Remote device state diagnosis and maintenance optimization method and system based on big data

Abstract

The application discloses a kind of remote equipment state diagnosis and maintenance optimization method and system based on big data.The method is collected equipment operating parameter by distributed sensor, after data is standardized pretreated, key features are extracted using K-SVD algorithm to construct optimized dictionary and using tracking algorithm;Double-flow sparse sensing LSTM model is designed, its main branch processes the time sequence variation law of sparse representation, auxiliary flow branch analyzes the use mode of dictionary base vector, and the processing capacity for sparse data is enhanced by improved gate structure;Sliding window mechanism and autoregressive operation are used to realize equipment state diagnosis and prediction, and corresponding maintenance instructions are triggered according to the results.The application significantly improves the accuracy of equipment state monitoring through innovative double-flow architecture and improved gate mechanism, and realizes efficient and real-time state diagnosis relying on data compression and hierarchical transmission strategy of edge node, which provides reliable technical support for intelligent maintenance decision of industrial equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment state monitoring, and particularly relates to a remote equipment state diagnosis and maintenance optimization method and system based on big data. BACKGROUND

[0002] With the development of industrial equipment towards large-scale and intelligentization, equipment state monitoring and maintenance have become the key link to ensure the safe and stable operation of the equipment. At present, remote equipment state monitoring technology has been widely applied in the industrial field. Through the distributed sensor network, equipment operation data is collected, and combined with big data analysis technology, real-time diagnosis and predictive maintenance of equipment state are realized.

[0003] However, the existing equipment state monitoring technology has the following shortcomings:

[0004] Firstly, the traditional data analysis method usually adopts a single data processing flow, which is difficult to consider both the dynamic change characteristics and the basic characteristics of the equipment state, resulting in limited accuracy of state recognition. Especially under complex working conditions, a single analysis method often cannot fully capture the multi-dimensional characteristics of the equipment state.

[0005] Secondly, the existing deep learning model has poor effect in processing high-dimensional sparse data generated by industrial equipment. The gating structure of the standard LSTM model is not optimized for sparse data characteristics, which makes it difficult to effectively extract and utilize the key information in sparse representation, affecting the perception ability and prediction accuracy of the model.

[0006] Thirdly, the commonly used fixed dictionary method lacks adaptability and cannot automatically adjust the feature extraction strategy according to the dynamic changes of the equipment working conditions. This rigid feature extraction method reduces the adaptability of the system to the changes of the equipment state, and also affects the reliability of the diagnosis results.

[0007] In addition, in practical applications, the transmission and processing efficiency of a large amount of equipment operation data is not high, and the system response speed is slow, which is difficult to meet the real-time monitoring requirements. At the same time, the existing technology has limited ability in predictive maintenance, and often cannot timely warn potential equipment failures.

[0008] Therefore, it is necessary to propose a new remote equipment state diagnosis and maintenance optimization method, which can effectively solve the above technical problems and improve the accuracy, real-time performance and reliability of equipment state monitoring. SUMMARY

[0009] The problem to be solved by the present application is to provide a remote equipment state diagnosis and maintenance optimization method and system based on big data, so as to realize accurate diagnosis and predictive maintenance of the equipment state.

[0010] In order to solve the above technical problems, the technical scheme adopted by the present application is: a remote equipment state diagnosis and maintenance optimization method based on big data, characterized by comprising the following steps:

[0011] S1. Collecting equipment key operation parameters and necessary system configuration information through distributed sensors, performing data compression through edge nodes and transmitting according to importance level;

[0012] S2. Standardizing and preprocessing the received equipment operation data, including outlier identification and missing data filling, to obtain standardized equipment operation parameter data stream;

[0013] S3. Based on the standardized equipment operation parameter data stream, using K-SVD algorithm to construct and optimize the dictionary, and using the dictionary to convert the high-dimensional standardized equipment operation parameter data stream into low-dimensional sparse representation through tracking algorithm, extracting key features representing equipment state from the sparse representation, and storing the key features, corresponding metadata and dictionary information into a distributed database;

[0014] S4. Using the key features in the distributed database to construct a double-flow sparse-aware LSTM model with sparse feature coding layer, wherein the main flow branch processes the time sequence change rule of the sparse representation, and the auxiliary flow branch analyzes the usage mode of the basis vector of the dictionary, and the LSTM unit of the double-flow sparse-aware LSTM model includes an improved input gate according to the sparsity adjustment factor, an improved forgetting gate according to the memory adjustment factor and an improved output gate according to the output adjustment factor;

