Electromechanical system health state sensing and early warning method and device based on block chain technology

Abstract

The invention relates to an electromechanical system health state sensing and early warning method based on a block chain technology. The method comprises the following steps: acquiring historical data monitored by an electromechanical system; classifying historical data, and storing the classified historical data in an HDFS, an HBase and an MySQL by adopting a block chain storage mechanism; preprocessing the data, and extracting a Mel spectrum characteristic coefficient; establishing a health state sensing model fusing the convolutional neural network and the Mel-spectrum feature coefficient, determining a specific sensing process, training the health state sensing model by using historical data, and determining a mapping relationship between physical quantity information and the health state; and acquiring real-time physical quantity information of the electromechanical system on line, taking the real-time physical quantity information as input of the trained health state sensing model so as to output a health state index of the electromechanical system in real time, and judging whether early warning is performed or not according to the health state index. The method has the beneficial effects that a closed-loop solution from data to early warning decision is provided for predictive maintenance of the electromechanical system, the operation and maintenance cost is remarkably reduced, and the equipment reliability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of health state assessment and management of electromechanical systems, and particularly relates to a method and device for health state perception and early warning of electromechanical systems based on blockchain technology. Background Art

[0002] For the health state detection of electromechanical systems, common methods include: traditional regular detection methods (visual inspection, non-destructive testing), data-driven intelligent evaluation technologies, predictive maintenance systems (such as PreMaint), index evaluation methods in patents, etc. The technical defects existing in such methods include: insufficient timeliness, low accuracy, dependence on manual experience, difficult data integration, poor model generalization ability, lagging maintenance strategies, etc. Specifically as follows: The traditional regular detection method has a problem of data delay, cannot capture sudden failures, and has weak generalization ability across devices; in the existing data-driven intelligent evaluation technologies, the existing algorithms have limited adaptability to complex working conditions, there are prediction deviations, and the feature alignment and correlation analysis of multiple data have not been fully clarified, which will affect the comprehensive evaluation effect; for the predictive maintenance system, in addition to the above problems, there is also a problem of low credibility: due to the black-box-based deep learning model, the credibility needs to be verified, and deploying sensors and edge computing devices requires high investment and high maintenance costs. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method and device for health state perception and early warning of electromechanical systems based on blockchain technology to overcome the above deficiencies in the prior art.

[0004] The technical solution of the present invention to solve the above technical problem is as follows:

[0005] A method for health state perception and early warning of electromechanical systems based on blockchain technology includes the following steps:

[0006] S10. Obtain the historical data monitored from the electromechanical system, which are respectively: historical physical quantity information and the corresponding health state information;

[0007] S20. Classify the historical data and store it in HDFS, HBase, and MySQL using the blockchain storage mechanism;

[0008] S30. Perform data preprocessing and extract Mel frequency spectrum feature coefficients;

[0009] S40. Establish a health state perception model integrating a convolutional neural network and Mel frequency spectrum feature coefficients, determine the specific perception process, then train the health state perception model with historical data, and determine the mapping relationship between physical quantity information and health state;

[0010] S50. Obtain the real-time physical quantity information of the electromechanical system online, and use it as the input of the trained health status perception model to output the health status index of the electromechanical system in real time, and determine whether to give an early warning based on the health status index.

[0011] Based on the above technical solutions, the present invention can be further improved as follows.

[0012] Further, the physical quantity information at least includes: voltage, current, power, rotational speed, acceleration, torque, temperature, and flow rate.

[0013] Further, the blockchain storage mechanism is specifically as follows:

[0014] First, encrypt and protect the data and generate a cryptographic hash digest to obtain a data format that meets the block requirements, and pack the obtained different data hash values into a block;

[0015] Then, use the data structure of the Merkle tree to store and connect the block header and the block body, and pack the root hash value obtained by iterating all transactions in the block body and store it in the block header;

[0016] Analyze the characteristics, advantages, and disadvantages of the consensus mechanism, and establish a consensus mechanism for nodes to reach an agreement on the distributed database during the dynamic transaction process of the blockchain;

[0017] Construct an automatically executable smart contract program to achieve data storage that is immutable, highly autonomous, and distributed consensus.

