Industrial equipment early warning system and method

Abstract

The invention discloses an industrial equipment early warning system and method, and the system comprises a multi-mode sensing network module which collects equipment operation data in real time, and carries out the preprocessing of the equipment operation data; the feature extraction module is used for performing feature extraction on the operation data of the multi-source equipment; the data fusion module is used for carrying out feature fusion on the extracted features and generating equipment health state feature vectors; the health state evaluation module is used for analyzing the feature vectors, evaluating the current health state of the equipment and predicting the change trend of the future 12-24 hours; the service life prediction module is used for predicting the residual service life of the equipment based on the historical operation data and the current state data of the equipment and generating a maintenance suggestion according to a prediction result; and the model optimization module is used for carrying out optimization processing on the optimized life prediction model. According to the invention, accurate prediction of the residual service life of the equipment can be realized.

Description

Technical Field

[0001] The invention relates to an industrial equipment early warning system, belonging to the technical field of industrial equipment monitoring and maintenance. Background Art

[0002] With the rapid development of industrial automation, the reliability and maintenance efficiency of industrial equipment have become key factors restricting the production efficiency and economic benefits of enterprises.

[0003] Traditional equipment maintenance relies primarily on scheduled inspections and post-failure repairs, which present numerous problems. Equipment failure predictions are inaccurate, with no early warnings, leading to frequent unplanned downtime. Maintenance strategies lack specificity, often relying on scheduled maintenance, resulting in wasted maintenance resources. Equipment health assessments rely on a single parameter, failing to fully reflect the equipment's actual operating status. Accurate predictions of the equipment's remaining useful life are lacking, making it impossible to formulate scientific maintenance plans. Monitoring data for each device is independent, failing to fully utilize the interconnected information across multiple devices.

[0004] The existing equipment monitoring systems based on the Internet of Things and artificial intelligence have limitations. The sensor types are single and only monitor a few parameters such as temperature and vibration, making it difficult to fully capture changes in equipment status; the data processing methods are simple, mainly relying on threshold judgment and simple statistical analysis, and lack the ability to model complex nonlinear relationships; the prediction model has limited accuracy and is difficult to adapt to different equipment types and operating environments; the system has poor scalability, making it difficult to achieve knowledge sharing and model optimization among multiple devices.

[0005] In terms of market demand, with the transformation and upgrading of the manufacturing industry and the development of intelligent manufacturing, predictive maintenance has become a key technology for improving equipment reliability and production efficiency. In manufacturing industries such as steel, automotive, and electronics, equipment failures result in significant downtime and losses, creating an urgent need for equipment health monitoring and fault prediction. Similarly, in energy industries such as electricity, oil, and natural gas, equipment reliability is directly linked to the security and stability of energy supply, creating a strong demand for highly reliable equipment monitoring systems. Furthermore, in transportation sectors such as railways, aviation, and shipping, as well as in medical equipment and smart buildings, there is a significant demand for equipment health monitoring and fault prediction. Summary of the Invention

[0006] The technical problem to be solved by the present invention is: how to achieve real-time monitoring of the health status of industrial equipment, fault prediction and life assessment.

[0007] To solve the above technical problems, the present invention provides an industrial equipment early warning system, comprising:

[0008] Multimodal sensor network module: collects equipment operation data in real time and pre-processes the equipment operation data;

[0009] Feature extraction module: extract features from multi-source equipment operation data;

[0010] Data fusion module: performs feature fusion on the extracted features to generate the equipment health status feature vector;

[0011] Health status assessment module: Analyzes feature vectors, evaluates the current health status of the device, and predicts the change trend in the next 12-24 hours;

[0012] Life prediction module: Based on the equipment's historical operating data and current status data, it predicts the equipment's remaining service life and generates maintenance recommendations based on the prediction results;

[0013] Model optimization module: used to optimize the life prediction model.

[0014] The aforementioned industrial equipment early warning system includes a multi-parameter sensor in the multimodal sensor network module, which dynamically adjusts the sampling frequency using an adaptive sampling method according to the equipment operating status. for:

[0015]

[0016] in, represents the information entropy of the sensor data at the current moment, and Represent the minimum and maximum information entropy of historical data, and represent the minimum and maximum sampling frequencies respectively, Indicates time.