[0015] S5. Using a sliding window mechanism to organize the key features in time sequence, inputting the organized feature sequence into the constructed double-flow sparse-aware LSTM model, and obtaining the diagnosis result of the current state of the equipment and the prediction result of the future state through autoregressive operation respectively;

[0016] S6. Setting a maintenance trigger condition according to the diagnosis result and the prediction result, and issuing a corresponding maintenance instruction when the trigger condition is met;

[0017] Wherein, the metadata includes sampling timestamp, sensor identifier, data acquisition cycle basic attribute information, which is used to ensure the traceability and time sequence correlation of equipment operation parameter data;

[0018] Wherein, the dictionary information includes basis vector matrix, dictionary size parameter and reconstruction error threshold, which is used to ensure the accuracy of sparse representation and the effectiveness of dictionary optimization;

[0019] The time sequence variation rule of the sparse representation refers to a feature obtained by analyzing sparse representations of multiple continuous time points, including sparse coefficient amplitude variation, non-zero coefficient position shift, and similarity between adjacent time points, and is used to depict a dynamic evolution process of the device state.

[0020] The usage mode of the basis vector of the dictionary includes basis vector activation frequency statistics, basis vector combination probability distribution, and basis vector activation sequence mode, and is used to represent basic state characteristics and typical working condition modes of the device.

[0021] The training data set is composed of device operation parameter data streams that are preprocessed by standardization and contain typical operation data samples of the device under different working conditions.

[0022] The variation trend of the reconstruction error with the dictionary size is analyzed, and the size value corresponding to the stable reconstruction error is selected as the final dictionary size.

[0023] Based on the dictionary size, the K-SVD algorithm is used to construct and optimize the dictionary until the reconstruction error is lower than a preset reconstruction error threshold or the element variation of the dictionary is smaller than a preset variation threshold. The dictionary generated in the K-SVD algorithm optimization process constitutes the basis vector matrix.

[0024] The tracking algorithm uses an orthogonal matching pursuit (OMP) algorithm. The algorithm projects the standardized device operation parameter data stream into the dictionary space through an iterative solving process, thereby obtaining the sparse representation, and controls the accuracy of the sparse representation by setting a reconstruction error threshold.

[0025] The key features extracted from the sparse representation include the sparse coefficients and the basis vector index information of the dictionary corresponding to the sparse coefficients, and the timestamp information in the metadata is associated.

[0026] The double-flow sparse-aware LSTM model is improved based on the gating structure of a standard LSTM model. Sparse degree adjustment factors, memory adjustment factors, and output adjustment factors are introduced into the input gate, the forgetting gate, and the output gate, respectively. The improved structure is as follows:

[0027] The input gate introduces a sparse degree adjustment factor α into the input gate calculation formula of the standard LSTM. t, whose calculation formula is:

[0028] i t = σ(W i x t + U i h t-1 + α t ⊙w i + b i )

[0029] wherein w i is an input feature weight vector, used to adjust the importance of different position features in the sparse representation, ⊙ represents element-level multiplication, W i is an input weight matrix, U i is an input gate hidden state weight matrix, h t-1 is the hidden state vector at t-1 time, b i is an input gate bias vector, and the sparse degree adjustment factor α t has a calculation formula as follows:

[0030]

[0031] wherein ε is a preset reconstruction error threshold, count(|x t |>ε) counts the number of sparse coefficients greater than the reconstruction error threshold in the current input, and dim(x t ) is the dimension of the sparse representation;

[0032] The forgetting gate introduces a memory adjustment factor β t on the basis of the forgetting gate calculation formula of the standard LSTM, and the calculation formula is:

[0033] f t = σ(W f x t + U f h t-1 + β t ⊙w f + b f )

[0034] wherein w f is a forgetting feature weight vector, used to adjust the retention degree of historical information in the sparse representation, W f is an input weight matrix, U f is a hidden state weight matrix of the forgetting gate, b f is a bias vector of the forgetting gate, and the memory adjustment factor β t has a calculation formula as follows:

[0035] β t = sim(s t , s t-1 )

[0036] wherein s t and s t-1 are the sparse representation of the state vector at the current time and the previous time, respectively;

[0037] The output gate introduces an output adjustment factor γ t based on the output gate calculation formula of the standard LSTM, and the calculation formula is:

[0038] o t = sigma(W o x t + U o h t-1 + γ t ⊙ w o + b o )

[0039] wherein w o is an output feature weight vector, used to adjust the output importance of different dimensions of features in the sparse representation, W o is an input weight matrix, U o is an implicit state weight matrix of the output gate, b o is a bias vector of the output gate, and the calculation formula of the output adjustment factor γ t is:

[0040] γ t = lambda * alpha t + (1-lambda) * beta t

[0041] wherein lambda is a weight parameter, used to balance the influence of the sparsity adjustment factor and the memory adjustment factor.

