Predictive maintenance method for intelligent factory Internet of Things equipment

Abstract

The invention relates to the technical field of industrial Internet of Things, in particular to a predictive maintenance method for intelligent factory Internet of Things equipment, which comprises the following steps of: acquiring equipment operation parameters, environment monitoring data and historical maintenance records, constructing a multi-dimensional feature data set, extracting equipment degradation features by adopting a topological graph attention mechanism and a Bayesian network, and establishing a multi-dimensional feature data set; the method realizes equipment health state modeling and fault probability prediction, combines a dynamic adjacency matrix and a multi-objective optimization algorithm, comprehensively optimizes maintenance cost, equipment fault risk and associated equipment influence, dynamically generates an optimal maintenance plan, carries out constraint optimization based on a mixed integer programming method, automatically generates a maintenance instruction sequence, and achieves the optimal maintenance of the equipment. Tasks are issued through the computerized maintenance management system, the PLC control system and the industrial Internet of Things gateway, and the execution state is monitored and maintained in real time. The intelligent level of equipment maintenance is effectively improved, non-planned shutdown is reduced, and the equipment reliability and the production efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet of Things, and particularly to a predictive maintenance method for Internet of Things devices in an intelligent factory. Background Art

[0002] With the development of industrial automation and intelligent manufacturing, Internet of Things technology has been widely applied in the equipment management of intelligent factories. Key equipment in industrial production, such as machine tools, transmission systems, and power devices, directly affects production efficiency and product quality in terms of their operating status.

[0003] In recent years, predictive maintenance has gradually become the mainstream development direction. It relies on multi-source data collection, health status assessment, and intelligent optimization scheduling. Through data analysis and machine learning models, it predicts the degradation trend of equipment, realizes precise maintenance, thereby reducing the failure rate, optimizing the allocation of maintenance resources, and improving the overall operation efficiency of the factory. However, the existing predictive maintenance methods have the following main technical problems: First, in terms of equipment status perception, traditional methods usually rely on single-sensor data, lacking the global integration of operating parameters, environmental factors, and maintenance records, resulting in unstable health assessment results and difficulty in accurately identifying the fault propagation path. Second, in terms of optimizing maintenance decisions, current methods lack a comprehensive consideration of the process coupling relationship between equipment and maintenance resource constraints. Maintenance plans are often triggered based on fixed rules or simple thresholds, leading to untimely maintenance of key equipment or unreasonable scheduling of maintenance resources. Summary of the Invention

[0004] Based on the above purpose, the present invention provides a predictive maintenance method for Internet of Things devices in an intelligent factory.

[0005] A predictive maintenance method for Internet of Things devices in an intelligent factory includes the following steps:

[0006] S1: Collect the operating parameter data, environmental monitoring data, and historical maintenance records of the target device to generate a multi-source heterogeneous data feature set;

[0007] S2: Perform implicit feature extraction on the multi-source heterogeneous data feature set to generate a set of device degradation feature vectors. The implicit feature extraction includes feature space mapping based on the device association topology structure;

[0008] S3: According to the set of device degradation feature vectors, combined with the historical fault data within the dynamic time window, generate a device fault probability prediction matrix;

[0009] S4: Based on the device fault probability prediction matrix and device maintenance cost parameters, construct a dynamic maintenance strategy decision model;

[0010] S5: Solve the dynamic maintenance policy decision model through a multi-objective optimization algorithm, generate an optimal maintenance instruction sequence, and send the optimal maintenance instruction sequence to the maintenance execution unit of the target device.

[0011] Optionally, the S1 includes:

[0012] S11: Collect operation parameter data through a sensor array deployed on the device body, where the operation parameter data includes vibration spectrum data, temperature gradient data, and current harmonic feature data;

[0013] S12: Preprocess the operation parameter data through an edge computing node to generate a standardized data stream;

[0014] S13: Collect environmental parameter data of the area where the device is located through an environmental monitoring terminal, where the environmental parameter data includes air particulate matter concentration, humidity fluctuation curve, and electromagnetic interference intensity, and align the environmental parameter data with the operation parameter data in terms of time stamps to generate a spatio-temporal correlation data block;

[0015] S14: Analyze the maintenance work order text in the device historical maintenance record, and extract maintenance feature vectors through natural language processing technology. The extraction includes:

[0016] Use the TF-IDF algorithm to vectorize the keywords in the maintenance work order text;

[0017] Based on the BiLSTM model, identify the mapping relationship between maintenance operation types and device components;

[0018] Construct a maintenance event-device status association matrix;

[0019] S15: Perform feature-level fusion on the standardized data stream, spatio-temporal correlation data block, and maintenance feature vectors to generate a multi-source heterogeneous data feature set. The fusion includes establishing a feature association weight distribution table based on the device topology relationship, performing tensor splicing on heterogeneous data sources, and then reducing the dimension through a feature selection algorithm, and outputting a three-dimensional feature tensor including the time dimension, device dimension, and feature dimension.

[0020] Optionally, the preprocessing of the operation parameter data in S12 includes:

[0021] S121: Perform wavelet packet decomposition and noise reduction on the vibration spectrum data, and extract the energy entropy feature of the specified frequency band;

[0022] S122: Use the dynamic baseline calibration algorithm to eliminate environmental interference on the temperature gradient data;

[0023] S123: Perform Fourier transform on the current harmonic feature data and match it with a preset harmonic feature template.