[0018] Further, the automatic execution process is as follows:

[0019] 1) Data encapsulation and uploading to the chain

[0020] After the data is encrypted by hybrid encryption, a CID identifier is generated. The smart contract automatically triggers the data sharding storage instruction, and distributes and stores the encrypted data to HDFS according to the PFS protocol. The metadata is packaged into a transaction block;

[0021] 2) Dynamic consensus verification

[0022] For high-frequency transactions, PoA is used to achieve second-level confirmation. For key data storage, BFT-PBFT is enabled to achieve cross-node consistency verification, and nodes verify the authenticity of the data hash through zk-SNARKs;

[0023] 3) Autonomous execution of smart contracts

[0024] Deploy a chain-triggered contract:

[0025] ① Storage verification contract: Automatically compare the Merkle Root- returned by the storage node with the record on the chain. If the difference rate > 0.1%, trigger the data repair instruction;

[0026] ② Permission Management Contract: Dynamically control data access based on attribute-based encryption. The access request triggers automatic key distribution and operation traceability.

[0027] ③ Abnormal Handling Contract: When 3 consecutive nodes report data anomalies, start multi-chain cross-verification and isolate the problem data segment.

[0028] 4) Self-maintaining Data Ecosystem

[0029] Build a hot and cold data hierarchical system. The smart contract automatically migrates low-frequency data to edge nodes according to the LU algorithm. The storage policy update takes effect through DPoS voting consensus, and the data operation logs are uploaded to the blockchain in real time.

[0030] Further, S20 is specifically as follows:

[0031] Classify historical data into structured data and unstructured data;

[0032] After the time-series data in the structured data is verified by the blockchain nodes, it is stored in MySQL by device ID + timestamp in a partitioned and sub-tabulated manner, and a composite index is established to optimize the query efficiency;

[0033] Store the relational feature data in the structured data through HBase columnar storage. Use RowKey to achieve second-level time range retrieval, and synchronize the data fingerprint to the blockchain for evidence storage;

[0034] First, encrypt the streaming data in the unstructured data in chunks, and then store it in HDFS through IPFS distributed storage, and generate a CID content identifier;

[0035] HBase stores the CID mapping table, records the data block location, hash value and access permission, and then realizes cross-node data traceability through the blockchain smart contract.

[0036] Further, the data preprocessing is specifically as follows:

[0037] Extract the stored data from HBase, MySQL, and HDFS, and verify the data hash value through the blockchain nodes to confirm the data integrity and timeliness, and filter out duplicate or tampered records;

[0038] Adopt a sliding time window mechanism to align the timestamps of the time-series data, eliminate noise through sliding mean filtering, dynamically identify and eliminate outliers based on the adaptive threshold filtering algorithm, and store it in the in-memory database after standardization;

[0039] Adopt wavelet denoising for preprocessing of the streaming data, extract the time-frequency domain feature vectors, and fuse them with the structured data at the feature level;

[0040] Extract statistical features, frequency-domain features, time-frequency domain features, and physical correlation features from historical data, construct a multi-dimensional feature pool based on these, and adopt a feature selection network driven by an attention mechanism to dynamically allocate feature weights through a multi-head self-attention mechanism.

[0041] Furthermore, extract Mel spectrogram feature coefficients:

[0042] Frame and window the vibration / acoustic signals of the electromechanical system, calculate the Mel spectrogram feature coefficients through Fourier transform and Mel filter bank mapping, and construct a time-frequency feature matrix as the input of the convolutional neural network.

[0043] Furthermore, the establishment process of the convolutional neural network is as follows:

[0044] 1) Design a multi-scale convolutional network architecture

[0045] The convolutional neural network adopts a two-channel parallel convolutional layer, which are 3×3 convolution and 5×5 convolution respectively, and applies a hierarchical learning rate strategy: the AdamW optimizer is used for the convolutional layer, the Nesterov momentum SGD is switched for the fully connected layer, and the training process is stabilized through gradient normalization. Stochastic Weight Averaging (SWA) is introduced to improve the generalization of the model, and the parameter sliding mean is synchronized every 5 epochs;

[0046] Insert DropBlock after the 3×3 convolution to randomly mask local regions of the feature map;

[0047] Implement MixUp cross-sample mixing, combined with the CutMix region replacement strategy, to generate mixed-label training data;

[0048] Introduce the SE attention module to dynamically enhance the feature weights of the fault-sensitive frequency bands;

[0049] After cross-layer feature splicing, connect to a dilated convolution to expand the temporal receptive field;

[0050] 2) Establish a deep supervision training mechanism

[0051] Define a mixed loss function: cross-entropy loss + KL divergence loss, and adopt gradient clipping;

[0052] Enhance dynamic data: online generate adversarial samples containing Gaussian noise and time-domain stretching;

[0053] 3) Online diagnosis and real-time optimization

[0054] Deploy the lightweight MobileNetV3 as an edge computing unit, and compress the benchmark model through knowledge distillation.