[0017] The aforementioned industrial equipment early warning system uses a multi-scale feature extraction method in the feature extraction module to extract features from multi-source data. The time domain feature calculation includes statistical features such as mean, variance, kurtosis, and skewness; the frequency domain feature is calculated by Fourier transform to calculate the power spectrum density and frequency band energy ratio; the time-frequency feature is extracted by wavelet transform and empirical mode decomposition, and the feature vector The formula is:

[0018]

[0019]

[0020]

[0021]

[0022] in, Represents the time domain feature vector, which is composed of time domain statistical features. is the mean, is the variance, is the skewness coefficient, is the kurtosis coefficient, is the peak factor, is the form factor;

[0023] represents the frequency domain eigenvector, is the power spectral density function, They are frequency boundary value 1, frequency boundary value 2, frequency boundary value 3, and frequency boundary value 4 respectively. is the probability density function, is the center frequency, is the frequency bandwidth;

[0024] represents the time-frequency domain eigenvector, is the wavelet packet energy, is the number of decomposition layers, For the Intrinsic mode functions;

[0025] Represents the depth vector feature.

[0026] In the aforementioned industrial equipment early warning system, the Dempster-Shafer evidence theory is used in the feature fusion module to assign weights according to the reliability of different sensor data. For fault type The basic probability distribution is , then the comprehensive fusion result for:

[0027]

[0028] in, represents the conflict factor, is an empty set, For evidence collection, is the set intersection operator.

[0029] The aforementioned industrial equipment early warning system uses a hybrid model combining deep neural networks and survival analysis in the health status assessment module to achieve real-time assessment of the equipment health status. The deep neural network includes an input layer, three hidden layers, and an output layer structure, and calculates the equipment health index through forward propagation. Output of the layer Expressed as:

[0030]

[0031] in, and Respectively The layer's weight matrix and bias vector, is the ReLU activation function, defined as , Output 0 when Time output , which can alleviate the vanishing gradient and speed up the convergence;

[0032] The health vector output by the neural network is used as the covariate of the Cox proportional hazard model and trained with the same loss function as the neural network. The failure risk function of the equipment is Defined as:

[0033]

[0034] in, is the benchmark risk function, It is an influencing factor. is the corresponding regression coefficient; the reliability function of the equipment and remaining life expectancy Can be expressed as:

[0035]

[0036]

[0037] in, is the reliability function, which indicates the equipment The probability of normal operation within the is the risk failure rate function, which represents the conditional probability of failure occurring in unit time; For time The reliability value at time The probability of normal operation at all times; The average remaining life, which means the equipment has been running The expected remaining useful life after time.

[0038] In the aforementioned industrial equipment early warning system, the life prediction model is based on a hybrid architecture of an LSTM model and a Transformer model;

[0039] The LSTM model captures time series features through the following recursive calculation:

[0040]

[0041] in, 、 、 They are forget gate, input gate and output gate respectively. is the cell state, is the candidate cell state, is hidden state, is the sigmoid function, is the input weight matrix, is the recursive weight matrix, are bias vectors, For time The input vector, and They are tanh function one and tanh function two, is the element-wise multiplication operator;

[0042] The Transformer model models long-term dependencies through a multi-head self-attention mechanism. The self-attention is calculated as follows:

[0043]

[0044] in, 、 、 are query, key, and value matrices respectively, is the dimension of the key. Multi-head attention is calculated by the following formula:

[0045]

[0046] in, represents the multi-head attention mechanism function, represents the connection function, is the number of attention heads, is the output linear transformation matrix, For the The parameter matrix of the attention head.

[0047] Both the LSTM model and the Transformer model adopt a cascade-parallel hybrid architecture.

[0048] The aforementioned industrial equipment early warning system has a cascade-parallel hybrid architecture represented as follows:

[0049] Cascade: LSTM first captures short-term / local temporal features and outputs hidden sequences ;

[0050] Parallel: Hidden Sequence After linear mapping, it is fed into the Transformer multi-head self-attention layer as query / key / value to extract long-term dependencies;

[0051] Concatenation: The final hidden state of the LSTM model Global representation of the Transformer model Connect to feature , for subsequent remaining useful life (RUL) distribution estimation.

[0052] in this way, The matrix actually shares the weight mapping with the LSTM output, and the variables are correlated rather than independent.

[0053] Then, the Monte Carlo method is used for sampling prediction, and the final output is the probability distribution of the remaining service life of the equipment.

