t ∈(0,1) represents the risk score of the device;

[0101] The last term is the rare anomaly activation constraint term. When the overall amplitude of the state is small, this term amplifies the output, which is beneficial for the identification of weak anomalies.

[0102] λ is the regularization coefficient, ∈ to prevent division by zero, typically set to 1e. -6 .

[0103] This structure makes the system more sensitive to high-risk, low-frequency fault events, rather than being "masked" by stable operation under most conditions.

[0104] All models run at the edge, with each round of computation completed within 80ms, fully meeting the requirements for industrial deployment.

[0105] Output

[0106] R t Risk score results are continuous values, representing the probability of the equipment currently malfunctioning (for direct use by the subsequent strategy judgment module);

[0107] F t The intermediate features of the state fusion are input into the policy optimization model in the next step, providing a contextual basis for policy generation.

[0108] This step innovatively embeds structural information into the weighted fusion and affine projection of historical states, and combines it with a sparse activation mechanism to achieve a risk score output with high robustness, high sensitivity, and high generalization ability in scenarios with limited edge resources.

[0109] Step 3: Reinforcement Learning Strategy Generation

[0110] This step, "Reinforcement Learning Strategy Generation," is a crucial link in the patent system's transition from "state awareness and risk assessment" to "proactive decision-making and maintenance actions." Its goal is to build upon the risk score R output in the previous step. t and eigenvector F t Automatically generate maintenance policy action A for the current device.t Simultaneously construct the policy function π(F) t ,R t This information is intended for subsequent learning and iterative optimization. In practical IoT device management scenarios, the operating status of devices and the external environment are highly volatile, maintenance resources are usually limited, and accidental or delayed maintenance can lead to high costs. Therefore, this step proposes a reinforced policy network that combines a risk-segmented policy response mechanism with policy confidence entropy constraints to achieve adaptive hierarchical response and policy stability adjustment of maintenance policies, ensuring a dynamic balance between real-time performance, energy efficiency, and operational effectiveness.

[0111] First, considering the risk score R t The interval properties (i.e., R) t The larger the space, the more dangerous it is. This is mapped to five levels of risk response actions, namely:

[0112] a1: Continue running;

[0113] a2: Slightly reduced load operation;

[0114] a3: Schedule manual inspection;

[0115] a4: Automatically enters maintenance preparation state;

[0116] a5: Stop the machine immediately.

[0117] We designed a policy function π(F) based on reinforcement learning. t ,R t Its function is to fuse the vector F based on the current state. t and risk score R t The output is the probability distribution in the action space, and the current policy action A is obtained by sampling through this distribution. t This function is implemented using a two-branch network:

[0118] State branch: Input F t Action preferences are extracted by adding ReLU to two linear layers;

[0119] Risk branch: Input R t After passing through a Sigmoid modulation layer, the action preference weights are dynamically recalibrated for risk.

[0120] The final output motion distribution π(F) t ,R t Normalization is achieved through Softmax.

[0121] During the policy optimization phase, a confidence entropy suppression term is used to prevent the model from exhibiting overly dispersed policy outputs (high uncertainty) or overly aggressive behavior under low-risk conditions. Our designed policy loss is:

[0122]

[0123] in:

[0124] The first term is the expected maintenance cost loss, where C(A) t ) represents the cost constant corresponding to the selected action, for example, C(a1) = 0, C(a5) = 10;

[0125] The second term is the policy smoothing term, where KL(·) represents the KL divergence between the current policy distribution and the distribution at the previous time step, which encourages the policy to remain stable when the operating conditions are stable.

[0126] The third term is the entropy penalty term, which only applies to R. t Activated when the risk threshold θ is less than the risk level, this is used to suppress policy uncertainty in low-risk states and reduce policy ambiguity. (representing the entropy of the policy distribution);

[0127] λ1 and λ2 are regularization coefficients, and θ is the risk tolerance threshold (e.g., θ = 0.35);

[0128] The core innovation of the above loss design lies in the fact that it not only considers the direct loss of the current action, but also incorporates the stability of strategy changes and the operational controllability in low-risk scenarios, thereby strengthening the system's ability to make rational strategic choices in "non-emergency situations".