[0055] Furthermore, the health state index is defined as:

[0056]

[0057] Among them, ω i is the feature weight, x i，normal is the normal operating condition reference value, x i，threshold is the warning threshold, and HI is the health status index;

[0058] Dynamically update the warning threshold: Early warning line = MA - 2σ, Fault line = MA - 4σ,

[0059] MA is the moving average of the health status index calculated based on a sliding time window, and σ is the standard deviation of the health status index calculated based on a sliding time window;

[0060] When the HI value of N consecutive sampling points is lower than the early warning line, a yellow warning is triggered. When it is lower than the fault line, a red alarm is triggered, and the evidence is automatically stored on the chain through a smart contract.

[0061] Based on the above technical solution, the present invention also provides a health status perception and warning device for an electromechanical system based on blockchain technology, including:

[0062] A sensor module for obtaining physical quantity information of the electromechanical system;

[0063] A data collection module for collecting historical data, including physical quantity information and the corresponding health status information;

[0064] A data storage module for classifying historical data and storing it in HDFS, HBase, and MySQL using a blockchain storage mechanism;

[0065] A data processing module for preprocessing the historical data stored in the data storage module, extracting Mel spectrum feature coefficients, then training the health status perception model with the historical data, and real-time outputting the health status index of the electromechanical system through the trained health status perception model when inputting real-time physical quantity information;

[0066] A display module for displaying the real-time health status index of the electromechanical system output by the health status perception model.

[0067] The beneficial effects of the present invention are as follows: Through the collaboration of data classification, blockchain evidence storage, and multi-type storage systems, the comprehensive improvement of the health status data of the electromechanical system in terms of security, reliability, and scalability is achieved. At the same time, the storage cost and performance efficiency are taken into account, providing standardization for data governance in the industrial Internet of Things scenario. Combining data preprocessing and MFCCs feature extraction significantly improves the accuracy, real-time performance, and robustness of the health status monitoring of the electromechanical system. By fusing a convolutional neural network and Mel-spectrum feature coefficients to construct a health status perception model and clarifying the process and mapping relationship, a closed-loop solution from data to early warning decision-making can be provided for the predictive maintenance of the electromechanical system, significantly reducing the operation and maintenance cost and improving the equipment reliability. Description of the Drawings

[0068] Figure 1 It is a flowchart of the method for perceiving and warning the health status of an electromechanical system based on blockchain technology in the present invention;

[0069] Figure 2 It is a blockchain data storage method;

[0070] Figure 3 It is a structural diagram of the system for perceiving and warning the health status of an electromechanical system based on blockchain technology in the present invention. Detailed Embodiments

[0071] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0072] Embodiment 1

[0073] As Figure 1 shown, the method for perceiving and warning the health status of an electromechanical system based on blockchain technology includes the following steps:

[0074] S10. Obtain the historical data monitored from the electromechanical system, including: historical physical quantity information and the corresponding health status information;

[0075] S20. Classify the historical data and store it in HDFS, HBase, and MySQL using the blockchain storage mechanism;

[0076] S30. Perform data preprocessing and extract Mel-spectrum feature coefficients;

[0077] S40. Establish a health status perception model that fuses a convolutional neural network and Mel-spectrum feature coefficients, determine the specific perception process, then train the health status perception model with historical data, and determine the mapping relationship between the physical quantity information and the health status;

[0078] S50. Obtain the real-time physical quantity information of the electromechanical system online, and use it as the input of the trained health status perception model to output the health status index of the electromechanical system in real time, and determine whether to give an early warning based on the health status index.

[0079] Embodiment 2

[0080] This embodiment is a further optimization based on Embodiment 1, and its specific content is as follows:

[0081] The physical quantity information monitored by the electromechanical system at least includes: voltage, current, power, rotation speed, acceleration, torque, temperature, and flow rate. Considering the on-site installation, temperature and humidity, oil pollution, power supply and other environmental conditions, it is necessary to process them subsequently.