[0054]

[0055] in, Indicates the remaining service life of the device. represents the expected value operator, Indicates the equipment failure time, Indicates the current time, Indicates the historical status data of the device. is the conditional survival function, is the integration variable.

[0056] In the aforementioned industrial equipment early warning system, the model optimization module uses a federated learning algorithm to achieve knowledge sharing and model optimization among multiple devices for the life prediction model. The objective function of federated learning is defined as:

[0057]

[0058] in, Represents the global objective function, which represents the optimization goal of the entire federated learning system; is the model parameter vector, which represents the neural network weights that need to be optimized; Indicates the total number of devices participating in federated learning, Representation device The amount of data, Indicates the total amount of data, Representation device dataset.

[0059] In the aforementioned industrial equipment early warning system, the federated learning optimization algorithm includes the following steps:

[0060] Initialize the global model parameters to , for each round of communication : Select a subset of devices ; Each selected device Compute local updates: ;in, is the global model parameter of the current round, For equipment The updated local model parameters, is the learning rate, For equipment Local objective function exist The gradient at

[0061] Perform aggregate updates: ;in, is the data volume of device j, used for weighted averaging;

[0062] Differential privacy protection is achieved by adding noise. The updated parameters are:

[0063]

[0064] in, is the updated global model parameter; is the normally distributed noise, is the noise scale, is the gradient clipping limit, is the identity matrix; privacy budget and noise scale The relationship is:

[0065]

[0066] in, is the number of communication rounds, is the failure probability, is a constant.

[0067] An industrial equipment early warning method includes: 1) using a multimodal sensor network to collect equipment operation data in real time and preprocessing the equipment operation data;

[0068] 2) Use the feature extraction module to extract features from multi-source equipment operation data, and then use the data fusion module to fuse the extracted features to generate the equipment health status feature vector;

[0069] 3) Use the health status assessment module to analyze feature vectors, evaluate the current health status of the device, and predict the change trend in the next 12-24 hours;

[0070] 4) Based on the equipment's historical operating data and current status data, the life prediction module is used to predict the equipment's remaining service life and generate maintenance recommendations based on the prediction results;

[0071] 5) Use the model optimization module to optimize the life prediction model.

[0072] The beneficial effects achieved by the present invention are as follows: An industrial equipment early warning system and method of the present invention uses multi-parameter sensors such as temperature, vibration, noise, and current to collect equipment operation data, generates equipment health status feature vectors through multi-scale feature extraction and fusion technology, combines a hybrid model combining deep neural network and survival analysis to perform health status assessment, and realizes accurate prediction of the remaining service life of the equipment based on the hybrid architecture of LSTM and Transformer.

[0073] Federated learning technology is also used to enable knowledge sharing and continuous model optimization across multiple devices, improving prediction accuracy and system adaptability. This invention can effectively address issues such as inaccurate fault prediction and high maintenance costs for industrial equipment, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a structural diagram of the industrial equipment early warning system in Example 1 of the present invention;

[0075] Figure 2 This is a diagram showing the data processing principle of the industrial equipment early warning system in Example 1 of the present invention. DETAILED DESCRIPTION

[0076] The present invention will be described in further detail below with reference to the accompanying drawings.

[0077] Example 1

[0078] like Figure 1 and Figure 2 As shown, this embodiment provides an industrial equipment early warning system, including:

[0079] Multimodal sensor network module: collects equipment operation data in real time and pre-processes the equipment operation data;

[0080] Feature extraction module: extract features from multi-source equipment operation data;

[0081] Data fusion module: performs feature fusion on the extracted features to generate the equipment health status feature vector;

[0082] Health status assessment module: Analyzes feature vectors, evaluates the current health status of the device, and predicts the change trend in the next 12-24 hours;

[0083] Life prediction module: Based on the equipment's historical operating data and current status data, it predicts the equipment's remaining service life and generates maintenance recommendations based on the prediction results;

[0084] Model optimization module: used to optimize the life prediction model.

[0085] In the multimodal sensor network module, including temperature, vibration, noise, current and other multi-parameter sensors, the sampling frequency is dynamically adjusted according to the operating status of the equipment using an adaptive sampling method. The sampling frequency range is 1Hz-1kHz. The adaptive sampling method is adjusted based on the rate of change of information entropy. The sampling frequency for:

[0086]

[0087] in, represents the information entropy of the sensor data at the current moment, and Represent the minimum and maximum information entropy of historical data, and represent the minimum and maximum sampling frequencies respectively, Indicates time.

