Mechanical equipment multi-state selective maintenance method based on deep learning

Abstract

The invention discloses a mechanical equipment multi-state selective maintenance method based on deep learning, and relates to the technical field of mechanical equipment intelligent prediction maintenance, and the method comprises the following steps: deploying a sensor at mechanical equipment to collect operation state data in real time, collecting maintenance operation logs and historical fault diagnosis record text data, and storing the collected maintenance operation logs and historical fault diagnosis record text data; generating a multi-source data stream through timestamp alignment, and preprocessing the multi-source data stream to obtain a preprocessed multi-modal data set; according to the method, the multi-modal data collected for the mechanical equipment has comprehensiveness and availability of data sources, the utilization rate of heterogeneous data is improved, the multi-modal feature data is fused by using the cross-modal association mapping model, hidden information can be mined, the invisible state can be recognized, the equipment operation condition is clearly mastered, and the method is suitable for popularization and application. And through combination of the STCN and the Weibull-deep learning hybrid model, the residual life of the mechanical equipment can be dynamically and accurately predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent predictive maintenance of mechanical equipment, and particularly to a selective maintenance method for multiple states of mechanical equipment based on deep learning. Background Technique

[0002] In industrial production, the stable operation of mechanical equipment is crucial for ensuring production efficiency, product quality, and the economic benefits of enterprises. Maintaining mechanical equipment is an important guarantee for ensuring the reliability and stability of mechanical equipment. Currently, maintaining mechanical equipment usually overly relies on empirical thresholds, regular inspections, and after-fact maintenance strategies, making it difficult to consider the actual operating conditions of mechanical equipment. Especially when facing multiple states of mechanical equipment in complex industrial scenarios, predictive maintenance is implemented based on sensor-based condition monitoring and traditional fault diagnosis methods, which have significant limitations.

[0003] For example, on the one hand, sensors collect the operating state data of mechanical equipment, and heterogeneous data such as the recording of maintenance logs and the diagnosis of historical fault types are processed independently, without establishing cross-modal correlation mappings, resulting in the fragmentation of multi-modal data and the inability to extract the hidden operating information and laws of mechanical equipment in the data, reducing the accuracy of multi-state analysis and maintenance decisions. On the other hand, the existing behavior patterns of regular maintenance and after-fact maintenance by operators are in an invisible state and cannot be flexibly adjusted according to the actual state of the equipment, easily causing problems of over-maintenance or under-maintenance, leading to an increase in maintenance costs and a decrease in equipment reliability.

[0004] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and thus it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] The object of the present invention is to provide a selective maintenance method for multiple states of mechanical equipment based on deep learning. The present invention solves the problems in the above background art by establishing a cross-modal correlation mapping model to fuse multi-modal data and identify and analyze the invisible state of the operation of mechanical equipment.

[0006] To achieve the above object, the present invention provides the following technical solution: A selective maintenance method for multiple states of mechanical equipment based on deep learning, comprising the following steps:

[0007] S1. Deploy sensors on the mechanical equipment to collect operation state data in real time, collect text data such as maintenance operation logs and historical fault diagnosis records, generate multi-source data streams through timestamp alignment, and preprocess the multi-source data streams to obtain a preprocessed multi-modal data set;

[0008] S2. Extract features from the multi-modal dataset, and use a deep learning network to construct a cross-modal association mapping model for multi-modal feature data fusion. By learning the internal relationships between different modal feature data, the operation information and rules of the mechanical equipment hidden in the data are mined, and the invisible state is identified and evaluated;

[0009] S3. Build a multi-state degradation model for the mechanical equipment using sensors with a spatio-temporal convolutional network STCN. The multi-state degradation model uses 3D convolution to extract the spatial distribution features of multi-sensors and introduces a bidirectional GRU to capture the temporal dependencies of the degradation trend in space and time, and then outputs degradation indicators. Based on the Weibull-deep learning hybrid model, dynamically predict the remaining life of the mechanical equipment;

[0010] S4. Develop a dynamic maintenance strategy for the selective maintenance of the multi-states of the mechanical equipment. Establish a maintenance cost-benefit quantification model by defining a multi-objective reward function, combine the proximal policy optimization PPO algorithm, and output executable maintenance actions according to the input state features of the mechanical equipment. Set up an elastic maintenance mechanism and optimize the maintenance strategy;

[0011] S5. Real-time monitor the operation status of the mechanical equipment and the information on the execution of dynamic maintenance decisions, so as to issue a warning in time when the abnormal status of the mechanical equipment is found, and feedback it to the cross-modal association mapping model, the multi-state degradation model and the maintenance cost-benefit quantification model for incremental learning and optimization adjustment.

[0012] Optionally, the generation steps of the source data stream are as follows:

[0013] Deploy multi-source sensors on the mechanical equipment to collect the operation status data of the mechanical equipment in real time and label it as D sensors , and at the same time collect the text data of the maintenance operation logs and historical fault diagnosis records of the mechanical equipment and label it as D text ;

[0014] Use the clock synchronization IEEE1588PTP protocol to label each piece of operation status data D sensors and text data D text with the corresponding timestamp;

[0015] Adopt the dynamic time warping DTW algorithm to align the timestamps of the operation status data D sensors and text data D text for data fusion, align the timestamps to calculate the optimal path, and the calculation formula of the optimal path is In the formula, D(i,j) represents the cumulative distance from time t i to time t j , D sensors (t i ) represents the operation status data Dsensors The value at time t i , D text (t j ) is represented as text data D text The value at time t j is represented as the Euclidean distance;

[0016] For the missing period data of the operating state data D sensors , cubic spline interpolation is used for data filling, and the calculation formula for interpolation filling is and t i ≤t≤t i+1 , where represents the data value at time t after interpolation of the operating state data D sensors , a, b, c, and d represent the coefficients of the cubic spline interpolation obtained by solving the adjacent 4-point data, and t i , t i+1 represent the start time and end time of the interpolation interval respectively;

[0017] Generate a multi-source data stream in a unified time series format, calibrated as D Multi-source , then the expression of the multi-source data stream is D Multi-source ={D1, D2,..., D k}}, D1={D sensors (t1), D text (t1), t1},... D k ={D sensors (t k ), D text (t k ), t k}}, where D k represents the new data with timestamp alignment of the kth operating state data D sensors and the text data D text .