[0024] Optionally, S2 includes:

[0025] S21: Analyze the three-dimensional feature tensor in the multi-source heterogeneous data feature set, construct a dynamic adjacency matrix based on the device topological relationship, and the node weights of the dynamic adjacency matrix are jointly calculated by the device physical connection tightness and the process coupling degree;

[0026] S22: Input the three-dimensional feature tensor and the dynamic adjacency matrix into a preset deep autoencoder, and perform cross-device feature propagation through the topological graph attention mechanism to generate a latent feature vector. The topological graph attention mechanism includes:

[0027] Calculate the feature similarity of adjacent device nodes as the attention weight;

[0028] Establish a cross-device feature propagation channel in the encoder hidden layer;

[0029] S23: Perform degradation trend enhancement processing on the latent feature vector, including:

[0030] Adopt residual connection to stack the device historical degradation features;

[0031] Calculate the cumulative offset of the feature vector through a time sliding window;

[0032] S24: Fuse the latent feature vector after degradation trend enhancement processing with the dynamic adjacency matrix to generate a set of device degradation feature vectors, and each vector in the set contains the device body degradation index and the associated device influence coefficient.

[0033] Optionally, S24 includes:

[0034] S241: Extract the weight parameters of the device nodes from the dynamic adjacency matrix to generate an associated influence coefficient matrix, where each coefficient represents the intensity value of the influence of the associated device on the current device degradation;

[0035] S242: Perform feature dimension expansion on the latent feature vector after degradation trend enhancement processing, add the device association topology dimension, and form a four-dimensional feature tensor.

[0036] S243: Perform Hadamard product operation on the associated influence coefficient matrix and the four-dimensional feature tensor to realize the quantization injection of topological association influence and generate a weighted feature tensor;

[0037] S244: Perform cross-dimensional feature compression on the weighted feature tensor.

[0038] S245: Perform temporal modeling on the compressed feature tensor through a gated recurrent unit (GRU) and output a set of device degradation feature vectors.

[0039] Optionally, S3 includes:

[0040] S31: Extract the fault event feature vectors associated with the current device from the historical fault database;

[0041] S32: Based on the parameters of the time sliding window and the window length adjustment rule, assign time-dependent weights to the historical fault data to generate a weighted fault feature sequence. The assignment rule includes:

[0042] Apply an exponential decay weight to the fault data within the time window according to the proximity of the occurrence time;

[0043] Apply a path attenuation coefficient to the cross-device propagation faults;

[0044] S33: Construct a Bayesian network prediction model and perform multi-modal input of the device degradation feature vector set and the weighted fault feature sequence;

[0045] S34: Dynamically update the conditional probability table of the Bayesian network through a sliding window mechanism;

[0046] S35: Perform Bayesian network inference and output the device fault probability prediction matrix.

[0047] Optionally, S4 includes:

[0048] S41: Integrate the device fault probability prediction matrix, the device maintenance cost parameters (the parsing results of the maintenance work order text in the historical maintenance records from S1), and the associated device impact coefficients to generate a structured decision input set.

[0049] S42: Using the generated structured decision input set as input data, construct a multi-objective optimization function, including minimizing the total maintenance cost, maximizing the suppression of fault risks, and maximizing the reduction of associated impacts;

[0050] S43: Define dynamic constraint conditions based on the structured decision input set. The constraint conditions include time window constraints, resource conflict constraints, and device mutual exclusion constraints;

[0051] S44: Dynamically adjust the model parameters according to the device degradation indicators in the real-time standardized data stream and the cumulative offset;

[0052] S45: Based on the multi-objective optimization function, dynamic constraint conditions, and real-time adjusted parameters, construct a dynamic maintenance strategy decision model, perform optimization and solution through a multi-objective optimization algorithm, and output a set of candidate strategies.

[0053] Optionally, S5 includes:

[0054] S51: Application of Genetic Algorithm: In the dynamic maintenance strategy decision model, multi-objective optimization involves minimizing maintenance costs, maximizing suppression of failure risks, and maximizing reduction of associated impacts, and is solved through genetic algorithm optimization.

[0055] S52: Determine the Execution Order of Maintenance Tasks: The optimal individual obtained by solving through genetic algorithm represents the set of devices to be maintained. To ensure the efficient execution of maintenance tasks, the execution order of maintenance tasks is optimized.

[0056] Optionally, the S5 further includes:

[0057] S53: Generate the Optimal Maintenance Instruction Sequence: Based on the execution order of maintenance tasks, generate a maintenance instruction sequence and construct a maintenance task instruction set;

[0058] S54: Issue Maintenance Instructions: The maintenance instructions are issued to the maintenance execution unit and monitored in real time to ensure that the tasks are executed according to the plan.

[0059] Advantages of the Present Invention:

[0060] In the present invention, by collecting operation parameters, environmental monitoring data, and historical maintenance records, a multi-source heterogeneous data feature set is constructed, and a topological graph attention mechanism and a Bayesian network are used for equipment degradation feature extraction and failure probability prediction. Based on the equipment topological association relationship, a dynamic adjacency matrix is constructed to identify the cross-device failure propagation path, and the time effect weight is calculated in combination with historical failure data to improve the accuracy of failure prediction. Compared with traditional maintenance methods that rely on single-sensor data, the present invention can construct a multi-dimensional, dynamic, and updatable equipment health assessment model, so as to realize more accurate equipment degradation trend analysis and intelligent maintenance decision support.

[0061] In the present invention, genetic algorithm is used for multi-objective optimization, comprehensively considering maintenance costs, equipment failure risks, and associated influence degrees, to ensure that the maintenance plan has optimal adaptability in different production environments. The maintenance task sorting is calculated through the weighted priority scheduling algorithm to ensure that key equipment is maintained first, avoid resource conflicts, and combine with the dynamic time window adjustment mechanism to intelligently compress and optimize the maintenance time window based on equipment degradation trends, production plans, and resource availability. Description of the Drawings

[0062] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative efforts.

[0063] Figure 1Schematic flowchart of the method according to an embodiment of the present invention;

[0064] Figure 2 Schematic flowchart of the S1 process according to an embodiment of the present invention. Detailed implementation manners

[0065] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the accompanying drawings are only for more specifically describing the embodiments and are not intended to specifically limit the present invention.

[0066] It should be noted that in the specification, references to "one embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc. indicate that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Additionally, when combining an embodiment to describe a specific feature, structure, or characteristic, implementing such a feature, structure, or characteristic in combination with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the relevant art.