[0088] The feature extraction module uses a multi-scale feature extraction method to extract features from multi-source data. Time domain feature calculation includes statistical features such as mean, variance, kurtosis, and skewness; frequency domain features are calculated through Fourier transform to calculate power spectrum density and band energy ratio; time-frequency features are extracted through wavelet transform and empirical mode decomposition. Feature vector The formula is:

[0089]

[0090]

[0091]

[0092]

[0093] in, Represents the time domain feature vector, which is composed of time domain statistical features. is the mean, is the variance, is the skewness coefficient, is the kurtosis coefficient, is the peak factor, is the form factor;

[0094] represents the frequency domain eigenvector, is the power spectral density function, They are frequency boundary value 1, frequency boundary value 2, frequency boundary value 3, and frequency boundary value 4 respectively. is the probability density function, is the center frequency, is the frequency bandwidth;

[0095] represents the time-frequency domain eigenvector, is the wavelet packet energy, is the number of decomposition layers, For the Intrinsic mode functions;

[0096] Represents the depth vector feature.

[0097] The feature fusion module uses the Dempster-Shafer evidence theory to assign weights according to the reliability of different sensor data. For fault type The basic probability distribution is , then the comprehensive fusion result for:

[0098]

[0099] in, represents the conflict factor, is an empty set, For evidence collection, is the set intersection operator.

[0100] The health status assessment module uses a hybrid model that combines deep neural networks with survival analysis to achieve real-time assessment of the health status of the equipment. The deep neural network consists of an input layer, three hidden layers, and an output layer structure, and calculates the equipment health index through forward propagation. Output of the layer Expressed as:

[0101]

[0102] in, and Respectively The layer's weight matrix and bias vector, is the ReLU activation function, defined as , when input Output 0 when Time output , which can alleviate the gradient disappearance and speed up the convergence.

[0103] The health vector output by the neural network is used as the covariate of the Cox proportional hazard model and trained with the same loss function as the neural network. The failure risk function of the equipment is Defined as:

[0104]

[0105] in, is the benchmark risk function, It is an influencing factor. is the corresponding regression coefficient; the reliability function of the equipment and remaining life expectancy Can be expressed as:

[0106]

[0107]

[0108] in, is the reliability function, which indicates the equipment The probability of normal operation within the is the risk failure rate function, which represents the conditional probability of failure occurring in unit time; For time The reliability value at time The probability of normal operation at all times; The average remaining life, which means the equipment has been running The expected remaining useful life after time.

[0109] The life prediction model is based on a hybrid architecture of the LSTM model and the Transformer model to accurately predict the remaining service life of the equipment. The LSTM model captures time series features through the following recursive calculation:

[0110]

[0111] in, 、 、 They are forget gate, input gate and output gate respectively. is the cell state, is the candidate cell state, is hidden state, is the sigmoid function, is the input weight matrix, is the recursive weight matrix, are bias vectors, For time The input vector, and They are tanh function one and tanh function two, is the element-wise multiplication operator.

[0112] The Transformer model models long-term dependencies through a multi-head self-attention mechanism. The self-attention is calculated as follows:

[0113]

[0114] in, 、 、 are query, key, and value matrices respectively, is the dimension of the key. Multi-head attention is calculated by the following formula:

[0115]

[0116] in, represents the multi-head attention mechanism function, represents the connection function, is the number of attention heads, is the output linear transformation matrix, For the The parameter matrix of the attention head.

[0117] Both the LSTM model and the Transformer model use a cascade-parallel hybrid architecture:

[0118] Cascade: LSTM first captures short-term / local temporal features and outputs hidden sequences ;

[0119] Parallel: Hidden Sequence After linear mapping, it is fed into the Transformer multi-head self-attention layer as query / key / value to extract long-term dependencies;

[0120] Concatenation: The final hidden state of the LSTM model Global representation of the Transformer model Connect to feature , for subsequent remaining useful life (RUL) distribution estimation.

[0121] in this way, The matrix actually shares the weight mapping with the LSTM output, and the variables are correlated rather than independent.

[0122] Then, the Monte Carlo method is used for sampling prediction, and the final output is the probability distribution of the remaining service life of the equipment.

[0123]

[0124] in, Indicates the remaining service life of the device. represents the expected value operator, Indicates the equipment failure time, Indicates the current time, Indicates the historical status data of the device. is the conditional survival function, is the integration variable.