[0082] Embodiment 3

[0083] As Figure 2 shown, this embodiment is a further optimization based on Embodiment 1 or 2, and its specific content is as follows:

[0084] The blockchain storage mechanism is a blockchain storage technology based on cryptography, Merkle tree, peer nodes, consensus mechanism, and smart contract. Design the key information of the block header and block body, specifically:

[0085] First, encrypt and protect the data and generate a cryptographic hash digest to obtain a data format that meets the block requirements, and pack the obtained different data hash values into the block;

[0086] Then, use the data structure of the Merkle tree to store and connect the block header and block body, and pack the root hash value obtained by iterating all transactions in the block body into the block header;

[0087] Analyze the characteristics, advantages and disadvantages of the consensus mechanism, and establish a consensus mechanism for nodes to reach an agreement on the distributed database during the dynamic transaction process of the blockchain;

[0088] Build an automatically executable smart contract program to achieve data storage that is immutable, highly autonomous, and distributed consensus;

[0089] The consensus mechanism includes: proof of work, practical Byzantine fault tolerance algorithm, proof of stake, delegated proof of stake, proof of authority, and Kafka, etc.

[0090] The automatic execution process is as follows:

[0091] 1. Data encapsulation and on-chain

[0092] The data of the electromechanical system is encrypted by hybrid encryption to generate a CID identifier. The smart contract automatically triggers the data sharding storage instruction, and distributes and stores the encrypted data to HDFS according to the PFS protocol. The metadata is packaged into a transaction block. The hybrid encryption is: AES-256 encrypts the ontology data + ECC encrypts the key. The metadata is: CD, timestamp, device ID;

[0093] 2. Dynamic Consensus Verification

[0094] Adopt a hierarchical consensus mechanism: High-frequency transactions use PoA (Proof of Authority) to achieve second-level confirmation (<3 seconds). The key data storage enables BFT-PBFT (Practical Byzantine Fault Tolerance) to achieve cross-node consistency verification. Nodes verify the authenticity of data hashes through zk-SNARKs (Zero-Knowledge Proof), avoiding performance losses caused by full data broadcasting;

[0095] 3. Autonomous Execution of Smart Contracts

[0096] Deploy a chain-triggered contract:

[0097] ① Storage verification contract: Automatically compare the Merkle Root returned by the storage node with the record on the chain. If the difference rate > 0.1%, trigger the data repair instruction;

[0098] ② Permission management contract: Dynamically control data access based on Attribute-Based Encryption (ABE). The access request triggers automatic key distribution and operation traceability;

[0099] ③ Abnormal handling contract: When N consecutive nodes report data anomalies, start multi-chain cross-verification and isolate the problem data segment. Here, N can be other numbers greater than or equal to 2 such as 3, 4, 5, etc.

[0100] 4. Self-Maintaining Data Ecosystem

[0101] Build a hot and cold data hierarchical system: The smart contract automatically migrates low-frequency data to edge nodes according to the LU algorithm. The storage policy update takes effect through DPoS voting consensus. The data operation log is uploaded to the chain in real time, supporting 7-layer traceability auditing, and the system autonomy rate reaches more than 92%.

[0102] Example 4

[0103] This example is a further optimization based on any one of Examples 1 to 3, and its specific content is as follows:

[0104] S20 is specifically:

[0105] Classify historical data into structured data and unstructured data;

[0106] For the structured data processing flow:

[0107] After the time-series data such as voltage and current in the structured data are verified by the blockchain nodes, they are stored in a MySQL database by device ID + timestamp in separate tables, and a composite index is established to optimize the query efficiency;

[0108] The relational feature data (such as health status index, fault code) in the structured data are stored in a columnar format by HBase. The RowKey (device ID time range) is used to achieve second-level time range retrieval, and the data fingerprint is synchronized to the blockchain for evidence storage;

[0109] The processing flow for unstructured data is as follows:

[0110] The streaming data such as vibration waveforms and infrared thermograms in the unstructured data are first encrypted in chunks, then distributedly stored in HDFS by IPFS, and a CID content identifier is generated;

[0111] HBase stores the CID mapping table, records the data block location, hash value, and access permissions, and then realizes cross-node data traceability through the blockchain smart contract.

[0112] Data classification can optimize data management, facilitate subsequent storage, query, and analysis, and improve processing efficiency. Using blockchain storage can eliminate single-point trust, enhance data traceability and credibility, and the coordinated use of multi-type storage systems can balance storage costs and query efficiency.