[0042] Further technical solutions are that the specific structure of the main branch and the auxiliary branch of the double-flow sparse perception LSTM model is:

[0043] The main branch contains a recurrent neural network layer for processing the sparse coefficient sequence, and the output of the recurrent neural network layer is converted into a dynamic feature vector through a fully connected layer;

[0044] The auxiliary branch contains a statistical analysis layer for analyzing the usage mode of basis vectors, and the statistical analysis layer calculates the activation frequency and combination relationship of the basis vectors, and the output thereof is converted into a basic state feature vector through a fully connected layer;

[0045] A feature fusion layer is arranged at the end of the model, which combines the dynamic feature vector and the basic state feature vector through feature splicing, and adopts a linear discriminant analysis method for dimension reduction to obtain a final state representation vector.

[0046] Further technical solutions are as follows: specific implementation of the sliding window mechanism and the autoregressive operation includes: the key feature sequence is organized by using a time window with a fixed size of T time steps, the key feature includes a sparse coefficient and a corresponding basis vector index, and the window is moved by one step each time new data is received;

[0047] Based on the key feature sequence in the current time window, the main stream branch analyzes the sparse coefficient sequence, the auxiliary stream branch analyzes the basis vector index sequence, and the current state representation vector is obtained through the feature fusion layer;

[0048] Multi-step prediction is performed in an autoregressive manner, and the state representation vector prediction result at t time is taken as the input at t+1 time, and the state prediction of future multiple time steps is completed by recursive calculation;

[0049] In the autoregressive process, the main stream branch predicts the future trend of the sparse feature, the auxiliary stream branch predicts the future activation mode of the basis vector of the dictionary, and the prediction results of the main stream branch and the auxiliary stream branch are fused at the end of the network to output the final device state prediction result and the current state diagnosis result.

[0050] Further technical solutions are as follows: the maintenance trigger condition includes:

[0051] When the state diagnosis result shows that the device is in an abnormal state, the system issues a maintenance instruction, and the maintenance instruction includes a description of the currently detected abnormal state;

[0052] When the sparse coefficient value in the key feature exceeds a preset feature threshold, the system issues a preventive maintenance suggestion;

[0053] When the autoregressive prediction result shows that the device state has a deterioration trend, the system issues an inspection notification.

[0054] The application also discloses a remote device state diagnosis and maintenance optimization system based on big data, characterized by comprising:

[0055] A data acquisition module is used to acquire device operation parameters and system configuration information through a distributed sensor network, and the data acquisition module includes an edge node used to perform data compression and hierarchical transmission;

[0056] A data preprocessing module is used to perform standardization processing on received device operation data, including an outlier identification unit and a missing data filling unit, and outputs a standardized device operation parameter data stream;

[0057] The feature extraction module comprises a dictionary construction unit configured to construct and optimize a dictionary based on a K-SVD algorithm, a sparse representation unit configured to convert a standardized device operation parameter data stream into a low-dimensional sparse representation through an orthogonal matching pursuit algorithm, and a feature storage unit configured to store extracted key features, metadata and dictionary information into a distributed database;

[0058] The state diagnosis module comprises a dual-stream neural network unit configured to construct a dual-stream sparse perception LSTM model with a sparse feature encoding layer, a time series processing unit configured to organize a key feature sequence through a sliding window mechanism, and a prediction analysis unit configured to perform an autoregressive operation and output device state diagnosis results and prediction results.

[0059] The maintenance decision module is configured to determine whether a maintenance trigger condition is met based on the state diagnosis results and the prediction results, and to issue a corresponding maintenance instruction when the condition is met.

[0060] The system modules exchange data and coordinate tasks through a distributed message middleware.

[0061] The present application has the following advantages:

[0062] (1) The dual-stream sparse perception LSTM model has a unique dual-stream architecture design, which divides the device state analysis into two branches of the main stream and the auxiliary stream, can capture the dynamic change features and the basic features of the device state at the same time, and significantly improves the accuracy of state recognition.

[0063] (2) The present application innovatively improves the LSTM gating structure, introduces a sparsity adjustment factor, a memory adjustment factor and an output adjustment factor, so that the model can more effectively process sparse data features, greatly improving the perception ability and prediction accuracy of the model for the device state.