[0125] The Transformer model affects the mean, variance, or mixing weights in the distribution parameters, thereby affecting the Monte Carlo sampling results; both share the same loss function during the training phase.

[0126] The model optimization module uses a federated learning algorithm to achieve knowledge sharing and model optimization among multiple devices for the life prediction model. The objective function of federated learning is defined as:

[0127]

[0128] in, Represents the global objective function, which represents the optimization goal of the entire federated learning system; is the model parameter vector, which represents the neural network weights that need to be optimized; Indicates the total number of devices participating in federated learning, Representation device The amount of data, Indicates the total amount of data, Representation device dataset.

[0129] The federated learning optimization algorithm includes the following steps:

[0130] Initialize the global model parameters to , for each round of communication : Select a subset of devices ; Each selected device Compute local updates: ;in, is the global model parameter of the current round, For devices The updated local model parameters, is the learning rate, For devices Local objective function exist The gradient at

[0131] Perform aggregate updates: ;in, is the data volume of device j, used for weighted averaging;

[0132] Differential privacy protection is achieved by adding noise. The updated parameters are:

[0133]

[0134] in, is the updated global model parameter; is the normally distributed noise, is the noise scale, is the gradient clipping limit, is the identity matrix; privacy budget and noise scale The relationship is:

[0135]

[0136] in, is the number of communication rounds, is the failure probability, is a constant.

[0137] The present invention utilizes the above-mentioned modules to realize the monitoring of the health status and life prediction of industrial equipment. It can be applied to many fields such as manufacturing, energy industry, transportation, medical equipment and intelligent buildings. It can effectively reduce the unplanned downtime of equipment, improve equipment utilization, reduce maintenance costs and extend the service life of equipment.

[0138] The following example illustrates the lifespan prediction and maintenance of steam turbine generator sets in thermal power plants. For these turbine generator sets, multimodal sensors include not only conventional temperature, vibration, and current sensors, but also differential pressure sensors, speed sensors, and turbine efficiency monitoring devices. High-temperature, high-pressure components, such as the turbine inlet valve, bearing housing, and generator stator windings, utilize high-temperature sensors to obtain reliable monitoring data even under extreme operating conditions.

[0139] An online gas chromatograph is used to monitor the dissolved gas components in transformer oil in real time, including the concentrations of hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), and acetylene (C2H2). Gas ratio analysis is used to predict potential transformer faults. Transformer oil analysis data is integrated with other sensor data using an uncertainty fusion algorithm:

[0140]

[0141] in, It is around the expected value The normal cloud model, is entropy, It is super entropy. is the weight coefficient.

[0142] In the feature extraction module, based on the characteristics of the vibration data of the power generation equipment, the wavelet packet transform is used to perform multi-resolution analysis on the vibration signal, and the frequency characteristics are extracted using the following formula:

[0143]

[0144] in, represents the wavelet energy coefficient, Is the signal on scale and location The wavelet coefficients at and Represent the number of decomposition layers and node numbers respectively, is the integral differential, It is a modular operation.

[0145] To address the issue of declining steam turbine efficiency, the relative efficiency index is calculated using an efficiency evaluation model based on the entropy method:

[0146]

[0147] in, represents the relative efficiency index, is a parameter The membership function of is the number of parameters, is the weight coefficient, which is calculated by entropy method. In the steam turbine scenario, key operating indicators are mapped item by item, including main steam pressure, main steam temperature, reheat steam temperature, extraction / exhaust pressure, feed water temperature, steam flow or heat rate, bearing and blade vibration amplitude, speed deviation and axial displacement, and lubricating oil temperature / pressure.

[0148] In the health status assessment module, a method combining fault tree analysis and Bayesian networks is used to construct a probabilistic graphical model containing 95 nodes, targeting the characteristics of steam turbine generator sets. Through structural mapping, the top event in the fault tree is converted into the root node of the Bayesian network, covering the possible failure modes of the generator set. Among them, the intermediate event is used as the intermediate node, and the bottom event is used as the leaf node. Each node corresponds to a Boolean variable representing the device status. In the logic gate conversion rule, the AND gate sets the child node to fail when all parent nodes fail. Its conditional probability expression is:

[0149]

[0150] For the OR gate, it is assumed that when any parent node fails, the child node must also fail. Its conditional probability expression is:

[0151]

[0152] Here, 1 indicates that the Boolean value is 1, that is, the state is faulty; 0 indicates that the Boolean value is 0, that is, the state is normal.