[0113] Example 5

[0114] This example is a further optimization based on any one of Examples 1 to 4, and the specific content is as follows:

[0115] The data preprocessing is specifically as follows:

[0116] 1. Data extraction and verification

[0117] Extract the stored data from HBase, MySQL, and HDFS, and verify the data hash value through the blockchain nodes to confirm the data integrity and timeliness, and filter out duplicate or tampered records;

[0118] 2. Streaming and batch integrated preprocessing

[0119] Use the sliding time window mechanism to align the timestamps of time-series data such as voltage, current, and temperature. The time window can be preferably 1 to 5 seconds, and the overlap rate can be preferably 30%. Eliminate noise through sliding mean filtering, eliminate the sensor sampling frequency difference, dynamically identify and remove outliers based on the adaptive threshold filtering algorithm, such as noise data such as sudden current jumps and sudden temperature rises, and store the standardized data in the in-memory database;

[0120] 3. Signal enhancement and feature extraction

[0121] Wavelet denoising is used for preprocessing streaming data such as vibration waveforms, and time-frequency domain feature vectors are extracted and fused at the feature level with structured data. Physical quantity fusion: multi-physical quantity coupling analysis is performed on power (voltage × current) and mechanical parameters (torque × rotational speed), and an energy efficiency index (η = output power / input power × 100%) is constructed. Time-frequency domain analysis: wavelet packet transform (WPT) is performed on the vibration signal (acceleration) and decomposed to 6 layers, and the energy entropy features of 16 sub-bands are extracted.

[0122] 4. Feature Engineering and Dimensionality Reduction

[0123] (1) Construct a multi-dimensional feature pool

[0124] Four types of features are extracted from historical data:

[0125] ① Statistical features (mean, variance, kurtosis);

[0126] ② Frequency domain features (FFT main frequency amplitude, harmonic component ratio);

[0127] ③ Time-frequency domain features (wavelet packet energy entropy, EMD marginal spectrum entropy);

[0128] ④ Physical correlation features (power-rotational speed curve slope, torque-flow correlation coefficient);

[0129] (2) Feature Selection and Fusion

[0130] An attention mechanism-driven feature selection network (AFSN) is used to dynamically allocate feature weights through a multi-head self-attention module (the multi-head self-attention mechanism is the core component of the Transformer architecture). For example:

[0131] Under high-temperature conditions, the covariance weight of temperature and current is increased to 0.8;

[0132] In high-vibration scenarios, the frequency domain energy entropy weight of the acceleration signal is dominant (weight 0.7).

[0133] There are many interferences in the physical quantity information signals of the original electromechanical system. Data preprocessing can remove irrelevant noise, enhance key information of the signals, and standardize / normalize the data format for subsequent feature extraction and time series analysis. The use of Mel spectrum feature analysis is because it contains richer frequency domain information, is sensitive to early weak faults, and at the same time, has higher computational efficiency and is more suitable for automated pipeline processing.

[0134] Example 6

[0135] This example is a further optimization based on any one of Examples 1 to 5, and its specific content is as follows:

[0136] Considering that the acquisition of device status data will be affected by external environments such as ambient temperature, humidity, vibration shock, and electromagnetic waves during the acquisition process, there will be a deviation between the acquired data and the real signal. In order to improve the signal-to-noise ratio, eliminate outliers in the original data, improve the quality of the data, and extract Mel Frequency Cepstral Coefficients (MFCC), specifically, the extraction of Mel Frequency Cepstral Coefficients is as follows: Frame and window the vibration / acoustic signals of the electromechanical system, such as a frame length of 25 ms and an overlap of 10 ms. After Fourier transform and Mel filter bank mapping, such as 40 triangular filters, calculate the Mel Frequency Cepstral Coefficients, such as taking the first 20 dimensions, and construct a time-frequency feature matrix as the input of the convolutional neural network.