[0018] Optionally, the acquisition logic of the multi-modal data set is as follows:

[0019] The preprocessing steps of the operating state data D sensors include using improved wavelet threshold denoising, outlier detection and correction based on the improved Z-Score method, filling missing values with the spatio-temporal KNN algorithm, and normalization processing. Among them, the adaptive threshold calculation formula for improved wavelet threshold denoising is and E n =∑|W n，m (D sensors )| 2 , n = 5, where λ n represents the adaptive threshold of the nth layer, σn is denoted as the standard deviation of noise at the n-th layer, N is denoted as the length of the signal, E n is denoted as the energy signal at the n-th layer, W n，m (D sensors ) is denoted as the operation status data D sensors the m-th wavelet coefficient at the n-th layer after being processed by the improved soft threshold function;

[0020] The calculation formula for outlier detection is and In the formula, z is denoted as the improved Z-Score value, is denoted as the operation status data D sensors the median of, MAD is denoted as the median absolute deviation of the operation status data D sensors the median absolute deviation of, median is denoted as the median function;

[0021] The expression of the spatio-temporal KNN algorithm is and X = 5, in the formula, is denoted as the operation status data D sensors the missing value to be filled in, x and X respectively denote the x-th neighbor, the number of X neighbors selected in the spatio-temporal KNN algorithm, ω x is denoted as the weight of the x-th neighbor, D sensors,x is denoted as the operation status data value of the x-th neighbor, α and β respectively denote the weight coefficients corresponding to the time and space distances, Δt and Δs respectively denote the time interval and the space interval;

[0022] The calculation formula for normalization is In the formula, D' sensors is denoted as the operation status data D sensors the data value after normalization, is denoted as the operation status data D sensors the mean of, is denoted as the operation status data D sensors the standard deviation of;

[0023] Text data D text The preprocessing steps of include tokenization and part-of-speech tagging using the BiLSTM-CRF model, and named entity recognition processing by rule matching. Among them, the loss function expression of the BiLSTM-CRF model is and In the formula, is denoted as the loss function value of the BiLSTM-CRF model, P(D' text |D text ) is denoted as the sequence annotation probability of the text data D text the sequence annotation probability of, D' textDenoted as text data D text The output label sequence, γ is denoted as the regularization coefficient, θ is denoted as the BiLSTM-CRF model parameters, |||| 2 Denoted as the square of the Euclidean distance, S(D' text |D text ) is denoted as the label score output by BiLSTM;

[0024] Using the operating state data D sensors The preprocessed normalized data D' sensors and the text data D text The preprocessed output label sequence D' text Construct a multimodal dataset, calibrated as D Multi-modalt , and D Multi-modalt ={D' sensors , D' text}, and store the data.

[0025] Optionally, the feature extraction steps of the multimodal dataset are as follows:

[0026] Input the normalized data D' sensors to the convolutional neural network CNN for convolution operation, extract the time-frequency domain features, and change the feature map dimension, output the time-frequency domain features, where the calculation formula of the CNN convolution operation is D' sensors ∈R 128×128×1 , W∈R L×L , and l1, l2∈L, L = 3, where F out (u, v) represents the value of the convolution output feature map at the position (u, v), L represents the side length of the convolution kernel, W(l1, l2) represents the weight of the convolution kernel at the position (l1, l2), D' sensors (u + l1, v + l2) represents the value of the input normalized data D' sensors feature map at (u, v), B0 represents the bias term of the CNN;

[0027] The expression for changing the feature map dimension is: input Output h vib , and h vib ∈R 256 , h vib represents the output time-frequency feature vector;

[0028] Input the label sequence D' text to the word embedding BERT model for training, generate the text feature vector using the attention pooling layer, and compress the text feature to 256 dimensions through the fully connected layer, output the text feature, where the expression for the BERT model word embedding is E text =BERT(D'text ) ∈ R M×768 , where E text represents the text vector matrix obtained by training the BERT model on the label sequence D' text , and M represents the number of texts in the label sequence D' text ;

[0029] The semantic aggregation expression of the pooling layer is and where χ M represents the M-th attention weight, Softmax represents the Softmax function, represents the learnable parameter of the attention pooling layer, represents the bias term of the attention pooling layer, and h text represents the text feature vector generated by the pooling layer;

[0030] The compression and dimensionality reduction expression of the fully connected layer is h' text = ω f h text + B f , and ω f ∈ R 768×256 , where h' text represents the text feature vector output after dimensionality reduction, ω f represents the weight matrix of the fully connected layer, and B f represents the bias term of the fully connected layer.

[0031] Optionally, the construction steps of the cross-modal association mapping model are as follows:

[0032] Input the output time-frequency domain feature vector h vib and the output text feature vector h' text into the cross-modal attention fusion model. After key-value projection and attention weight calculation processing, feature fusion is obtained. Among them, the expression of key-value projection is Q = W Q h vib , K = W K h' text , V = W V h' text , and W Q , W K , W V ∈ R 256×256 , where Q, K, and V respectively represent the query vector, key vector, and value vector, and W Q , W K , W V respectively represent the corresponding query vector, key vector, and value vector projection matrices;

[0033] The calculation formula for the attention weight is In the formula, A represents the attention weight matrix, and T represents the transpose;

[0034] The calculation formula for feature fusion is h fusion = h vib + A·V, and A·V ∈ R 256 , where h fusion represents the time-frequency domain feature vector h vib and the text feature vector h' text the fused feature vector;

[0035] Input the fused feature h fusion into the fully connected layer of the dual-path network, and define the loss function and the joint optimization total loss function to identify the hidden state. Among them, the expression of the fully connected layer is Y hidden = sigmoid(ω2·ReLU(ω1·h fusion + b1)+ b2), and ω1 ∈ R 256×128 、ω2 ∈ R 128×C , where Y hidden represents the output of the dual-path fully connected layer, sigmoid represents the activation function, ω1 and ω2 respectively represent the weights of the dual-path fully connected layer, b1 and b2 respectively represent the bias terms of the dual-path fully connected layer, and C represents the number of hidden state categories;

[0036] The calculation formula for the loss function is In the formula, represents the multi-task binary cross-entropy loss function, represents the true label of the Cth hidden state of the output of the dual-path fully connected layer, represents the prediction probability of the cross-modal association mapping model for the true label of the Cth hidden state ;

[0037] The expression of the total loss function is In the formula, represents the total loss function for identifying the hidden state, represents the square of the L2 norm.

[0038] Optionally, the construction steps of the multi-state degradation model are as follows:

[0039] Input the operation state data D sensors collected by multi-source sensors during the operation of the mechanical equipment into the spatio-temporal convolutional network STCN;

[0040] Use 3D convolution to process the operation state data D sensors and extract the multi-sensor spatial distribution features, calibrated as F sensors(t), where t represents the time step;

[0041] A bidirectional GRU is introduced to capture the temporal dependence of the degradation trend. Among them, the expression for calculating the hidden state from the forward direction of the forward GRU is In the formula, represents the forward hidden state, and GRU represents the bidirectional GRU model. represents the hidden state at the previous time step t - 1; the expression for calculating the hidden state from the reverse direction of the backward GRU is In the formula, represents the forward and backward hidden state. represents the hidden state at the next time step t + 1; the expression for concatenating the hidden states in two directions to obtain the final output hidden state is In the formula, H t represents the final output hidden state;

[0042] Integrating spatial and temporal features, an index reflecting the degradation state of the mechanical equipment is output. Then, the expression for the output of the degradation state index is g(t) = Sigmoid(ω g H t + B g ). In the formula, g(t) represents the degradation index, Sigmoid represents the Sigmoid activation function, ω g represents the weight corresponding to the hidden state H t , and B g represents the bias term corresponding to the hidden state H t .