[0067] Generally, terms can be understood, at least in part, from their use in the context. For example, at least in part depending on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or can be used to describe a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey a set of exclusive factors, but rather, at least in part depending on the context, can allow for the existence of other factors that may not be explicitly described.

[0068] As Figure 1 - Figure 2 shown, a predictive maintenance method for an Internet of Things device in an intelligent factory includes the following steps:

[0069] S1: Collect the operating parameter data, environmental monitoring data, and historical maintenance records of the target device to generate a multi-source heterogeneous data feature set;

[0070] S2: Perform implicit feature extraction on the multi-source heterogeneous data feature set to generate a set of device degradation feature vectors. The implicit feature extraction includes feature space mapping based on the device association topology structure;

[0071] S3: According to the set of device degradation feature vectors and in combination with the historical fault data within a dynamic time window, generate a device fault probability prediction matrix;

[0072] S4: Based on the device fault probability prediction matrix and the device maintenance cost parameters, construct a dynamic maintenance strategy decision model;

[0073] S5: Solve the dynamically maintained policy decision model through a multi-objective optimization algorithm, generate an optimal maintenance instruction sequence, and send the optimal maintenance instruction sequence to the maintenance execution unit of the target device.

[0074] S1 includes:

[0075] S11: Collect operation parameter data through a sensor array deployed on the device body. The operation parameter data includes vibration spectrum data, temperature gradient data, and current harmonic characteristic data;

[0076] The sensor array includes an accelerometer, a thermocouple, and a Hall current sensor, which are respectively used to collect vibration spectrum data, temperature gradient data, and current harmonic characteristic data;

[0077] S12: Preprocess the operation parameter data through an edge computing node to generate a standardized data stream. The preprocessing includes:

[0078] Perform wavelet packet decomposition and noise reduction on the vibration spectrum data, and extract the energy entropy characteristics of the specified frequency band;

[0079] Use a dynamic baseline calibration algorithm for the temperature gradient data to eliminate environmental interference;

[0080] Perform Fourier transform on the current harmonic characteristic data and match it with a preset harmonic characteristic template;

[0081] S13: Collect environmental parameter data of the area where the device is located through an environmental monitoring terminal. The environmental parameter data includes air particulate matter concentration, humidity fluctuation curve, and electromagnetic interference intensity, and align the environmental parameter data with the operation parameter data by timestamp to generate a spatio-temporal correlation data block;

[0082] The environmental monitoring terminal includes a laser particle sensor, a humidity sensor, and a spectrum analyzer, which are respectively used to collect air particulate matter concentration, humidity fluctuation curve, and electromagnetic interference intensity;

[0083] S14: Analyze the maintenance work order text in the device historical maintenance record, and extract maintenance feature vectors through natural language processing technology. The extraction includes:

[0084] Perform keyword vectorization on the maintenance work order text using the TF-IDF algorithm;

[0085] Based on the BiLSTM model, identify the mapping relationship between maintenance operation types and device components;

[0086] Construct a maintenance event-device status association matrix;

[0087] S14: Analyze the unstructured text data in the device historical maintenance record. The steps are as follows:

[0088] Set the maintenance work order text set: D = {d 1 , d 2 ,..., d N};

[0089] (1) Vectorize the keywords of the maintenance work order text: Use TF-IDF to calculate the text keyword vector, expressed as: Among them, TF(d i , t) is the frequency of the word t in the document d i , DF(t) is the number of documents containing the word t, W(d i , t) is the TF-IDF weight of the word t in the document d i , d i is the i-th maintenance work order text, belonging to the maintenance work order set D, and t is a certain word in the maintenance work order text (such as "replace", "bearing");

[0090] (2) Map the maintenance operation type and equipment components, using the BiLSTM model, expressed as:

[0091] Among them, X t is the input of the maintenance work order text, and H t is the maintenance feature vector output by BiLSTM, is the forward LSTM, processing the text sequence from left to right, is the reverse LSTM, processing the text sequence from right to left, H t-1 is the hidden state of the forward LSTM at time t - 1, and H t+1 is the hidden state of the reverse LSTM at time t + 1;

[0092] Generate the maintenance event - equipment status association matrix through BiLSTM, expressed as:

[0093] A ij = σ(W h H i + W e P j + b);

[0094] Among them, A ij is the association probability between the maintenance event e i and the equipment status s j , e i is the i-th maintenance event, s j is the j-th equipment status, H i is the BiLSTM feature vector of the maintenance event e i , P j is the embedding vector of the equipment status s j , W h and We is a mapping matrix used to calculate the matching degree between maintenance events and device states, and b is a bias term used to adjust the calculation of the matching degree;

[0095] S15: Perform feature-level fusion on the standardized data stream, spatio-temporal correlation data block, and maintenance feature vector to generate a multi-source heterogeneous data feature set. The fusion includes establishing a feature association weight distribution table based on the device topology relationship, performing tensor splicing on heterogeneous data sources, and then reducing the dimension through a feature selection algorithm to output a three-dimensional feature tensor containing the time dimension, device dimension, and feature dimension.

[0096] In S12, preprocess the operating parameter data, including:

[0097] S121: Perform wavelet packet decomposition and noise reduction on the vibration spectrum data, and extract the energy entropy feature of the specified frequency band;

[0098] Perform wavelet packet decomposition on the vibration spectrum data, expressed as: X vib (t) = ∑ j ∑ k w j,k ψ j,k (t), where X vib (t) is the vibration spectrum data, w j,k is the wavelet packet decomposition coefficient, and ψ j,k (t) is the j-th layer and k-th wavelet basis function;

[0099] Calculate the frequency band energy entropy E j , expressed as: E j = -∑ k P j,k logP j,k , where, is the normalized energy distribution;

[0100] S122: Use the dynamic baseline calibration algorithm to eliminate environmental interference for the temperature gradient data;

[0101] The dynamic baseline calibration of the temperature gradient data is expressed as: Let the temperature data be: X temp (t) = T obs (t) - T baseline (t), where T obs (t) is the temperature measured by the sensor, and T baseline (t) is the dynamic baseline based on the long-term average of the environment, where N is the baseline window length;

[0102] S123: Perform Fourier transform on the current harmonic feature data and match it with a preset harmonic feature template.