[0129] The architecture uses the Proximal Policy Optimization (PPO) algorithm for iterative optimization during the training phase and a distilled version during the deployment phase, keeping the model inference speed within 50ms and adapting to edge node environments.

[0130] Output

[0131] The optimal maintenance action selected at the current moment is determined by the policy distribution π(F). t ,R t Sampled from )

[0132] π(F t ,R t ): The current distribution of maintenance strategies, used for updates in the next step of execution and feedback training.

[0133] Step 4: Execution Control and Feedback Learning

[0134] This step, "Execution Control and Feedback Learning," is responsible for implementing the maintenance action A output by the reinforcement strategy module from the previous step. tIt is actually executed in the device system, and the subsequent device operating status is collected to construct a feedback score r for policy training and optimization. t .

[0135] First, execute control action A. t The actual method is as follows:

[0136] For actions a1 (continue running) or a2 (reduced load running), the edge node sends power or speed setting commands to the PLC via Modbus-TCP or OPC UA interface;

[0137] For a3 (manual notification), the maintenance scheduling module is invoked to generate a maintenance work order and send an SMS or App notification;

[0138] For a4 (maintenance preparation) or a5 (shutdown), an execution command is sent to the equipment safety power-off module via serial port, with a fan / oil pump shutdown command added if necessary.

[0139] After the action is executed, the system re-collects the current state vector S from the device after a fixed delay Δt (e.g., 60 seconds). t+Δt The data acquisition method is the same as in step 1 (edge ​​nodes extract sensor values ​​via MODBUS polling and assemble them into S). t+Δt vector).

[0140] We construct a feedback scoring function r t Evaluate the effectiveness of implementation:

[0141] r t =(||S t ||2-||S t+Δt ||2)-β·C(A t )

[0142] in:

[0143] ||S t ||2 and||S t+Δt ||2 represents the state modulus before and after the action, which represents the overall fluctuation range of the device. The smaller the value, the more stable it is.

[0144] C(A t ) represents the estimated maintenance cost of the current action (e.g., C(a1) = 0, C(a5) = 10);

[0145] β is the cost adjustment coefficient, which is empirically set to 0.3 to 0.5.

[0146] The score is then combined with the current action and state features to form a training sample:

[0147] τ t =(F t ,Rt A t ,r t )

[0148] The sample has a complete structure and clear temporal logic, and is a policy network π(F) t ,R t The training input provides the basis for feedback.

[0149] To enable the system to respond more quickly to high-risk situations and prevent sample redundancy, we introduce a feedback confidence weighting factor w. t Used for sampling during subsequent training:

[0150] w t =γ·R t +(1-γ)

[0151] in:

[0152] R t It is a risk score;

[0153] γ∈[0,1], it is recommended to set it to 0.7 to give high-risk samples a higher priority in training;

[0154] The implementation method is that the samples in the sample buffer are in the form of w t It is used as a sampling weight in the construction of policy training batches.

[0155] Finally, all samples τ t The data is stored in a circular buffer on the edge device (e.g., a FIFO buffer of length 1000). A policy update training will be triggered whenever the number of samples meets the policy update condition (e.g., 32 samples).

[0156] Output

[0157] r t The feedback score for this round of maintenance actions indicates the net benefit of this action to the system's state.

[0158] τ t Standard feedback sample units are used for periodic training of the reinforcement strategy network.

[0159] Step 5: Blockchain records shared with multiple parties

[0160] This step involves encapsulating the maintenance feedback results and policy execution sample information generated in the previous step into a trusted structure and writing them into the blockchain network. This ensures immutable recording of equipment maintenance activities, auditable processes, and traceable events. Furthermore, considering the multiple stakeholders in the equipment management system, including operators, manufacturers, and third-party maintenance providers, we designed a sharing mechanism based on access policies to ensure that different roles access data with authorization, supporting multi-party collaboration throughout the equipment's lifecycle.

[0161] First, regarding τ t Encapsulate and construct the data structure for on-chain processing. Its fields are as follows:

[0162] action_id:A t Action number;

[0163] risk_score:R t Risk assessment before strategy implementation;

[0164] reward:r t Feedback and rating;

[0165] policy_id: by π(F) t ,R t The version hash in this round is used to identify the policy version;

[0166] state_fingerprint: F is compressed by the following network ψ(·) t and R t After joint encoding, SHA256 hashing is performed to obtain the state digest.