[0043] Optionally, the steps for constructing the multi-state degradation model are as follows:

[0044] Input the operation state data D of the mechanical equipment collected by multi-source sensors during operation sensors into the spatio-temporal convolutional network STCN;

[0045] Use 3D convolution to process the operation state data D sensors to extract the spatial distribution features of multi-sensors, calibrated as F sensors (t), where t represents the time step;

[0046] A bidirectional GRU is introduced to capture the temporal dependence of the degradation trend. Among them, the expression for calculating the hidden state from the forward direction of the forward GRU is In the formula, represents the forward hidden state, and GRU represents the bidirectional GRU model. represents the hidden state at the previous time step t - 1; the expression for calculating the hidden state from the reverse direction of the backward GRU is In the formula, Expressed as the hidden state before and after, Expressed as the hidden state at the next time step t+1; the hidden states in both directions are concatenated to form the final output hidden state expression as In the formula, H t Expressed as the final output hidden state;

[0047] Integrating spatial and temporal features, an index reflecting the degradation state of the mechanical equipment is output, and the index output expression of the degradation state is g(t) = Sigmoid(ω g H t +B g ), where g(t) represents the degradation index, Sigmoid represents the Sigmoid activation function, ω g Represents the weight corresponding to the hidden state H t and B g Represents the bias term corresponding to the hidden state H t .

[0048] Optionally, the selective maintenance system for multi-state mechanical equipment based on deep learning includes: a multi-source heterogeneous data acquisition and preprocessing module, which collects and preprocesses the dispersed and heterogeneous multi-source data on the mechanical equipment, and integrates it into a multi-modal data set with a unified time stamp to provide a high-quality data basis for subsequent modules for data analysis;

[0049] A cross-modal feature fusion and implicit state mining module, which learns the internal relationship between different modal feature data through feature extraction and fusion of the multi-modal data set, and then mines the operation information and rules of the mechanical equipment hidden in the multi-modal data, and identifies and evaluates the invisible state of the equipment;

[0050] A mechanical equipment life prediction module, which dynamically predicts the remaining life of the mechanical equipment by establishing a multi-state degradation model and using a Weibull-deep learning hybrid model, analyzes the equipment operation state and degradation trend, and accurately predicts the remaining life of the mechanical equipment;

[0051] A dynamic maintenance optimization module, which comprehensively considers maintenance costs and benefits to establish a maintenance cost-benefit quantification model, and sets up an elastic maintenance mechanism to formulate an optimal dynamic maintenance strategy for the multi-state selective maintenance of mechanical equipment;

[0052] A monitoring warning and incremental learning module, which monitors the operation state of the mechanical equipment and the information on the execution of dynamic maintenance decisions in real time, discovers abnormal situations in equipment operation in a timely manner, and continuously optimizes the models in other modules through incremental learning.

[0053] A computer device, comprising: a memory and a processor; the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned selective maintenance method for multiple states of mechanical equipment based on deep learning are implemented.

[0054] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned selective maintenance method for multiple states of mechanical equipment based on deep learning are implemented.

[0055] In the above technical solution, the technical effects and advantages provided by the present invention:

[0056] The present invention has the comprehensiveness and availability of data sources for the multi-modal data collected from mechanical equipment, improves the utilization rate of heterogeneous data, and uses a cross-modal association mapping model to fuse multi-modal feature data, which can not only mine hidden information, but also identify invisible states, clearly master the operating conditions of the equipment, and through the spatio-temporal convolutional network STCN combined with the Weibull-deep learning hybrid model, can dynamically and accurately predict the remaining life of mechanical equipment. By adopting a multi-objective reward function to establish a maintenance cost-benefit quantification model, combined with the proximal policy optimization PPO algorithm to formulate a dynamic maintenance strategy, and setting up an elastic maintenance mechanism, the balance between maintenance cost and benefit is achieved, effectively improving maintenance efficiency and economy, and performing incremental learning and optimization adjustment on relevant models, enabling the system to continuously adapt to equipment changes, continuously improving the accuracy and reliability of maintenance, reducing equipment failure risks, and ensuring the stable operation of mechanical equipment. Description of the Drawings

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0058] Figure 1 It is a flowchart of the selective maintenance method for multiple states of mechanical equipment based on deep learning of the present invention.

[0059] Figure 2 It is a schematic diagram of the modules of the selective maintenance system for multiple states of mechanical equipment based on deep learning of the present invention. Detailed Embodiments

[0060] Now, the exemplary embodiments will be described more fully with reference to the accompanying drawings. However, the exemplary embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these exemplary embodiments are provided so that the present disclosure will be more complete and comprehensive, and the concept of the exemplary embodiments will be fully conveyed to those skilled in the art.

[0061] Example 1

[0062] The present invention provides a selective maintenance method for multiple states of mechanical equipment based on deep learning as shown in Figure 1 and includes the following steps:

[0063] S1. Deploy sensors on the mechanical equipment to collect operation status data in real time, collect text data of maintenance operation logs and historical fault diagnosis records, generate a multi-source data stream through timestamp alignment, and preprocess the multi-source data stream to obtain a preprocessed multi-modal data set. Among them, the operation status data is subjected to cleaning, denoising, missing value and normalization processing, and the text data is subjected to word segmentation, part-of-speech tagging and named entity recognition processing;

[0064] Specifically, the generation steps of the multi-source data stream are as follows:

[0065] Deploy multi-source sensors on the mechanical equipment, including vibration sensors, temperature sensors, pressure sensors, acoustic emission sensors, etc., to collect the operation status data of the mechanical equipment in real time, and calibrate it as D sensors , and at the same time collect the text data of the maintenance operation logs and historical fault diagnosis records of the mechanical equipment, and calibrate it as D text ;

[0066] Use the clock synchronization IEEE1588PTP protocol to label each piece of operation status data D sensors and text data D text collected with the corresponding timestamp to ensure that the data has information in the time dimension;