[0137] The dimension of the original time-domain signal is high and redundant. The Mel Frequency Cepstral Coefficient method performs frame processing on the signal through short-time Fourier transform, focuses on the key frequency bands, and extracts the feature vectors representing the signal changes, which can greatly reduce the computational amount. Example 7

[0138] This embodiment is a further optimization based on any one of Embodiments 1 to 6, and its specific content is as follows:

[0139] The establishment process of the convolutional neural network is as follows:

[0140] 1) Design a multi-scale convolutional network architecture

[0141] The convolutional neural network adopts a dual-channel parallel convolutional layer, which are 3×3 convolution and 5×5 convolution respectively. The 3×3 small kernel captures local frequency domain features, and the 5×5 large kernel extracts the broadband energy distribution. The SE (Squeeze-Excitation) attention module is introduced to dynamically enhance the feature weights of the fault-sensitive frequency bands; after cross-layer feature splicing, dilated convolution is connected (such as: dilation = 2) to expand the temporal perception field; a hierarchical learning rate strategy is adopted: the convolutional layer uses the AdamW optimizer (such as: initial LR = 1e -4 , weight decay 0.01), and the fully connected layer switches to Nesterov momentum SGD (such as: LR = 1e -3 ), and the training process is stabilized through gradient normalization (such as: Norm = 1.0). The SWA (Stochastic Weight Averaging) is introduced to improve the generalization of the model, and the sliding mean of the parameters is synchronized every 5 epochs;

[0142] Structured regularization: Insert DropBlock (such as: block size = 7, keep prob = 0.8) after the 3×3 convolution to randomly mask the local area of the feature map and enhance the spatial robustness;

[0143] Data-level constraints: Implement MixUpi cross-sample mixing (e.g., = 0.4), combined with the CutMix region replacement strategy, to generate mixed-label training data to suppress overfitting;

[0144] 2) Establish a deep supervision training mechanism

[0145] Define a mixed loss function: cross-entropy loss (health status classification) + KL divergence loss (MFCC feature distribution alignment), and prevent model oscillation through gradient clipping (e.g., threshold 1.0);

[0146] Enhance dynamic data augmentation: Online generate adversarial samples with Gaussian noise (e.g., SNR ≥ 15dB) and time-domain stretching (e.g., ±10%) to improve the generalization ability of the model;

[0147] 3) Online diagnosis and real-time optimization Deploy the lightweight MobileNetV3 as an edge computing unit, and compress the benchmark model through knowledge distillation, such as compressing 80%, to achieve real-time inference (<50ms / sample). Experiments show that this method achieves 98.6% accuracy on the bearing fault dataset, a 23.5% improvement compared to the traditional SVM method.

[0148] Achieve high-precision and adaptive device health status assessment through multi-modal feature fusion and deep learning. The specific effects are as follows:

[0149] Achieve accurate diagnosis: Combine the advantages of signal processing and deep learning to improve the fault classification accuracy;

[0150] Efficient deployment: The lightweight design supports edge computing and meets the industrial real-time requirements;

[0151] Interpretability: Physical features and data-driven models complement each other to enhance the credibility of the results;

[0152] Generalization ability: Training with historical data enables the model to adapt to multi-device and multi-condition scenarios.

[0153] Deploy automatic mixed precision training (AMP), and achieve FP16 / FP32 mixed computing through Tensor Core. Combine gradient accumulation (update every 4 batches) to reduce the video memory occupancy, and the training speed of ResNet-50 is increased by 2.3 times. Experiments show that this algorithm achieves 78.9% top-1 accuracy on the ImageNet dataset, and the convergence speed is increased by 40% compared to the baseline training scheme.

[0154] The health status index is defined as:

[0155]

[0156] where, ω iis the feature weight (optimized by the random forest algorithm), x i，normal is the normal operating condition reference value, x i，threshold is the warning threshold, HI is the health status index;

[0157] Dynamically update the warning threshold: warning line = MA - 2σ, fault line = MA - 4σ,

[0158] MA is the moving average of the health status index calculated based on the sliding time window, and σ is the standard deviation of the health status index calculated based on the sliding time window;

[0159] When the HI values of N consecutive sampling points are lower than the warning line, a yellow warning is triggered, and when lower than the fault line, a red alarm is triggered, and the evidence is automatically stored on the chain through a smart contract. Here, N can be other numbers greater than or equal to 2 such as 3, 4, 5, etc.

[0160] Dynamic decision-making and on-chain interaction

[0161] Trigger the smart contract according to the health status index (HI value):

[0162] Healthy state (H≥0.9): Update the blockchain device status ledger;

[0163] Abnormal warning (Hl<0.7): Automatically generate an alarm event and associate the CID of the original data in the storage layer for traceability.

[0164] The processing result is pushed to the application layer through the Kafka message queue, and the end-to-end delay ≤ 800ms.