[0103] Fourier transform of the current harmonic characteristic data to calculate the harmonic spectrum: Among them, X harm (f) is the harmonic spectrum amplitude at frequency f during the operation of the device, and x harm (t) is the time-domain current signal measured during the operation of the device, is the Fourier transform operator, which converts the time-domain signal to the frequency domain;

[0104] Calculate the harmonic characteristic matching degree, expressed as: Among them, X ref (f) is the preset harmonic characteristic template, S harm is the harmonic characteristic matching degree, X harm (f i ) is the harmonic spectrum amplitude at frequency f under the current operating state of the device, X i at, X ref (f i ) is the preset harmonic characteristic template, and the reference harmonic spectrum amplitude at frequency f i at, f i is the harmonic frequency component.

[0105] S2 includes:

[0106] S21: Analyze the three-dimensional feature tensor in the multi-source heterogeneous data set, construct a dynamic adjacency matrix based on the device topology relationship, and the node weights of the dynamic adjacency matrix are jointly calculated by the physical connection tightness and process coupling degree of the device;

[0107] Calculate the physical connection tightness of the device, expressed as: Among them, is the physical connection tightness (range 0-1) between device i and device j, L ij is the connection strength (such as the number of connection components, the number of cables) between device i and j, d ij is the physical distance between device i and j, and ∈ is a small value to prevent the denominator from being zero;

[0108] Calculate the process coupling degree of the device, expressed as: Among them, is the process coupling degree (range 0-1) between device i and j, C ij (t) is the process cooperation degree (such as material transfer, data interaction frequency) between device i and j at time t, C max (t) is the maximum process cooperation degree in the device network at time t;

[0109] Construct a dynamic adjacency matrix, expressed as: Among them, A ij represents the dynamic adjacency weight between device i and device j, λ 1and λ 2 are adjustment parameters used to balance the effects of physical connection tightness and process coupling degree;

[0110] S22: Input the three-dimensional feature tensor and the dynamic adjacency matrix into a preset deep autoencoder, and perform cross-device feature propagation through the topological graph attention mechanism to generate a latent feature vector. The topological graph attention mechanism includes:

[0111] Calculate the feature similarity of adjacent device nodes as the attention weight;

[0112] Calculate the feature similarity of adjacent device nodes, expressed as:

[0113] where α ij is the attention weight of device i to device j, F i is the feature vector of device i, sim(F i , F j ) is the feature similarity calculation function, and Γ(i) is the neighbor set of device i;

[0114] Establish a cross-device feature propagation channel in the encoder hidden layer, expressed as:

[0115] where, is the hidden feature representation of device i in the l+1 layer, W (l) is the weight matrix of the l layer, is the hidden feature of device j in the l layer;

[0116] S23: Perform degradation trend enhancement processing on the latent feature vector, including:

[0117] Adopt residual connection to stack the device historical degradation features (from historical maintenance records), expressed as:

[0118] Z i = H i + γR i , where Z i is the latent feature vector of device i after degradation trend enhancement, H i is the original latent feature vector of device i, R i is the historical degradation feature of device i (from maintenance records), and γ is an adjustment parameter used to balance the effects of the current feature and the historical degradation feature;

[0119] Calculate the cumulative offset of the feature vector through a time sliding window, expressed as:

[0120] where, D i is the feature cumulative offset of device i within the time window, W is the sliding window size, Zi (t) is the degradation feature vector of device i at time t;

[0121] S24: Integrate the latent feature vector after enhanced degradation trend processing with the dynamic adjacency matrix to generate a set of device degradation feature vectors. Each vector in the set contains the device body degradation index and the associated device influence coefficient.

[0122] S24 includes:

[0123] S241: Extract the weight parameters of device nodes from the dynamic adjacency matrix to generate an associated influence coefficient matrix, where each coefficient represents the intensity value of the degradation influence of an associated device on the current device. The steps are as follows:

[0124] (1) Extract the associated weights from the dynamic adjacency matrix: Select the target device and its adjacent devices, and obtain their connection weights. Combine the physical connection tightness and process coupling degree between devices to calculate the degradation influence degree of a device on its adjacent devices;

[0125] (2) Normalization processing: Use the normalization method to make the sum of the associated influence coefficients of each device equal to 1 to ensure the comparability of the influence degrees between different devices;

[0126] (3) Output: Generate an associated influence coefficient matrix, where each element of the matrix represents the intensity of the degradation influence of a device on another device;

[0127] S242: Perform feature dimension expansion on the latent feature vector after enhanced degradation trend processing, add the device association topology dimension, and form a four-dimensional feature tensor (time × device × feature × topology). The steps are as follows:

[0128] Time dimension expansion: Maintain the structure of the original time series to ensure the integrity of time series information.

[0129] Device dimension expansion: The degradation feature vector of each device is retained, and the information of associated devices is added to its feature dimension.

[0130] Topology dimension expansion: According to the dynamic adjacency matrix, expand the feature vector of each device and add a topology structure dimension so that the feature tensor can express both the characteristics of the device itself and the influence of associated devices.

[0131] Output: Form a four-dimensional feature tensor with the data structure of (time × device × feature × topology).