[0067] Adopt the dynamic time warping DTW algorithm to align the operation status data D sensors and text data D text for data fusion by aligning the timestamps, and calculate the optimal path by aligning the timestamps. The calculation formula of the optimal path is In the formula, D(i,j) represents the cumulative distance from time t i to time t j , D sensors (t i ) represents the value of the operation status data D sensors at time t i , D text (t j ) represents the value of the text data D text at time t j , |||| represents the Euclidean distance;

[0068] For the operation status data D sensorsFor the missing period data, cubic spline interpolation is used to fill the data. The calculation formula for interpolation filling is and t i ≤t≤t i+1 , where in the formula represents the operating status data D sensors the data value at time t after interpolation. a, b, c, and d represent the coefficients of the cubic spline interpolation obtained from the adjacent 4-point data. t i and t i+1 represent the start time and end time of the interpolation interval respectively;

[0069] Generate a multi-source data stream in a unified time series format, calibrated as D Multi-source , then the expression of the multi-source data stream is D Multi-source ={D1, D2,..., D k}}, D1={D sensors (t1), D text (t1), t1},... D k ={D sensors (t k ), D text (t k ), t k}, where D k represents the new data with the kth operating status data D sensors and the text data D text time stamps aligned.

[0070] Specifically, the acquisition logic of the multi-modal data set is as follows:

[0071] The preprocessing steps of the operating status data D sensors include using improved wavelet threshold denoising, outlier detection and correction based on the improved Z-Score method, filling missing values with the spatio-temporal KNN algorithm, and normalization processing. Among them, the adaptive threshold calculation formula for improved wavelet threshold denoising is and E n =∑|W n，m (D sensors )| 2 , n = 5, where λ n represents the adaptive threshold of the nth layer, σ n represents the noise standard deviation of the nth layer, N represents the length of the signal, E n represents the energy signal of the nth layer, and W n，m (D sensors ) represents the mth wavelet coefficient of the nth layer of the operating status data D sensors after being processed by the improved soft threshold function;

[0072] The calculation formula for outlier detection is and In the formula, z represents the improved Z-Score value, represents the median of the operating status data D sensors and MAD represents the median absolute deviation of the operating status data D sensors and median represents the median function;

[0073] The expression of the spatio-temporal KNN algorithm is and X = 5. In the formula, represents the missing value to be filled in the operating status data D sensors , x and X respectively represent the xth neighbor and the number of X neighbors selected in the spatio-temporal KNN algorithm, and ω x represents the weight of the xth neighbor, D sensors,x represents the operating status data value of the xth neighbor, and α and β respectively represent the weight coefficients corresponding to the time and space distances, and Δt and Δs respectively represent the time interval and the space interval;

[0074] The calculation formula for normalization is In the formula, D' sensors represents the data value of the operating status data D sensors after normalization, μ Dsensors represents the mean of the operating status data D sensors , and σ Dsensors represents the standard deviation of the operating status data D sensors ;

[0075] The preprocessing steps of the text data D text include word segmentation and part-of-speech tagging using the BiLSTM-CRF model, and named entity recognition processing by rule matching. Among them, the loss function of the BiLSTM-CRF model is expressed as and In the formula, represents the loss function value of the BiLSTM-CRF model, P(D' text |D text ) represents the sequence labeling probability of the text data D text , D' text represents the output label sequence of the text data D text , γ represents the regularization coefficient, θ represents the BiLSTM-CRF model parameters, |||| 2 represents the square of the Euclidean distance, and S(D' text |D text ) represents the label score output by BiLSTM;

[0076] Using the operating status data D sensors The preprocessed normalized data D' sensors and the text data D text The preprocessed output label sequence D' text Construct a multi-modal data set, labeled as D Multi-modalt and D Multi-modalt ={D' sensors , D' text}, and perform data storage.

[0077] Furthermore, by deploying sensors on mechanical equipment to collect operating status data in real time and collecting relevant text data, a multi-source data stream is generated using the PTP protocol and the DTW algorithm, achieving microsecond-level alignment of the multi-source data stream. Then, different types of data are preprocessed separately. By using the improved wavelet threshold denoising to improve the signal-to-noise ratio, accurately detect outliers, and normalize the data, a preprocessed multi-modal data set is obtained, solving the problem of separated features and fragmented features in the traditional heterogeneous data processing process. Furthermore, it realizes that the originally scattered multi-modal data can be associated, and has a unified format and scale, providing more accurate and comprehensive data support for subsequent multi-state analysis and maintenance decision-making.

[0078] S2. Extract features from the multi-modal data set, and use a deep learning network to construct a cross-modal association mapping model for multi-modal feature data fusion. By learning the internal relationships between different modal feature data, the operating information and laws of mechanical equipment hidden in the data are mined, and the invisible state is identified and evaluated. Among them, feature extraction is to use a convolutional neural network CNN to extract time-frequency domain features from the normalized data, and use the word embedding BERT technology to convert the text data into vectors;

[0079] Specifically, the feature extraction steps of the multi-modal data set are as follows:

[0080] Input the normalized data D' sensors into the convolutional neural network CNN for convolution operation, extract time-frequency domain features, and change the feature mapping dimension, and output time-frequency domain features. Among them, the calculation formula of the CNN convolution operation is D' sensors ∈R 128×128×1 , W∈R L×L , and l1, l2∈L, L = 3. In the formula, F out (u, v) represents the value of the convolution output feature map at the position (u, v), L represents the side length of the convolution kernel, W(l1, l2) represents the weight of the convolution kernel at the position (l1, l2), and D' sensors (u + l1, v + l2) represents the input normalized data D' sensorsThe value of the feature map at (u, v), where B0 represents the bias term of the CNN;

[0081] The expression for the change in the dimension of the feature map is: input Output h vib , and h vib ∈R 256 , h vib represents the time-frequency feature vector of the output;

[0082] Input label sequence D' text is fed into the word embedding BERT model for training. The attention pooling layer is used to generate the text feature vector, and the fully connected layer compresses the text feature to 256 dimensions to output the text feature. Among them, the expression for the word embedding of the BERT model is E text = BERT(D' text ) ∈R M×768 , where E text represents the text vector matrix obtained by the BERT model training on the label sequence D' text , and M represents the number of texts in the label sequence D' text ;

[0083] The semantic aggregation expression of the pooling layer is and where χ M represents the M-th attention weight, Softmax represents the Softmax function, represents the learnable parameter of the attention pooling layer, represents the bias term of the attention pooling layer, and h text represents the text feature vector generated by the pooling layer;

[0084] The compression and dimensionality reduction expression of the fully connected layer is h' text = ω f h text + B f , and ω f ∈R 768×256 , where h' text represents the text feature vector of the output after dimensionality reduction, ω f represents the weight matrix of the fully connected layer, and B f represents the bias term of the fully connected layer.