[0153] The construction of the conditional probability table uses the Leaky Noisy-OR model to deal with uncertainty:

[0154]

[0155] in, Indicates the reason The weight that causes the failure. In the probability update, the variational inference algorithm is used:

[0156]

[0157] in, represents the optimized latent variable variational distribution, is a hidden variable, It means "proportional to", that is, the proportional relationship after ignoring the normalization constant. Indicates except Other hidden variables ( ) takes the expectation, is the observed variable, is a joint probability distribution. Variational inference provides a way to integrate the "health vector learned by the network" and the "Cox risk function" into the same training framework, allowing them to be jointly optimized end-to-end.

[0158] In the life prediction module, based on the long-term operating characteristics of the steam turbine generator set, the equipment life curve is divided into three stages: healthy period, degradation period, and failure period. Different prediction models are used for each stage. Among them, the Wiener process model is used to describe the equipment performance degradation trajectory in the degradation period:

[0159]

[0160] in, Indicates time The performance degradation of Indicates the initial performance status value, is the drift coefficient, is the diffusion coefficient, is the standard Brownian motion. The drift coefficient in the degradation phase is , diffusion coefficient The hidden vector of the hybrid network By using linear head prediction, parameter sharing is achieved. When approaching the threshold, the system switches to the RUL distribution prediction mode, and the hybrid network outputs the complete remaining life probability.

[0161] This embodiment was verified for 12 months on a 600MW generator set at a large thermal power plant. It successfully predicted potential faults such as cracks in the steam turbine inlet valve stem, aging of the generator stator winding insulation, and contamination of the bearing lubricating oil. The average early warning lead time reached 15 days, with an accuracy rate of 94.7%. During this period, it helped the power plant reduce unplanned equipment downtime by 78%, effectively improving the operational reliability and economy of the generator set. Especially during the two high-load power generation periods during the flood season, the system avoided emergency shutdowns due to sudden failures by adjusting the maintenance plan in real time, saving the power plant approximately 5.5 million yuan in maintenance costs and extending the overhaul cycle of the generator set from the standard 18-month maintenance cycle to 24 months.

[0162] Furthermore, the forecasting model was continuously optimized during operation, adaptively adjusting for the impact of seasonal load fluctuations and ambient temperature on equipment performance. This reduced the average absolute error of the forecasting model from an initial 7.8% to 3.2%. Federated learning with other power plants on the grid enabled the sharing of equipment operating experience and failure cases, further improving the model's generalization and forecasting accuracy.

[0163] Example 2

[0164] This embodiment provides an industrial equipment early warning method, comprising: 1) using a multimodal sensor network to collect equipment operation data in real time and preprocessing the equipment operation data;

[0165] The multimodal sensing network includes multimodal sensors, and the data of each sensor is pre-processed by the edge computing unit and uploaded to the central processing unit via industrial Ethernet or 5G network.

[0166] The multimodal sensing network dynamically adjusts the sampling frequency according to the operating status of the equipment, and the sampling frequency range is 1Hz-1kHz; the multimodal sensor includes a temperature sensor, a vibration sensor, an acoustic emission sensor and a current sensor.

[0167] 2) Use the feature extraction module to extract features from multi-source equipment operation data, and then use the data fusion module to fuse the extracted features to generate the equipment health status feature vector;

[0168] The feature extraction module uses a multi-scale feature extraction method to extract features, and the features include:

[0169] Time domain statistical characteristics: mean, variance, kurtosis, skewness;

[0170] Frequency domain characteristics: power spectrum density, frequency band energy ratio;

[0171] Time-frequency characteristics: wavelet packet energy, EMD decomposition;

[0172] Deep features: latent features extracted by the autoencoder;

[0173] The feature extraction process adopts a parallel computing architecture with a processing delay of less than 100ms and adaptive adjustment of feature dimensions to ensure a balance between the integrity of feature information and computational efficiency.

[0174] 3) Use the health status assessment module to analyze feature vectors, evaluate the current health status of the device, and predict the change trend in the next 12-24 hours;

[0175] The health status assessment module uses a hybrid model that combines a deep neural network with survival analysis. The neural network structure includes an input layer (64 neurons), three hidden layers (128, 256, and 128 neurons, respectively, using Reluctant Unit (ReLU) activation function), and an output layer (equipment health score). The health status assessment model integrates historical equipment maintenance records, operating parameters, and environmental factors, achieving a 95% health status assessment accuracy and providing a 24-48 hour warning lead time.