[0132] S243: Perform the Hadamard product operation on the associated influence coefficient matrix and the four-dimensional feature tensor to realize the quantitative injection of topological association influence and generate a weighted feature tensor. The steps are as follows:

[0133] Define the Hadamard product operation: perform element-wise multiplication according to the device nodes to fuse the four-dimensional feature tensor and the associated influence coefficient matrix;

[0134] Calculate the weighted feature tensor, denoted as: T weighted = T 4D ° A assoc , where T weighted is the weighted feature tensor, T 4D is the four-dimensional feature tensor (time × device × feature × topology), and A assoc is the associated influence coefficient matrix, and ° represents element-wise multiplication according to the device nodes;

[0135] Output: Generate a weighted feature tensor, where the feature values of each device are affected by the topological structure, reflecting the degradation propagation relationship between devices;

[0136] S244: Perform cross-dimensional feature compression on the weighted feature tensor, and the processing steps are as follows:

[0137] Perform sliding window average pooling along the time dimension: Adopt the sliding window strategy to perform average pooling in the time dimension to reduce noise and smooth the time series features. By selecting the window size, ensure that the time series pattern is retained.

[0138] Select the Top-K most correlated influence feature channels along the topology dimension: Calculate the influence coefficient of each device in the topology dimension, and select the K channels with the greatest influence to reduce the computational complexity. Through feature screening, only retain the most representative topological relationship information.

[0139] Output: Generate a compressed feature tensor, and the data structure is (time × device × feature).

[0140] S245: Perform temporal modeling on the compressed feature tensor through a gated recurrent unit (GRU), and output a set of device degradation feature vectors, and the steps are as follows:

[0141] GRU processes temporal data: Adopt the GRU network to learn on the time series to model the device degradation trend.

[0142] Extract the device degradation index: The GRU hidden state is used to represent the degradation degree of the current device, serving as the device ontology degradation index.

[0143] Extract the influence coefficients of associated devices: Extract the influence coefficients of the Top-K associated devices from the weighted feature tensor to quantify the degradation contribution of other devices to the target device.

[0144] Output the degradation index of the device body and the influence coefficient of associated devices. The degradation index of the device body represents the current device degradation level characterized by the GRU hidden state, and the influence coefficient of associated devices is the Top-K associated device weight value extracted from the weighted feature tensor.

[0145] S3 includes:

[0146] S31: Extract the fault event feature vectors associated with the current device from the historical fault database. The specific steps are as follows:

[0147] S311: Based on the process coupling degree threshold (≥0.5) in the dynamic adjacency matrix (from S21), screen the associated devices with effective fault propagation paths and generate a list of associated devices;

[0148] S312: Extract the historical fault records of all devices in the list of associated devices from the historical fault database. The records include the fault type, occurrence time, and fault propagation path;

[0149] S313: Encode the fault type into a numerical vector with the same dimension as the device degradation feature vector set (from S24). The encoding rules include:

[0150] Assign orthogonal encoding vectors to mechanical faults (such as bearing wear, gear fracture);

[0151] Use one-hot encoding for electrical faults (such as overload, insulation failure);

[0152] Decompose the compound fault into a linear combination of basic fault types;

[0153] Output: A set of fault event feature vectors (dimension: fault type × time × device);

[0154] S32: Based on the parameters of the time sliding window and the window length adjustment rule, assign time-dependent weights to the historical fault data to generate a weighted fault feature sequence. The assignment rules include:

[0155] Apply an exponential decay weight to the fault data within the time window according to the occurrence time;

[0156] Apply a path decay coefficient to the cross-device propagation fault (calculated based on the node weights in the dynamic adjacency matrix of S21);

[0157] The specific steps are as follows:

[0158] S321: According to the time sliding window parameters (window length T w and step size ΔT) of S23, intercept the historical fault data within the window;

[0159] S322: Calculate the time-dependent decay weight, expressed as: where λ is the attenuation factor (default value 0.1), and t fault is the fault occurrence time;

[0160] S323: Calculate the path attenuation coefficient, expressed as: where w node is the node weight in the S21 dynamic adjacency matrix, and k is the process sensitivity factor (default value 2);

[0161] S324: Apply a comprehensive weight to the eigenvector of each fault event, expressed as: f weighted = f raw ·w time ·w path ;

[0162] Output: Weighted fault feature sequence (dimension: time window × fault type × weight value);

[0163] S33: Construct a Bayesian network prediction model, and perform multi-modal input on the equipment degradation eigenvector set and the weighted fault feature sequence. The specific steps are as follows:

[0164] S331: Define the hierarchical structure of the Bayesian network:

[0165] Root node: Equipment body degradation index (from the "equipment body degradation index" field in the equipment degradation eigenvector set of S24);

[0166] Intermediate node: Associated equipment influence coefficient (from the "associated equipment influence coefficient" field of S24);

[0167] Leaf node: Weighted fault feature sequence (from S32);

[0168] S332: Establish the conditional probability relationship between nodes:

[0169] The conditional probability table (CPT) from the root node to the intermediate node is dynamically generated by the equipment topology relationship (dynamic adjacency matrix);

[0170] The CPT from the intermediate node to the leaf node is obtained through training with historical fault data;

[0171] S333: Inject a prior distribution correction factor: After normalizing the weighted fault feature sequence by the time window dimension, it is used as the prior probability distribution correction amount for the leaf node;

[0172] Output: Initialized Bayesian network prediction model;

[0173] S34: Dynamically update the conditional probability table of the Bayesian network through a sliding window mechanism. The specific steps are as follows:

[0174] S341: Monitor the standardized data stream of S12 in real time and trigger the time window reset when any of the following conditions is detected;

[0175] Sudden change in device load (instantaneous change rate ≥ 30;

[0176] Operation mode switching signal (such as production batch change instruction);

[0177] S342: Calculate the error rate E based on the cumulative offset of S23 r , and adjust the time window step size ΔT, expressed as:

[0178]

[0179] S342: Update the Bayesian network conditional probability table:

[0180] For the root node CPT, update it using the maximum likelihood estimation of the device degradation index within the sliding window;

[0181] For the intermediate node CPT, adjust it based on the sliding average of the associated device influence coefficient,

[0182] Output: Optimized Bayesian network prediction model;

[0183] S35: Perform Bayesian network inference and output the device failure probability prediction matrix. The specific steps are as follows:

[0184] S351: Perform Bayesian network inference and calculate the occurrence probability of each device D i in the failure mode F j , expressed as:

[0185]

[0186] where C k is the associated device influence state corresponding to the intermediate node;

[0187] S352: Calculate the failure propagation chain gradient, expressed as:

[0188] indicating the influence intensity of device i's failure on device m's failure probability;

[0189] S353: Construct the matrix elements:

[0190] Row dimension: Device numbers (D 1 , D 2 ,..., D N );

[0191] Column dimension: Failure modes (F 1 , F 2 ,..., F M);

[0192] Element value: P(F j |D i );

[0193] Additional dimension: Chain gradient matrix G N×N ;

[0194] Output: Equipment failure probability prediction matrix (main matrix N×M + gradient matrix N×N);

[0195] Example (injection molding machine production line):

[0196] S31 Fault data screening:

[0197] Dynamic adjacency matrix display: The process coupling degree between the injection molding machine (equipment A) and the mold temperature controller (equipment B) = 0.7, and the coupling degree with the conveyor belt (equipment C) = 0.3;

[0198] Only select the historical fault data of equipment B (such as "heating tube failure", "circulation pump blockage");

[0199] S32 Weight calculation:

[0200] Fault occurrence time difference Δt = 8 hours, aging weight w time = e -0.1×8 = 0.449;

[0201] Node weight of equipment B w node = 0.7, path weight w path = 1 / (1 + e -2×0.7 ) = 0.802;

[0202] Comprehensive weight = 0.449 × 0.802 = 0.360;

[0203] S35 Matrix output:

[0204] The "screw jamming" fault probability of equipment A = 58.3%;

[0205] Gradient matrix display: The gradient G of the "heating tube failure" of equipment B on the failure of equipment A B→A = 0.62.

[0206] S4 includes:

[0207] S41: Integrate the equipment failure probability prediction matrix, equipment maintenance cost parameters (parsing results of maintenance work order texts in the historical maintenance records from S1), and the associated equipment influence coefficient to generate a structured decision input set. The integration includes:

[0208] Extract equipment failure probability prediction data: Extract the failure occurrence probability information of each device from the equipment failure probability prediction matrix to quantify its potential risk level. In addition, extract the failure propagation influence degree between devices from the chain probability gradient to measure the possible impact on other devices after a failure occurs.

[0209] Parse the maintenance work order text in the historical maintenance records: Extract the equipment maintenance cost parameters from the maintenance work order text, including:

[0210] Standard man-hour cost: Calculate the average maintenance time cost of different types of devices based on the maintenance task types and execution times in the historical maintenance work orders.

[0211] Spare part replacement cost: Extract the spare part replacement information in the historical maintenance records and calculate the spare part procurement and replacement costs.

[0212] Downtime loss cost: Combine the historical maintenance data and production plan to quantify the production losses caused during the equipment maintenance process.

[0213] Extract equipment correlation influence information: Extract the correlation equipment influence coefficient from the equipment topology structure and failure propagation path. This coefficient reflects the influence degree of the maintenance or failure of a certain device on other devices, ensuring that the optimization model fully considers the mutual influence between devices when formulating maintenance strategies.

[0214] S42: Use the generated structured decision input set as input data to construct a multi-objective optimization function, including minimizing the total maintenance cost, maximizing the suppression of failure risks, and maximizing the reduction of correlation influences;

[0215] S43: Define dynamic constraint conditions based on the structured decision input set. The constraint conditions include time window constraints, resource conflict constraints, and equipment mutual exclusion constraints;

[0216] S44: Dynamically adjust the model parameters according to the equipment degradation indicators in the real-time standardized data stream and the cumulative offset, including:

[0217] Adjust the maintenance priority according to the equipment degradation indicators: The degradation characteristics of the equipment can be calculated through the cumulative offset. When the degradation degree of a certain device exceeds the set threshold, its maintenance priority needs to be increased to ensure that high-risk devices are maintained first.

[0218] Adjust the maintenance time window according to the real-time standardized data stream: The real-time standardized data stream can reflect the current operating state of the equipment and provide information about production plans, equipment utilization rates, etc. Based on this data, the model dynamically adjusts the maintenance time window. For example, arrange more maintenance tasks during periods of low production demand to reduce the impact on production.

[0219] Adjust and optimize weights according to the equipment degradation trend: When the equipment degradation trend intensifies, the model needs to increase the weight of the fault risk suppression target to ensure that maintenance tasks can be executed in a timely manner and reduce the likelihood of sudden failures.

[0220] S45: Based on the multi-objective optimization function, dynamic constraint conditions, and real-time adjustment parameters, construct a dynamic maintenance strategy decision model, and perform optimization and solution through the multi-objective optimization algorithm to output a set of candidate strategies.

[0221] Generate a set of candidate maintenance strategies: The result of the optimization solution outputs a set of maintenance strategies that meet the constraint conditions. Each maintenance strategy includes:

[0222] List of equipment to be maintained and execution timing: Clearly define the maintenance execution time for each device to ensure that maintenance tasks are carried out according to the plan.

[0223] Estimated maintenance cost and failure risk reduction rate: Calculate the cost consumption of the maintenance strategy and evaluate its effect on reducing the failure risk.

[0224] Index for reducing the impact on associated equipment: Analyze the degree of impact of the maintenance task on other equipment to ensure that the maintenance strategy can minimize the failure correlation effect between equipment to the greatest extent.

[0225] S5 includes:

[0226] S51: Application of genetic algorithm: In the dynamic maintenance strategy decision model, multi-objective optimization involves minimizing maintenance cost, maximizing failure risk suppression, and maximizing reduction of associated impact. Optimization and solution are performed through the genetic algorithm. The genetic algorithm is a global optimization method suitable for optimization problems under complex constraints, as follows:

[0227] (1) Initialize the population: The execution situation of the maintenance task is represented by binary coding. Each individual corresponds to a maintenance strategy, where:

[0228] Gene bit 1 indicates that the device performs maintenance during the current maintenance cycle;

[0229] Gene bit 0 indicates that the device does not perform maintenance.