[0085] Specifically, the construction steps of the cross-modal association mapping model are as follows:

[0086] The output time-frequency domain feature vector h vib and the output text feature vector h' textInput into the cross-modal attention fusion model, and through key-value projection and attention weight calculation processing, feature fusion is obtained. Among them, the expression of key-value projection is Q = W Q h vib 、K = W K h' text 、V = W V h' text ,and W Q 、W K 、W V ∈R 256×256 ,wherein, Q, K, and V respectively represent the query vector, key vector, and value vector, and W Q 、W K 、W V respectively represent the corresponding query vector, key vector, and value vector projection matrices;

[0087] The calculation formula of the attention weight is wherein, A represents the attention weight matrix, and T represents the transpose;

[0088] The calculation formula of feature fusion is h fusion = h vib + A·V, and A·V∈R 256 ,wherein, h fusion represents the feature vector after the fusion of the time-frequency domain feature vector h vib and the text feature vector h' text ;

[0089] Input the fusion feature h fusion into the double-path network fully connected layer, and define the loss function and the joint optimization total loss function to identify the hidden state. Among them, the expression of the fully connected layer is Y hidden = sigmoid(ω2·ReLU(ω1·h fusion + b1)+ b2), and ω1∈R 256×128 、ω2∈R 128×C ,wherein, Y hidden represents the output of the double-path fully connected layer, sigmoid represents the activation function, ω1 and ω2 respectively represent the weights of the double-path fully connected layer, b1 and b2 respectively represent the bias terms of the double-path fully connected layer, and C represents the number of hidden state categories;

[0090] The calculation formula of the loss function is wherein, represents the multi-task binary cross-entropy loss function, represents the true label of the Cth hidden state of the output of the double-path fully connected layer, represents the true label of the Cth hidden state by the cross-modal correlation mapping model Prediction probability;

[0091] The total loss function expression is In the formula, represents the total loss function for identifying the hidden state, represents the square of the L2 norm.

[0092] Furthermore, by extracting features from the multi-modal dataset, when using the convolutional neural network CNN to extract time-frequency domain features for the normalized data, using the word embedding BERT technology to convert text data into vectors, and using the cross-modal attention mechanism to achieve deep feature association, the operating information of mechanical equipment can be comprehensively and deeply mined, a cross-modal association mapping model can be established, the internal relationship between different modal feature data can be learned, and the heterogeneity of CNN time-frequency features and BERT text features can be data-fused using a unified 256-dimensional embedding space, which can identify and evaluate the hidden state of mechanical equipment, provide a key basis for the operation and maintenance of the equipment, and identify the hidden state of mechanical equipment, improve the accuracy of identifying and evaluating the hidden state of mechanical equipment, help to discover potential problems in advance, reduce the equipment failure risk, and ensure the stable operation of the equipment.

[0093] S3. By using the spatio-temporal convolutional network STCN for mechanical equipment sensors to construct a multi-state degradation model, the multi-state degradation model uses 3D convolution in space and time respectively to extract the spatial distribution features of multi-sensors and introduces bidirectional GRU to capture the temporal dependence of the degradation trend and output the degradation index, and dynamically predicts the remaining life of mechanical equipment based on the Weibull-deep learning hybrid model;

[0094] Specifically, the construction steps of the multi-state degradation model are as follows:

[0095] Input the operation state data D of the mechanical equipment collected by multi-source sensors during the operation process sensors into the spatio-temporal convolutional network STCN;

[0096] Use 3D convolution to process the operation state data D sensors to extract the spatial distribution features of multi-sensors and calibrate them as F sensors (t), where t represents the time step;

[0097] Introduce bidirectional GRU to capture the temporal dependence relationship of the degradation trend. Among them, the expression for the forward GRU to calculate the hidden state from the forward direction is In the formula, represents the forward hidden state, GRU represents the bidirectional GRU model, represents the hidden state at the previous t-1 time step; the expression for the backward GRU to calculate the hidden state from the backward direction is In the formula, represents the forward and backward hidden states, Denoted as the hidden state at the next time step t+1; the hidden states in two directions are concatenated, and the expression for the final output hidden state is In the formula, H t Denoted as the final output hidden state;

[0098] Integrating spatial and temporal features, an index reflecting the degradation state of the mechanical equipment is output. Then, the expression for the index output of the degradation state is g(t) = Sigmoid(ω g H t +B g ), where g(t) represents the degradation index, Sigmoid represents the Sigmoid activation function, ω g Denotes the weight corresponding to the hidden state H t and B g Denotes the bias term corresponding to the hidden state H t .

[0099] Specifically, the dynamic prediction logic of the remaining life of the mechanical equipment is as follows:

[0100] The mapping shape parameter δ and scale parameter ξ of the hidden state H t are output through STCN. Among them, the expression for the mapping shape parameter δ is δ(t) = ecp(ω δ H t +B δ ), and δ > 0. In the formula, δ(t) represents the change of the shape parameter of the Weibull distribution with time t, ecp represents the exponential function, ω δ Denotes the weight matrix that maps the hidden state variable H t to the shape parameter δ(t), and B δ Denotes the bias vector that maps the hidden state variable H t to the shape parameter c(t);

[0101] The expression for the scale parameter ξ is ξ(t) = ecp(ω ξ H t +B ξ ), and ξ > 0. In the formula, ξ(t) represents the change of the scale parameter of the Weibull distribution with time t, ω ξ Denotes the weight matrix that maps the hidden state variable H t to the scale parameter ξ(t), and B ξ Denotes the bias vector that maps the hidden state variable H t to the scale parameter ξ(t);

[0102] Based on the survival function of the Weibull distribution, the remaining useful life RUL of the mechanical equipment is calculated and dynamically iteratively predicted. Among them, the expression for the survival function is Wherein, R(t) represents the survival function of the Weibull distribution, which is used to represent the probability that the device is still operating normally at time t;

[0103] The expression for dynamic iterative prediction is Wherein, represents the predicted value of the degradation index g(t), STCN represents the spatio-temporal convolutional network model, and θ represents the STCN network parameters;

[0104] Taking the joint optimization of degradation index fitting and Weibull parameters as the loss function, optimizing the Weibull-deep learning hybrid model to dynamically predict the remaining life of mechanical equipment. Among them, the expression of the loss function is Wherein, represents the loss function for jointly optimizing the physical consistency of degradation index fitting and Weibull parameters, δ0 represents the physical prior value of the Weibull shape parameter, and ξ0 represents the physical prior value of the Weibull scale parameter.