[0165] List the specific content of the mapping relationship as an example:

[0166]

[0167] Example 7

[0168] As Figure 3 shown, the electromechanical system health status perception and warning device based on blockchain technology includes:

[0169] A sensor module for obtaining physical quantity information of the electromechanical system;

[0170] A data collection module for collecting historical data, including physical quantity information and the corresponding health status information;

[0171] A data storage module for classifying historical data and storing it in HDFS, HBase, and MySQL using the blockchain storage mechanism;

[0172] A data processing module is used to preprocess the historical data stored in the data storage module, extract Mel spectrum feature coefficients, then train the health status perception model with the historical data, and when real-time physical quantity information is input, the health status index of the electromechanical system is output in real time through the trained health status perception model;

[0173] A display module is used to display the real-time health status index of the electromechanical system output by the health status perception model.

[0174] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for perceiving and warning the health status of an electromechanical system based on blockchain technology, characterized in that It includes the following steps: S10. Obtain the historical data monitored by the electromechanical system, which are respectively: historical physical quantity information and the corresponding health status information; S20. Classify the historical data and store it in HDFS, HBase and MySQL using the blockchain storage mechanism; S30. Perform data preprocessing and extract Mel spectrogram feature coefficients; S40. Establish a health status perception model integrating a convolutional neural network and Mel spectrogram feature coefficients, establish a specific perception process, determine the mapping relationship between physical quantity information and health status, and then train the health status perception model with historical data; S50. Online obtain the real-time physical quantity information of the electromechanical system and use it as the input of the trained health status perception model to output the health status index of the electromechanical system in real time, and judge whether to give an early warning based on the health status index.

2. The method for perceiving and warning the health status of the electromechanical system based on blockchain technology according to claim 1, characterized in that: The physical quantity information at least includes: voltage, current, power, rotation speed, acceleration, torque, temperature and flow rate.

3. The method for perceiving and warning the health status of an electromechanical system based on blockchain technology according to claim 2, wherein, The blockchain storage mechanism is specifically as follows: First, encrypt and protect the data and generate a cryptographic hash digest to obtain a data format that meets the block requirements, and pack the obtained different data hash values into a block; Then use the Merkle tree data structure to store and connect the block header and block body, and pack the root hash value obtained by iterating all transactions in the block body and store it in the block header; Analyze the characteristics, advantages and disadvantages of the consensus mechanism, and establish a consensus mechanism for nodes to reach an agreement on the distributed database during the dynamic transaction process of the blockchain; Build an automatically executed smart contract program to achieve tamper-proof, highly autonomous and distributed consensus data storage.

4. The method for perceiving and warning the health state of the electromechanical system based on the blockchain technology according to claim 3, characterized in that, The automatic execution process is as follows: 1) Data encapsulation and on-chain After the data is encrypted by hybrid encryption, a CID identifier is generated, and the smart contract automatically triggers the data sharding storage instruction, and distributes and stores the encrypted data to HDFS according to the PFS protocol, and the metadata is packaged into a transaction block; 2) Dynamic consensus verification For high-frequency transactions, PoA is used to achieve second-level confirmation. For key data storage, BFT-PBFT is enabled to achieve cross-node consistency verification, and nodes verify the authenticity of the data hash through zk-SNARKs; 3) Autonomous execution of smart contracts Deploy chain-triggered contracts: ① Storage verification contract: Automatically compare the Merkle Root returned by the storage node with the record on the chain. If the difference rate > 0.1%, trigger the data repair instruction; ② Permission management contract: Dynamically control data access based on attribute-based encryption, and the access request triggers automatic key distribution and operation trace recording; ③ Abnormal handling contract: When 3 consecutive nodes report data anomalies, start multi-chain cross-verification and isolate the problem data segment; 4) Self-maintaining data ecosystem Build a hot and cold data hierarchical system, and the smart contract automatically migrates low-frequency data to edge nodes according to the LU algorithm. The update of the storage strategy takes effect through DPoS voting consensus, and the data operation log is uploaded to the chain in real time.

5. The method for perceiving and warning the health status of the electromechanical system based on blockchain technology according to claim 3, characterized in that, S20 is specifically as follows: Classify the historical data into structured data and unstructured data; After the time-series data in the structured data is verified by the blockchain node, it is stored in MySQL by device ID + timestamp in a sub-library and sub-table manner, and a composite index is established to optimize the query efficiency; Store the relational feature data in structured data through HBase columnar storage, use RowKey to achieve second-level time range retrieval, and synchronize the data fingerprint to the blockchain for evidence storage; First, block-encrypt the streaming data in unstructured data, and then store it in HDFS through IPFS distributed storage, and generate a CID content identifier; HBase stores the CID mapping table, records the data block location, hash value and access rights, and then realizes cross-node data traceability through the blockchain smart contract.