[0167] The compression network ψ(·) is configured as follows:

[0168] The input dimension is 13 (12-dimensional F). t +1 dimension R t );

[0169] The first layer is a linear transformation, with an 8-dimensional output and ReLU activation function.

[0170] The second layer is a linear transformation that outputs a 4-dimensional value as a state representation summary.

[0171] Finally, perform the standard hash: SHA256(ψ(F) t ,R t This forms the state_fingerprint.

[0172] Then construct the feedback value weight ω of this data. t Used for on-chain priority determination:

[0173]

[0174] in:

[0175] The first item reflects the effectiveness of the strategy;

[0176] The second item reflects the on-chain priority of high-risk execution;

[0177] λ is a high-risk enhancement factor, with a recommended value of 1.0;

[0178] θ is the high-risk threshold, and a value of 0.6 is recommended.

[0179] The system employs a lightweight consensus mechanism (such as PBFT) for ordered writes, with each edge node submitting all pending writes every ΔT seconds. By sorting ω t Arrange in descending order, add the first K items to the block, and queue or discard the rest.

[0180] Finally, to achieve controllable sharing among multiple parties, we assign each record... Add an access permission tag κ t It consists of the following:

[0181]

[0182] This indicates that the data was encrypted by the node using the public key of partner j;

[0183] It is written into the smart contract control field access_policy, and can only be read if the authorized party's address matches.

[0184] This step will maintain the behavior result r t and complete feedback sample τ t Encapsulated as blockchain data units By leveraging a feedback-driven ranking consensus mechanism and compressed state digest technology, on-chain write efficiency is effectively improved. Furthermore, key policies enable multi-party controllable access, providing a fundamental guarantee for achieving transparent, trustworthy sharing, and accurate traceability of device lifecycle data.

[0185] The technical solutions described in the above embodiments of this application have at least the following technical effects or advantages: By performing data preprocessing and preliminary analysis on edge computing nodes, this solution can significantly reduce response latency in device management, ensuring that devices can be monitored and fault warnings provided in real time during actual operation. The introduction of edge computing not only solves the network bandwidth problem in existing systems but also enables dynamic optimization based on real-time device data and environmental changes, ensuring the efficiency and real-time performance of the device management system. In terms of device fault prediction, this invention employs an intelligent decision-making algorithm, integrating historical data, real-time data, and factors such as device type and operating environment to provide more accurate fault warnings and maintenance suggestions. Furthermore, by introducing decentralized blockchain technology, it ensures high transparency and credibility in data storage and sharing within the device management system.

[0186] This application also provides an embodiment of an IoT-based electromechanical equipment management method and system, such as... Figure 2 As shown, an IoT-based electromechanical equipment management system includes:

[0187] The data acquisition module is used to acquire various physical signals from the device and preprocess the various physical signals into a first state vector;

[0188] The risk assessment module generates a risk score and a feature vector using a first state vector and a working condition label. The risk score is obtained by remapping historical states through an affine transformation mapping function, and the feature vector is obtained by concatenating the current state and the fused state. The current state is obtained from the original data, and the fused state is obtained by dynamically activating historical states.

[0189] A strategy generation module is used to generate a first maintenance action and construct a strategy function to obtain the maintenance strategy distribution at the current time.

[0190] The execution feedback module performs a first maintenance action and collects a second state vector after execution. It constructs a feedback score and inputs the second state vector to obtain the feedback score. The feedback score, the first maintenance action, and the second state vector are used to form a training sample.

[0191] The information encapsulation and sharing module encapsulates feedback scores and training samples into blockchain data units, drives ranking consensus and state digests through feedback scores, and achieves shared access through a key policy.

[0192] The working principle is as follows: By performing data preprocessing and preliminary analysis on edge computing nodes, the response latency in device management is reduced; edge computing is introduced and dynamic optimization is performed based on real-time data of the device and environmental changes; by integrating historical data, real-time data of the device, as well as factors such as device type and operating environment, more accurate fault warnings and maintenance suggestions are provided; and decentralized blockchain technology is introduced to ensure information storage and sharing.