[0105] Furthermore, a multi-state degradation model is constructed using the spatio-temporal convolutional network STCN to accurately characterize the multi-state degradation characteristics of mechanical equipment during operation, thereby precisely capturing the spatial and temporal characteristics in the multi-state degradation process of mechanical equipment, enhancing the understanding and grasp of the equipment degradation law, being able to describe and quantify the degradation process of mechanical equipment in detail, outputting degradation indicators, and further precisely capturing the spatial and temporal characteristics in the multi-state degradation process of mechanical equipment, enhancing the understanding and grasp of the equipment degradation law; by using the Weibull-deep learning hybrid model to dynamically predict the remaining life of mechanical equipment, the dynamic prediction of the remaining life of mechanical equipment is realized, providing a scientific basis for the maintenance and management of the equipment, significantly improving the accuracy and dynamics of the prediction of the remaining life of mechanical equipment, helping to reasonably arrange the equipment maintenance plan, reduce the maintenance cost, and improve the reliability and availability of the equipment.

[0106] S4. Develop a dynamic maintenance strategy for the selective maintenance of the multi-state of mechanical equipment. By defining a multi-objective reward function to establish a maintenance cost-benefit quantification model, combined with the proximal policy optimization PPO algorithm, output executable maintenance actions according to the input state characteristics of the mechanical equipment, set up an elastic maintenance mechanism and optimize the maintenance strategy;

[0107] Specifically, the steps for establishing the maintenance cost-benefit quantification model are as follows:

[0108] Determine multiple objectives of maintenance cost, equipment availability, and failure risk that need to be considered during the maintenance process of mechanical equipment;

[0109] Design corresponding reward functions for each objective, and this reward function will be used as the basis for evaluating the quality of the maintenance strategy;

[0110] Using the weighted summation method, the reward functions of multiple objectives are integrated into an objective reward function for constructing and maintaining a cost-benefit quantification model, which is used to evaluate the costs and risk benefits under different maintenance strategies;

[0111] Combined with the Proximal Policy Optimization (PPO) algorithm, the state characteristics of mechanical equipment are input into the PPO network, and the PPO loss function is defined to output the dynamic space of maintenance.

[0112] Specifically, the setting steps of the elastic maintenance mechanism are as follows:

[0113] Analyze and determine the elastic factors such as equipment state fluctuations, maintenance resource availability, and production plan changes that affect the maintenance decisions of mechanical equipment;

[0114] According to the elastic factors, formulate corresponding elastic rules for immediate maintenance, deferred maintenance, and partial maintenance, and dynamically adjust the maintenance cycle;

[0115] Integrate the elastic rules into the existing maintenance strategies, and set the trigger conditions for emergency maintenance and establish the equipment maintenance priorities, that is, by continuously optimizing the elastic rules, improve the adaptability and effectiveness of the maintenance strategies.

[0116] Furthermore, by defining a multi-objective reward function to establish a maintenance cost-benefit quantification model, combined with the Proximal Policy Optimization (PPO) algorithm, and outputting executable maintenance actions according to the state characteristics input by the mechanical equipment, the functions of accurately outputting maintenance actions for different equipment states, setting an elastic maintenance mechanism, and optimizing the maintenance strategy are realized. Furthermore, the effects of better meeting the multi-state maintenance requirements of mechanical equipment, scientifically balancing the maintenance costs and benefits, and optimizing the maintenance strategy to improve the equipment maintenance efficiency and reliability are achieved.

[0117] S5. Real-time monitor the operating status of the mechanical equipment and the information on the execution of dynamic maintenance decisions, so as to issue a warning in time when the abnormal state of the mechanical equipment is found, and feedback it to the cross-modal correlation mapping model, multi-state degradation model, and maintenance cost-benefit quantification model for incremental learning and optimization adjustment, which is used to continuously improve the performance and accuracy of the models.

[0118] Furthermore, by real-time monitoring the operating status of the mechanical equipment and the information on the execution of dynamic maintenance decisions and feedbacking it to the cross-modal correlation mapping model, multi-state degradation model, and maintenance cost-benefit quantification model for incremental learning and optimization adjustment, the effects of continuously improving the model performance, generalization, and accuracy are achieved, enabling the model to better adapt to various changes in the operation process of the mechanical equipment, providing more reliable support for dynamic maintenance decisions, thereby reducing the maintenance costs, improving the equipment reliability and production efficiency.

[0119] Embodiment 2

[0120] The present invention provides a selective maintenance system for multiple states of mechanical equipment based on deep learning as follows Figure 2 A multi-source heterogeneous data acquisition and preprocessing module, which acquires and preprocesses the dispersed and heterogeneous multi-source data on the mechanical equipment, and integrates them into a multi-modal data set with a unified timestamp, providing a high-quality data basis for subsequent modules for data analysis;

[0121] A cross-modal feature fusion and implicit state mining module, which learns the internal relationship between different modal feature data through feature extraction and fusion of the multi-modal data set, and then mines the operation information and rules of the mechanical equipment hidden in the multi-modal data, and identifies and evaluates the invisible state of the equipment, so as to accurately judge the actual operation status of the equipment;

[0122] A mechanical equipment life prediction module, which dynamically predicts the remaining life of the mechanical equipment by establishing a multi-state degradation model and using a Weibull-deep learning hybrid model, analyzes the operation state and degradation trend of the equipment, accurately predicts the remaining life of the mechanical equipment, and provides an important basis for maintenance decision-making;

[0123] A dynamic maintenance optimization module, which comprehensively considers the maintenance cost and benefit to establish a maintenance cost-benefit quantification model, and sets up an elastic maintenance mechanism to formulate an optimal dynamic maintenance strategy for the multi-state selective maintenance of the mechanical equipment, improving the maintenance efficiency and economy;

[0124] A monitoring, warning and incremental learning module, which monitors the operation status of the mechanical equipment and the information on the execution of the dynamic maintenance decision in real time, discovers abnormal situations in the equipment operation in time, ensures the safe and stable operation of the equipment, and continuously optimizes the models in other modules through incremental learning, improving the performance and accuracy of the system.