[0230] Generate the initial population by extracting information on high-priority devices from historical maintenance records to ensure that the early population has a tendency towards better solutions.

[0231] (2) Calculate the fitness function: The fitness function is determined by the multi-objective optimization function. Calculate the fitness value F of each individual, expressed as:

[0232] F = w 1 ∑C 维护 + w 2 ∑P 故障 (1 - xi ) + w 3 ∑G i→j x i (1 - x j );

[0233] Among them, the weight parameter w 1 , w 2 , w 3 Adopts a dynamic adjustment mechanism to optimize the weight ratio in real time according to the changes in the device status.

[0234] (3) Selection, crossover and mutation operations:

[0235] Selection strategy: Adopts tournament selection, and selects individuals with higher fitness from the population as parents.

[0236] Crossover strategy: Adopts single-point crossover, and the crossover point is dynamically adjusted based on the device priority to increase the probability of maintaining high-priority devices.

[0237] Mutation strategy: Adopts adaptive mutation, mutates individuals with a small probability, and preferentially adjusts the maintenance status of critical devices during the mutation process.

[0238] (4) Iteration termination condition: When the population evolves to a certain number of generations or the fitness value converges, stop the iteration, and select the optimal individual as the final optimization result.

[0239] S52: Determine the execution order of maintenance tasks: The optimal individual obtained by solving through the genetic algorithm represents the set of devices to be maintained. To ensure the efficient execution of maintenance tasks, optimize the execution order of maintenance tasks as follows:

[0240] (1) Basis for task sorting:

[0241] Sorting by device failure probability: Devices with high failure probability are given priority for maintenance;

[0242] Sorting by device association impact: Critical devices that affect other devices are given priority for execution;

[0243] Avoiding maintenance resource conflicts: Ensure that high-priority tasks are executed under the condition of available resources.

[0244] (2) Task sorting method: Adopts a weighted priority scheduling algorithm to calculate the device priority score, expressed as:

[0245] S i =α 1 P 故障,i +α 2 U i -α 3 R i ;

[0246] Among them, S i represents the priority score of device i, P 故障,i is the failure probability of device i, U i is the maintenance priority adjustment coefficient of device i, R i is the maintenance resource requirement of device i, α 1 is the failure probability weight coefficient, α 2 is the maintenance priority weight coefficient, α 3 is the resource requirement weight coefficient;

[0247] Calculate the priority scores of all devices, sort them by score, and form the final execution order of maintenance tasks.

[0248] S5 also includes:

[0249] S53: Generate the optimal maintenance instruction sequence: Based on the execution order of maintenance tasks, generate the maintenance instruction sequence and construct the maintenance task instruction set, the steps are as follows:

[0250] (1) Maintenance instruction format: Each maintenance instruction contains the following information:

[0251] Device number: Uniquely identify the maintenance device;

[0252] Maintenance task: Replace parts, lubricate, calibrate, detect, etc.;

[0253] Execution time: Specific maintenance time period;

[0254] Resource allocation: Allocate maintenance personnel, tools and spare parts;

[0255] Task priority: Indicate the urgency of the task.

[0256] (2) Instruction set construction: Organize all maintenance instructions into a maintenance task instruction set according to the time sequence and resource constraints to ensure no resource conflicts between tasks;

[0257] S54: Issue maintenance instructions: Issue the maintenance instructions to the maintenance execution unit and perform real-time monitoring to ensure that the tasks are executed according to the plan, the steps are as follows:

[0258] (1) Maintenance instruction issuance: According to the device type, the maintenance instructions are issued through different communication protocols:

[0259] The intelligent maintenance management system processes manual maintenance tasks;

[0260] The PLC control system analyzes the device-level maintenance instructions;

[0261] The industrial Internet of Things gateway remotely issues maintenance tasks.

[0262] (2) Maintenance task monitoring:

[0263] Status feedback: Real-time record of the maintenance task status, including "pending execution", "executing", "completed", "abnormal";

[0264] Abnormality detection: Detect abnormal situations during task execution, such as resource shortage, equipment unavailability, etc.;

[0265] Execution confirmation: After the maintenance task is completed, update the maintenance database and record the maintenance log.

[0266] The present invention covers any alternatives, modifications, equivalent methods and solutions made within the spirit and scope of the present invention. For the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention without these detailed descriptions. In addition, well-known methods, processes, procedures, components and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0267] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A predictive maintenance method for IoT equipment in a smart factory, characterized in that: The following steps are involved: S1: Collect the operating parameter data, environmental monitoring data and historical maintenance records of the target equipment to generate a multi-source heterogeneous data feature set; S2: performing implicit feature extraction on the multi-source heterogeneous data feature set to generate a set of device degradation feature vectors, wherein the implicit feature extraction includes feature space mapping based on a device association topology structure; S3: Generate an equipment failure probability prediction matrix based on the equipment degradation feature vector set and historical failure data in a dynamic time window; S4: constructing a dynamic maintenance strategy decision model based on the equipment failure probability prediction matrix and equipment maintenance cost parameters; S5: Solving the dynamic maintenance strategy decision model through a multi-objective optimization algorithm to generate an optimal maintenance instruction sequence, and sending the optimal maintenance instruction sequence to the maintenance execution unit of the target device.