[0176] 4) Based on the equipment's historical operating data and current status data, the life prediction module is used to predict the equipment's remaining service life and generate maintenance recommendations based on the prediction results;

[0177] The life prediction module adopts a hybrid architecture of LSTM and Transformer. The LSTM layer is used to capture time series features, and the Transformer layer is used to model long-term dependencies. The life prediction model input includes equipment operating parameters, maintenance records, environmental data and historical failure data, and the output is the probability distribution of the remaining service life of the equipment. The prediction model has a mean absolute error of less than 5% on the test set and can predict equipment failures 7-14 days in advance.

[0178] 5) Use the model optimization module to optimize the life prediction model.

[0179] The model optimization module uses federated learning technology to achieve knowledge sharing among multiple devices and optimize the lifespan prediction model. It establishes a knowledge base for equipment health status, which includes failure modes, maintenance strategies, and optimization solutions. Data security is achieved through differential privacy protection, with a privacy budget set to 0.1. An asynchronous update strategy is used for knowledge sharing, improving the stability and real-time performance of the industrial equipment early warning system.

[0180] After deployment at industrial sites, the accuracy of equipment failure prediction reaches over 92%, and the average warning lead time reaches 48 hours; it can reduce unplanned equipment downtime by 65%, maintenance costs by 40%, and extend equipment service life by 25%.

[0181] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0182] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0183] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0184] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An industrial equipment early warning system, characterized in that: include: Multimodal sensor network module: collects equipment operation data in real time and pre-processes the equipment operation data; Feature extraction module: extract features from multi-source equipment operation data; Data fusion module: performs feature fusion on the extracted features to generate the equipment health status feature vector; Health status assessment module: Analyzes feature vectors, evaluates the current health status of the device, and predicts the change trend in the next 12-24 hours; Life prediction module: Based on the equipment's historical operating data and current status data, it predicts the equipment's remaining service life and generates maintenance recommendations based on the prediction results; Model optimization module: used to optimize the life prediction model.

2. An industrial equipment early warning system according to claim 1, characterized in that: In the multimodal sensor network module, there are multi-parameter sensors, which dynamically adjust the sampling frequency using the adaptive sampling method according to the operating status of the equipment. for: in, represents the information entropy of the sensor data at the current moment, and Represent the minimum and maximum information entropy of historical data, and represent the minimum and maximum sampling frequencies respectively, Indicates time.

3. The industrial equipment early warning system according to claim 1, characterized in that: In the feature extraction module, a multi-scale feature extraction method is used to extract features from multi-source data. The time domain feature calculation includes statistical features such as mean, variance, kurtosis, and skewness. Frequency domain features are calculated by Fourier transform to calculate power spectrum density and frequency band energy ratio; time-frequency features are extracted by wavelet transform and empirical mode decomposition, and feature vector The formula is: in, Represents the time domain feature vector, which is composed of time domain statistical features. is the mean, is the variance, is the skewness coefficient, is the kurtosis coefficient, is the peak factor, is the form factor; represents the frequency domain eigenvector, is the power spectral density function, They are frequency boundary value 1, frequency boundary value 2, frequency boundary value 3, and frequency boundary value 4 respectively. is the probability density function, is the center frequency, is the frequency bandwidth; represents the time-frequency domain eigenvector, is the wavelet packet energy, is the number of decomposition layers, For the Intrinsic mode functions; Represents the depth vector feature.

4. The industrial equipment early warning system according to claim 1, characterized in that: In the feature fusion module, the Dempster-Shafer evidence theory is used to assign weights according to the reliability of different sensor data. For fault type The basic probability distribution is , then the comprehensive fusion result for: in, represents the conflict factor, is an empty set, For evidence collection, is the set intersection operator.