[0125] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to obtain a formula closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0126] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0127] It should be understood that in various embodiments of the present application, the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0128] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0129] As described above, the above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A selective maintenance method for multi-states of mechanical equipment based on deep learning, characterized in that, It includes the following steps: S1. Deploy sensors on the mechanical equipment to collect operation status data in real time, collect maintenance operation logs and historical fault diagnosis record text data, generate a multi-source data stream by aligning timestamps, and preprocess the multi-source data stream to obtain a preprocessed multi-modal data set; S2. Extract features from the multi-modal data set, and use a deep learning network to construct a cross-modal association mapping model for multi-modal feature data fusion. By learning the internal relationships between different modal feature data, excavate the operation information and rules of the mechanical equipment hidden in the data, and identify and evaluate the invisible state; S3. Construct a multi-state degradation model for the mechanical equipment using the spatio-temporal convolutional network STCN. The multi-state degradation model uses 3D convolution to extract the multi-sensor spatial distribution features in space and time respectively, and introduces a bidirectional GRU to capture the temporal dependence of the degradation trend and output the degradation index, and dynamically predict the remaining life of the mechanical equipment based on the Weibull-deep learning hybrid model; S4. Develop a dynamic maintenance strategy for the selective maintenance of the multi-states of the mechanical equipment. Establish a maintenance cost-benefit quantification model by defining a multi-objective reward function, combine the proximal policy optimization PPO algorithm, output executable maintenance actions according to the input state features of the mechanical equipment, set up an elastic maintenance mechanism and optimize the maintenance strategy; S5. Monitor the operation status of the mechanical equipment and the information on the execution of dynamic maintenance decisions in real time, so as to issue a warning in time when the abnormal state of the mechanical equipment is found, and feedback it to the cross-modal association mapping model, multi-state degradation model and maintenance cost-benefit quantification model for incremental learning and optimization adjustment.

2. The selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 1, characterized in that The generation steps of the source data stream are as follows: Multi-source sensors are deployed on mechanical equipment to collect real-time operating status data of the mechanical equipment and calibrate it as D sensors , and collect the text data of the maintenance operation log and historical fault diagnosis record of the mechanical equipment, and mark it as D text ; Use the clock synchronization IEEE1588PTP protocol to label each piece of operation status data D sensors and text data D text with the corresponding timestamp; Using the dynamic time warping (DTW) algorithm, align the running state data D sensors and the text data D text to perform data fusion on the timestamps, calculate the optimal path by aligning the timestamps, and the calculation formula for the optimal path is In the formula, D(i,j) represents the cumulative distance from time t i to time t j , D sensors (t i ) represents the value of the running state data D sensors at time t i , D text (t j ) represents the value of the text data D text at time t j , and || || represents the Euclidean distance; For the missing period data of the operating status data D sensors data filling is performed using cubic spline interpolation, and the calculation formula for interpolation filling is and t i ≤t≤t i+1 , where represents the data value of the operating status data D sensors after interpolation at time t, a, b, c, and d represent the coefficients of the cubic spline interpolation obtained by solving from adjacent 4-point data, t i , t i+1 represent the start time and end time of the interpolation interval respectively; Generate a multi-source data stream in a unified time series format, calibrated as D Multi-source , then the expression of the multi-source data stream is D Multi-source = {D1, D2, …, D k}, D1 = {D sensors (t1), D text (t1), t1}, … D k = {D sensors (t k ), D text (t k ), t k}, where D k represents the new data with the time stamps of the k-th running state data D sensors and the text data D text aligned.

3. The selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 2, characterized in that, The acquisition logic of the multi-modal data set is as follows: Operation status data D sensors The preprocessing steps of sensors include using improved wavelet threshold denoising, outlier detection and correction based on the improved Z-Score method, filling missing values with the spatio-temporal KNN algorithm, and normalization. Among them, the adaptive threshold calculation formula for improved wavelet threshold denoising is and E n = ∑|W n，m (D sensors )| 2 , n = 5. In the formula, λ n represents the adaptive threshold of the nth layer, σ n represents the noise standard deviation of the nth layer, N represents the length of the signal, and E n represents the energy signal of the nth layer. W n，m (D sensors ) represents the mth wavelet coefficient of the nth layer after the operation status data D sensors is processed by the improved soft threshold function; The calculation formula for outlier detection is and In the formula, z represents the improved Z-Score value, represents the operating status data D sensors of the median, and MAD represents the median absolute deviation of the operating status data D sensors of the median absolute deviation, and median represents the median function; The expression of the spatio-temporal KNN algorithm is and X = 5, where represents the missing value to be filled in the operation status data D sensors in, x and X respectively represent the x-th neighbor, and the number of X neighbors selected in the spatio-temporal KNN algorithm, ω x represents the weight of the x-th neighbor, D sensors,x represents the operation status data value of the x-th neighbor, α and β respectively represent the weight coefficients corresponding to the time and space distances, and Δt and Δs respectively represent the time interval and the space interval; The normalization calculation formula is In the formula, D' sensors represents the operation status data D sensors as the normalized data value, μ Dsensors represents the operation status data D sensors as the mean value, σ Dsensors represents the operation status data D sensors as the standard deviation; Text data D text The preprocessing steps of include tokenization and part-of-speech tagging using the BiLSTM-CRF model, and named entity recognition processing by rule matching. Among them, the loss function of the BiLSTM-CRF model is expressed as In the formula,[[]] denotes the loss function value of the BiLSTM-CRF model, P(D' text |D text ) denotes the sequence labeling probability of the text data D text , D' text denotes the output label sequence of the text data D text , γ denotes the regularization coefficient, θ denotes the BiLSTM-CRF model parameters, || || 2 denotes the square of the Euclidean distance, and S(D' text |D text ) denotes the label score output by the BiLSTM; Using the operating status data D sensors The preprocessed normalized data D' sensors and the text data D text The preprocessed output label sequence D' text Construct a multi-modal data set, labeled as D Multi-modalt and D Multi-modalt ={D' sensors , D' text}, and perform data storage.

4. The selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 3, characterized in that, The feature extraction steps of the multi-modal data set are as follows: Input normalized data D' sensors Go through convolutional operations in the convolutional neural network CNN to extract time-frequency domain features, change the dimension of the feature map, and output time-frequency domain features. Among them, the calculation formula for the CNN convolutional operation is D' sensors ∈R 128×128×1 , W ∈ R L×L , and l1, l2 ∈ L, L = 3. In the formula, F out (u, v) represents the value of the convolutional output feature map at the position (u, v), L represents the side length of the convolutional kernel, W(l1, l2) represents the weight of the convolutional kernel at the position (l1, l2), D' sensors (u + l1, v + l2) represents the value of the input normalized data D' sensors feature map at (u, v), and B0 represents the bias term of the CNN; The expression for the change in the feature mapping dimension is: input output h vib , and h vib ∈R 256 , h vib is represented as the time-frequency feature vector of the output; Input label sequence D' text is fed into the word embedding BERT model for training. The attention pooling layer is used to generate text feature vectors, and the fully connected layer compresses the text features to 256 dimensions to output text features. Among them, the expression of the word embedding of the BERT model is E text = BERT(D' text ) ∈ R M×768 , where E text represents the text vector matrix obtained by training the label sequence D' text by the BERT model, and M represents the number of texts in the label sequence D' text ; The semantic aggregation expression of the pooling layer is χ M = Softmax(ω R0 ·E text + B R0 ), and where χ M represents the M-th attention weight, Softmax represents the Softmax function, represents the learnable parameter of the attention pooling layer, represents the bias term of the attention pooling layer, and h text represents the text feature vector generated by the pooling layer; The compression and dimensionality reduction expression of the fully connected layer is h' text = ω f h text + B f , and ω f ∈ R 768×256 , where h' text represents the text feature vector of the output after dimensionality reduction, ω f represents the weight matrix of the fully connected layer, and B f represents the bias term of the fully connected layer.