2. The predictive maintenance method for smart factory IoT equipment according to claim 1 is characterized in that: The S1 includes: S11: collecting operating parameter data through a sensor array deployed on the device body, wherein the operating parameter data includes vibration spectrum data, temperature gradient data, and current harmonic characteristic data; S12: preprocessing the operating parameter data through an edge computing node to generate a standardized data stream; S13: Collect environmental parameter data of the area where the equipment is located through the environmental monitoring terminal, the environmental parameter data including air particulate matter concentration, humidity fluctuation curve and electromagnetic interference intensity, and align the environmental parameter data with the operating parameter data by time stamp to generate a spatiotemporal correlation data block; S14: parsing the maintenance work order text in the historical maintenance record of the equipment, and extracting the maintenance feature vector by natural language processing technology, wherein the extraction includes: Use the TF-IDF algorithm to vectorize keywords in the maintenance work order text; Identify the mapping relationship between maintenance operation types and equipment components based on the BiLSTM model; Construct maintenance event-equipment status association matrix; S15: The standardized data stream, spatiotemporal correlation data block and maintenance feature vector are subjected to feature-level fusion to generate a multi-source heterogeneous data feature set. The fusion includes establishing a feature correlation weight distribution table based on the device topology relationship, and performing tensor splicing on the heterogeneous data source and then reducing the dimension through a feature selection algorithm to output a three-dimensional feature tensor including time dimension, device dimension and feature dimension.

3. The predictive maintenance method for smart factory IoT equipment according to claim 2 is characterized in that: The step S12 pre-processes the operating parameter data, including: S121: performing wavelet packet decomposition and noise reduction on the vibration spectrum data to extract energy entropy features of a specified frequency band; S122: Using a dynamic baseline calibration algorithm to eliminate environmental interference on the temperature gradient data; S123: Perform Fourier transform on the current harmonic feature data and match it with a preset harmonic feature template.

4. The predictive maintenance method for smart factory IoT equipment according to claim 3 is characterized in that: The S2 includes: S21: parsing the three-dimensional feature tensor in the multi-source heterogeneous data feature set, and constructing a dynamic adjacency matrix based on the equipment topology relationship, wherein the node weights of the dynamic adjacency matrix are jointly calculated by the physical connection tightness of the equipment and the process coupling degree; S22: Input the three-dimensional feature tensor and the dynamic adjacency matrix into a preset deep autoencoder, perform cross-device feature propagation through a topological graph attention mechanism, and generate an implicit feature vector. The topological graph attention mechanism includes: Calculate the feature similarity of adjacent device nodes as attention weight; Establish a cross-device feature propagation channel in the encoder hidden layer; S23: performing degradation trend enhancement processing on the latent feature vector, including: Use residual connection to superimpose the historical degradation characteristics of the equipment; Calculate the cumulative offset of the feature vector through the time sliding window; S24: Fusion the latent feature vector after degradation trend enhancement processing and the dynamic adjacency matrix to generate a set of equipment degradation feature vectors, each vector in the set includes a degradation index of the equipment itself and an impact coefficient of the associated equipment.

5. The predictive maintenance method for smart factory IoT equipment according to claim 4 is characterized in that: The S24 includes: S241: extracting weight parameters of device nodes from the dynamic adjacency matrix to generate an association influence coefficient matrix, wherein each coefficient represents an intensity value of the influence of an associated device on the current device degradation; S242: Expanding the feature dimension of the latent feature vector after the degradation trend enhancement processing, increasing the device association topology dimension, and forming a four-dimensional feature tensor; S243: performing a Hadamard product operation on the correlation influence coefficient matrix and the four-dimensional feature tensor to achieve quantitative injection of topological correlation influence and generate a weighted feature tensor; S244: performing cross-dimensional feature compression on the weighted feature tensor; S245: Perform time series modeling on the compressed feature tensor through a gated recurrent unit, and output a set of device degradation feature vectors.

6. The predictive maintenance method for smart factory IoT equipment according to claim 5 is characterized in that: The S3 includes: S31: extracting a fault event feature vector associated with the current device from a historical fault database; S32: Based on the parameters of the time sliding window and the window length adjustment rules, the historical fault data is assigned time weights to generate a weighted fault feature sequence, wherein the assignment rules include: Apply exponential decay weights to fault data within the time window according to the time of occurrence; Apply path attenuation coefficients to propagate faults across devices; S33: constructing a Bayesian network prediction model, and performing multi-modal input of the equipment degradation feature vector set and the weighted fault feature sequence; S34: Dynamically update the conditional probability table of the Bayesian network through a sliding window mechanism; S35: Execute Bayesian network reasoning and output a device failure probability prediction matrix.

7. The predictive maintenance method for smart factory IoT equipment according to claim 6 is characterized in that: The S4 includes: S41: Integrate the equipment failure probability prediction matrix, equipment maintenance cost parameters and related equipment impact coefficients to generate a structured decision input set; S42: Using the generated structured decision input set as input data, construct a multi-objective optimization function, including minimizing the total maintenance cost, maximizing the failure risk suppression, and maximizing the reduction of associated impacts; S43: defining dynamic constraints based on the structured decision input set, where the constraints include time window constraints, resource conflict constraints, and device mutual exclusion constraints; S44: dynamically adjust model parameters based on real-time standardized data streams and accumulated offsets; S45: Based on the multi-objective optimization function, dynamic constraints and real-time adjustment parameters, a dynamic maintenance strategy decision model is constructed, and an optimization solution is performed through a multi-objective optimization algorithm to output a candidate strategy set.

8. The predictive maintenance method for smart factory IoT equipment according to claim 7 is characterized in that: The S5 includes: S51: Application of Genetic Algorithm: In the dynamic maintenance strategy decision model, multi-objective optimization involves minimizing maintenance cost, maximizing failure risk suppression, and maximizing associated impact reduction, which are optimized and solved by genetic algorithm; S52: Determine the execution order of the maintenance tasks: obtain the set of equipment to be maintained by solving the genetic algorithm, and optimize the execution order of the maintenance tasks.

9. The predictive maintenance method for smart factory IoT equipment according to claim 8, characterized in that: The S5 further includes: S53: Generate an optimal maintenance instruction sequence: Generate a maintenance instruction sequence based on the execution order of the maintenance tasks, and construct a maintenance task instruction set; S54: issuing maintenance instructions: the maintenance instructions are issued to the maintenance execution unit and are monitored in real time.

Images