5. The industrial equipment early warning system according to claim 1, characterized in that: In the health status assessment module, a hybrid model combining deep neural network and survival analysis is used to achieve real-time assessment of the health status of the equipment. The deep neural network includes an input layer, three hidden layers and an output layer structure, and calculates the equipment health index through forward propagation. Output of the layer Expressed as: in, and Respectively The layer's weight matrix and bias vector, is the ReLU activation function, defined as , Output 0 when Time output , which can alleviate the vanishing gradient and speed up the convergence; The health vector output by the neural network is used as the covariate of the Cox proportional hazard model and trained with the same loss function as the neural network. The failure risk function of the equipment is Defined as: in, is the benchmark risk function, It is an influencing factor. is the corresponding regression coefficient; the reliability function of the equipment and remaining life expectancy Can be expressed as: in, is the reliability function, which indicates the equipment The probability of normal operation within the system; is the risk failure rate function, which represents the conditional probability of failure occurring in unit time; For time The reliability value at time The probability of normal operation at all times; The average remaining life, which means the equipment has been running The expected remaining useful life after time.

6. The industrial equipment early warning system according to claim 1, characterized in that: The lifespan prediction model is based on a hybrid architecture of LSTM model and Transformer model; The LSTM model captures time series features through the following recursive calculation: in, 、 、 They are forget gate, input gate and output gate respectively. is the cell state, is the candidate cell state, is hidden state, is the sigmoid function, is the input weight matrix, is the recursive weight matrix, are bias vectors, For time The input vector, and They are tanh function one and tanh function two, is the element-wise multiplication operator; The Transformer model models long-term dependencies through a multi-head self-attention mechanism. The self-attention is calculated as follows: in, 、 、 are query, key, and value matrices respectively, is the dimension of the key. Multi-head attention is calculated by the following formula: in, represents the multi-head attention mechanism function, represents the connection function, is the number of attention heads, is the output linear transformation matrix, For the The parameter matrix of the attention head. Both the LSTM model and the Transformer model adopt a cascade-parallel hybrid architecture.

7. An industrial equipment early warning system according to claim 6, characterized in that: The cascade-parallel hybrid architecture is represented as: Cascade: LSTM first captures short-term / local temporal features and outputs hidden sequences ; Parallel: Hidden Sequence After linear mapping, it is fed into the Transformer multi-head self-attention layer as query / key / value to extract long-term dependencies; Concatenation: The final hidden state of the LSTM model Global representation of the Transformer model Connect to feature , for subsequent remaining useful life (RUL) distribution estimation. in this way, The matrix actually shares the weight mapping with the LSTM output, and the variables are correlated rather than independent. Then, the Monte Carlo method is used for sampling prediction, and the final output is the probability distribution of the remaining service life of the equipment. in, Indicates the remaining service life of the device. represents the expected value operator, Indicates the equipment failure time, Indicates the current time, Indicates the historical status data of the device. is the conditional survival function, is the integration variable.

8. The industrial equipment early warning system according to claim 1, characterized in that: The model optimization module uses a federated learning algorithm to achieve knowledge sharing and model optimization among multiple devices for the life prediction model. The objective function of federated learning is defined as: in, Represents the global objective function, which represents the optimization goal of the entire federated learning system; is the model parameter vector, which represents the neural network weights that need to be optimized; Indicates the total number of devices participating in federated learning, Representation device The amount of data, Indicates the total amount of data, Representation device dataset.

9. An industrial equipment early warning system according to claim 8, characterized in that: The federated learning optimization algorithm includes the following steps: Initialize the global model parameters to , for each round of communication : Select a subset of devices ; Each selected device Compute local updates: ;in, is the global model parameter of the current round, For equipment The updated local model parameters, is the learning rate, For equipment Local objective function exist The gradient at Perform aggregate updates: ;in, is the data volume of device j, used for weighted averaging; Differential privacy protection is achieved by adding noise. The updated parameters are: in, is the updated global model parameter; is the normally distributed noise, is the noise scale, is the gradient clipping limit, is the identity matrix; privacy budget and noise scale The relationship is: in, is the number of communication rounds, is the failure probability, is a constant.

10. An industrial equipment early warning method, characterized in that: include: 1) Use multimodal sensor networks to collect equipment operation data in real time and pre-process the equipment operation data; 2) Use the feature extraction module to extract features from multi-source equipment operation data, and then use the data fusion module to fuse the extracted features to generate the equipment health status feature vector; 3) Use the health status assessment module to analyze feature vectors, evaluate the current health status of the device, and predict the change trend in the next 12-24 hours; 4) Based on the equipment's historical operating data and current status data, the life prediction module is used to predict the equipment's remaining service life and generate maintenance recommendations based on the prediction results; 5) Use the model optimization module to optimize the life prediction model.

Images