5. The selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 4, characterized in that, The construction steps of the cross-modal association mapping model are as follows: The output time-frequency domain feature vector h vib and the output text feature vector h' text are input into the cross-modal attention fusion model. After key-value projection and attention weight calculation processing, feature fusion is obtained. Among them, the expression of key-value projection is Q = W Q h vib , K = W K h' text , V = W V h' text , and W Q , W K , W V ∈ R 256×256 . In the formula, Q, K, and V respectively represent the query vector, key vector, and value vector, and W Q , W K , W V respectively represent the corresponding query vector, key vector, and value vector projection matrices; The calculation formula of the attention weight is where A represents the attention weight matrix, and T represents the transpose; The calculation formula for feature fusion is h fusion = h vib + A·V, and A·V ∈ R 256 , where h fusion represents the time-frequency domain feature vector h vib and the text feature vector h' text the fused feature vector; Input the fused feature h fusion into the fully-connected layer of the dual-path network, and define the loss function and the combined optimization total loss function to identify the hidden state. Among them, the expression of the fully-connected layer is Y hidden = sigmoid(ω2·ReLU(ω1·h fusion + b1)+ b2), and ω1 ∈ R 256×128 、ω2 ∈ R 128×C , where Y hidden represents the output of the dual-path fully-connected layer, sigmoid represents the activation function, ω1 and ω2 respectively represent the weights of the dual-path fully-connected layer, b1 and b2 respectively represent the bias terms of the dual-path fully-connected layer, and C represents the number of hidden state categories; The calculation formula of the loss function is In the formula,[[]] represents the multi-task binary cross-entropy loss function,[[]] represents the true label of the hidden state of the C-th class in the output of the two-way fully connected layer,[[]] represents the predicted probability of the true label of the hidden state of the C-th class by the cross-modal correlation mapping model ; The total loss function expression is In the formula,[[]]END]] represents the total loss function for identifying the invisible state,[[]]END]] represents the square of the L2 norm.[[]]END]] 6. The selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 5, wherein The construction steps of the multi-state degradation model are as follows: Collect the operation status data D of mechanical equipment during operation by multi-source sensors sensors Input it into the spatio-temporal convolutional network STCN; Process the operating state data D using 3D convolution sensors to extract the multi-sensor spatial distribution features and calibrate them as F sensors (t), where t represents the time step; A bidirectional GRU is introduced to capture the temporal dependencies of the degradation trend. Among them, the expression for the forward GRU to calculate the hidden state from the forward direction is In the formula, is expressed as the forward hidden state, and GRU is expressed as the bidirectional GRU model. is expressed as the hidden state at the previous time step t - 1; The expression for the backward GRU to calculate the hidden state from the reverse calculation is In the formula,[[]] represents the forward and backward hidden states,[[]] represents the hidden state at the next time step t+1; the expression for concatenating the hidden states in both directions into the final output hidden state is In the formula, H t represents the final output hidden state. Integrating spatial and temporal features, an indicator reflecting the degradation state of mechanical equipment is output. The output expression of the degradation state indicator is g(t) = Sigmoid(ω g H t + B g ). In the formula, g(t) represents the degradation indicator, Sigmoid represents the Sigmoid activation function, ω g represents the weight corresponding to the hidden state H t , and B g represents the bias term corresponding to the hidden state H t .

7. A selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 6, characterized in that The construction steps of the multi-state degradation model are as follows: Collect the operating state data D of the mechanical equipment during operation by multi-source sensors sensors Input it into the spatio-temporal convolutional network STCN; Process the operating state data D using 3D convolution sensors to extract the multi-sensor spatial distribution features and calibrate them as F sensors (t), where t represents the time step; Introduce a bidirectional GRU to capture the temporal dependencies of the degradation trend. Among them, the expression for the forward GRU to calculate the hidden state from the forward direction is In the formula,[[]]END]] is expressed as the forward hidden state, and GRU is expressed as the bidirectional GRU model.[[]]END]] is expressed as the hidden state at the previous time step t-1;[[]]END]] The expression for the backward GRU to calculate the hidden state from the reverse calculation is In the formula,[[]] is expressed as the hidden states before and after,[[]] is expressed as the hidden state at the next time step t + 1; the expression for concatenating the hidden states in both directions into the final output hidden state is In the formula, H t is expressed as the final output hidden state; Integrating spatial and temporal features, an indicator reflecting the degradation state of mechanical equipment is output. The output expression of the degradation state indicator is g(t) = Sigmoid(ω g H t + B g ). In the formula, g(t) represents the degradation indicator, Sigmoid represents the Sigmoid activation function, ω g represents the weight corresponding to the hidden state H t , B g represents the bias term corresponding to the hidden state H t .

8. A selective maintenance method for multiple states of mechanical equipment based on deep learning according to claim 7, characterized in that, The selective maintenance system for multi-states of mechanical equipment based on deep learning includes: a multi-source heterogeneous data acquisition and preprocessing module, which collects and preprocesses the scattered and heterogeneous multi-source data on the mechanical equipment, and integrates them into a multi-modal data set with a unified timestamp, providing a high-quality data basis for subsequent modules for data analysis; A cross-modal feature fusion and hidden state mining module, which learns the internal relationships between different modal feature data by extracting and fusing the features of the multi-modal data set, and then excavates the operation information and rules of the mechanical equipment hidden in the multi-modal data, and identifies and evaluates the invisible state of the equipment; A mechanical equipment life prediction module, which dynamically predicts the remaining life of the mechanical equipment by establishing a multi-state degradation model and using a Weibull-deep learning hybrid model, analyzes the operation state and degradation trend of the equipment, and accurately predicts the remaining life of the mechanical equipment; A dynamic maintenance optimization module, which comprehensively considers maintenance costs and benefits to establish a maintenance cost-benefit quantification model, and sets up an elastic maintenance mechanism to formulate an optimal dynamic maintenance strategy for the selective maintenance of the multi-states of the mechanical equipment; The monitoring and early warning and incremental learning module monitors the operating status of mechanical equipment and the information on the implementation of dynamic maintenance decisions in real time, discovers abnormal conditions during equipment operation in a timely manner, and continuously optimizes the models in other modules through incremental learning.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of a selective maintenance method for multiple states of mechanical equipment based on deep learning according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of a selective maintenance method for multiple states of mechanical equipment based on deep learning according to any one of claims 1 to 8 are implemented.

Images