6. The method for perceiving and warning the health status of an electromechanical system based on blockchain technology according to any one of claims 1 to 5, characterized in that, The data preprocessing is specifically as follows: Extract the stored data from HBase, MySQL, and HDFS, and verify the data hash value through the blockchain node to confirm the data integrity and timeliness, and filter out duplicate or tampered records; Adopt a sliding time window mechanism to align the timestamps of time series data, eliminate noise through sliding mean filtering, dynamically identify and remove outliers based on the adaptive threshold filtering algorithm, and store them in the in-memory database after standardization; Adopt wavelet denoising to preprocess the streaming data, extract the time-frequency domain feature vectors, and fuse them with the structured data at the feature level; Extract statistical features, frequency domain features, time-frequency domain features, and physical association features from historical data, and construct a multi-dimensional feature pool based on this. Adopt a feature selection network driven by the attention mechanism, and dynamically allocate feature weights through the multi-head self-attention mechanism.

7. The method for perceiving and warning the health status of an electromechanical system based on blockchain technology according to claim 5, wherein Extract Mel spectrum feature coefficients: Frame and window the vibration / acoustic signals of the electromechanical system, calculate the Mel spectrum feature coefficients through Fourier transform and Mel filter bank mapping, and construct a time-frequency feature matrix as the input of the convolutional neural network.

8. The method for perceiving and warning the health status of an electromechanical system based on blockchain technology according to any one of claims 1 to 7, characterized in that, The establishment process of the convolutional neural network is as follows: 1) Design a multi-scale convolutional network architecture The convolutional neural network adopts a two-channel parallel convolutional layer, which are 3×3 convolution and 5×5 convolution respectively, and applies a hierarchical learning rate strategy: the AdamW optimizer is used for the convolutional layer, and the Nesterov momentum SGD is switched for the fully connected layer. And stabilize the training process through gradient normalization, introduce SWA to improve the model generalization ability, and synchronize the parameter sliding mean every 5 epochs; Insert DropBlock after the 3×3 convolution to randomly mask the local area of the feature map; Implement MixUpi cross-sample mixing, cooperate with the CutMix region replacement strategy to generate mixed label training data; Introduce the SE attention module to dynamically enhance the feature weights of the fault-sensitive frequency bands; After cross-layer feature splicing, connect to the dilated convolution to expand the time series perception field; 2) Establish a deep supervision training mechanism Define a mixed loss function: cross-entropy loss + KL divergence loss, and adopt gradient clipping; Enhance dynamic data: online generate adversarial samples containing Gaussian noise and time domain stretching; 3) Online diagnosis and real-time optimization Deploy the lightweight MobileNetV3 as the edge computing unit, and compress the benchmark model through knowledge distillation.

9. The method for perceiving and warning the health state of the electromechanical system based on the blockchain technology according to any one of claims 1 to 8, characterized in that The health status index is defined as: Among them, ω i is the feature weight, x i，normal is the normal operating condition reference value, x i，threshold is the warning threshold, and HI is the health status index; Dynamically update the alarm threshold: warning line = MA - 2σ, fault line = MA - 4σ, MA is the moving average of the health status index calculated based on the sliding time window, and σ is the standard deviation of the health status index calculated based on the sliding time window; When the HI values of consecutive N sampling points are lower than the warning line, a yellow warning is triggered. When they are lower than the fault line, a red alarm is triggered, and the evidence is automatically stored on the chain through a smart contract.

10. An electromechanical system health status perception and early warning device based on blockchain technology, characterized in that, It includes: A sensor module for obtaining physical quantity information of the electromechanical system; A data collection module for collecting historical data, including physical quantity information and corresponding health status information; A data storage module for classifying historical data and storing it in HDFS, HBase, and MySQL using a blockchain storage mechanism; A data processing module for preprocessing the historical data stored in the data storage module, extracting Mel spectrum feature coefficients, then training a health status perception model with the historical data, and when real-time physical quantity information is input, the health status index of the electromechanical system is output in real time through the trained health status perception model; A display module for displaying the real-time health status index of the electromechanical system output by the health status perception model.

Images