[0193] 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 method for managing electromechanical equipment based on the Internet of Things, characterized in that, include: Collect raw data, including vibration signals, current signals, equipment housing temperature, and operating condition labels; The first state vector is obtained by preprocessing the original data; The system inputs a first state vector and a working condition label, and outputs a risk score and a feature vector. The feature vector is obtained by concatenating the current state and the fused state. The current state is obtained from the original data, and the fused state is obtained by dynamically activating historical states. The risk score is calculated by inputting the feature vector into a scoring function. Specifically, the fused state is obtained by remapping historical states through an affine transformation mapping function based on the current working condition label. Each working condition corresponds to a set of linear parameters. The similarity between the current state and historical states is calculated through a working condition-related matrix, and weights are obtained through normalization. Finally, the weighted summation of different historical states is obtained. The scoring function is as follows: ； in, Weights for fully connected networks; For the Sigmoid function, the output is... This indicates the risk score of the device; the last item is the rare abnormal activation constraint. The regularization coefficient is... To prevent division by zero, set to ; A policy function is constructed, and a first maintenance action is generated based on risk score and feature vector. The first maintenance action is obtained by the distribution in the action space, which is mapped to five levels of risk response actions. The policy function realizes the probability distribution of the output in the action space through a dual-branch network, and introduces a policy smoothing term and an entropy penalty term. Execute the first maintenance action. After the execution is completed, collect the current state vector of the device as the second state vector. Construct a feedback scoring function and input the second state vector to obtain the feedback score. Use the feedback score, the first maintenance action and the second state vector to form a training sample. Feedback scores and training samples are encapsulated into blockchain data units, and shared access is achieved through a key policy.

2. The method for managing electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The preprocessing includes data windowing, data alignment, and feature extraction.

3. The method for managing electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The first state vector includes the average value of the vibration signal, the standard deviation of the vibration signal, the average value of the current signal, the rate of change of current within the window time, the average value of temperature, and the rate of change of temperature.

4. The method for managing electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The five-level risk response actions include: continue operation; slightly reduce load during operation; schedule manual inspection; automatically enter maintenance preparation state; and immediately shut down.

5. The method for managing electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The dual-branch network includes a state branch and a risk branch. The state branch extracts action preferences through two linear layers plus ReLU, and the risk branch performs risk dynamic recalibration on the action preference weights through a Sigmoid modulation layer.

6. The method for managing electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The policy smoothing term represents the KL divergence between the current policy distribution and the distribution at the previous time step, which is used to keep the policy stable under stable operating conditions.

7. The method for managing electromechanical equipment based on the Internet of Things according to claim 1, characterized in that, The training samples incorporate a feedback confidence weighting factor, which is calculated based on the risk score and used for sampling during subsequent training.

8. An electromechanical equipment management system based on the Internet of Things, characterized in that, include: The data acquisition module is used to acquire various physical signals from the device and preprocess the various physical signals into a first state vector; The risk assessment module is used to generate a risk score and a feature vector using a first state vector and a working condition label; The feature vector is obtained by concatenating the current state and the fused state; the current state is obtained from the original data, and the fused state is obtained by dynamically activating historical states; the risk score is calculated by inputting the feature vector into a scoring function; specifically, the fused state is obtained by remapping historical states through an affine transformation mapping function based on the current working condition label. Each working condition corresponds to a set of linear parameters. The similarity between the current state and historical states is calculated through a working condition-related matrix, and weights are obtained through normalization. Finally, the weighted summation of different historical states is obtained; the scoring function is as follows: ； in, Weights for fully connected networks; For the Sigmoid function, the output is... This indicates the risk score of the device; the last item is the rare abnormal activation constraint. The regularization coefficient is... To prevent division by zero, set to ; The strategy generation module is used to construct a strategy function, generate the first maintenance action based on the risk score and feature vector, and obtain the maintenance strategy distribution at the current time. The execution feedback module performs a first maintenance action and collects a second state vector after execution. It constructs a feedback score and inputs the second state vector to obtain the feedback score. The feedback score, the first maintenance action, and the second state vector are used to form a training sample. The information encapsulation and sharing module encapsulates feedback scores and training samples into blockchain data units and enables shared access through a key strategy.

Images