[0064] (3) The present application adopts an adaptive dictionary learning mechanism based on the K-SVD algorithm, automatically determines the optimal dictionary size by analyzing the reconstruction error, and realizes efficient sparse representation combined with the orthogonal matching pursuit algorithm, which has stronger adaptability and robustness.

[0065] (4) The present application uses edge nodes for data compression and hierarchical transmission in terms of data processing efficiency, significantly reduces the data transmission burden, improves the system response speed, and solves the problem of low data processing efficiency in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 It is a flow chart of a remote device state diagnosis and maintenance optimization method based on big data.

[0067] Figure 2 It is a feature extraction flow chart.

[0068] Figure 3 A dual-flow sparse-aware LSTM model structure is provided.

[0069] Figure 4 A system architecture diagram is provided. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0071] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the application, but the application can also be implemented in other manners different from those described herein, and a person of ordinary skill in the art can make similar generalizations without departing from the concept of the application, and therefore the application is not limited to the specific embodiments disclosed below.

[0072] With reference to Figure 1 The main steps of the method provided by the application are as follows:

[0073] S1. Collecting key operation parameters and necessary system configuration information of a distributed sensor acquisition device, performing data compression through an edge node, and transmitting the data according to importance levels;

[0074] S2. Standardizing and pre-processing the received device operation data, including identifying abnormal values and filling in missing data, to obtain a standardized device operation parameter data stream;

[0075] S3. Based on the standardized device operation parameter data stream, constructing and optimizing a dictionary by using a K-SVD algorithm, and converting the high-dimensional standardized device operation parameter data stream into a low-dimensional sparse representation by using the dictionary through a tracking algorithm, extracting key features capable of representing device states from the sparse representation, and storing the key features, corresponding metadata, and dictionary information in a distributed database;

[0076] S4. Using the key features in the distributed database to construct a dual-flow sparse-aware LSTM model with a sparse feature coding layer, wherein a main flow branch processes the time sequence variation law of the sparse representation, and an auxiliary flow branch analyzes the usage mode of basis vectors of the dictionary, and an LSTM unit of the dual-flow sparse-aware LSTM model includes an improved input gate according to a sparsity adjustment factor, an improved forgetting gate according to a memory adjustment factor, and an improved output gate according to an output adjustment factor;

[0077] S5. The key features are time-sequentially organized by using a sliding window mechanism, and the organized feature sequence is input into the constructed double-flow sparse perception LSTM model to obtain the diagnosis result of the current state and the prediction result of the future state of the equipment through self-recurrence operation respectively;

[0078] S6. A maintenance trigger condition is set according to the diagnosis result and the prediction result, and a corresponding maintenance instruction is issued when the trigger condition is met;

[0079] The method for collecting and transmitting the equipment operation parameters in the step S1 is as follows:

[0080] S101: Based on the equipment characteristics and monitoring requirements, a distributed sensor network is arranged at the key parts of the equipment. Specifically, sensors of corresponding types are installed at the key monitoring points of vibration, temperature, pressure and the like of the equipment; preferably, the sampling frequency of the sensors is determined according to the dynamic characteristics of the measured parameters, the sampling frequency of the vibration signal is not less than 1 kHz, and the sampling frequency of the slow-varying parameters such as temperature and pressure is not less than 1 Hz; each sensor has a unique identification code for subsequent data tracing and correlation analysis.

[0081] S102: The original data collected by each sensor is transmitted to the nearby edge node through an industrial field bus. Preferably, the edge node adopts an industrial-grade embedded computing platform and has local storage and computing capabilities; the edge node and the sensor communicate with each other through an industrial field bus such as RS485 or CAN to ensure the real-time and reliability of data transmission.

[0082] S103: The collected original data is compressed at the edge node. Preferably, the following compression strategy is adopted: for fast-varying parameters such as vibration signals, wavelet transform is used for multi-scale decomposition of the signals, and important coefficients are retained; for slow-varying parameters, a piecewise linear fitting method is used for data compression; the data compression process ensures that the information loss is controlled within a preset range.

[0083] S104: The compressed data is classified according to importance. Preferably, the importance of the data is determined based on the following principles: parameters related to key performance indicators of the equipment are first-class; feature parameters reflecting potential faults of the equipment are second-class; auxiliary monitoring parameters are third-class. Different levels of data use different transmission strategies: first-class data is transmitted in real time; second-class data is transmitted quasi-in real time with a delay of not more than 5 minutes; third-class data is transmitted periodically in batches.

[0084] S105: The classified data is transmitted to the remote monitoring center through an encrypted channel. Preferably, the encrypted transmission adopts SSL / TLS protocol to ensure the security of data transmission; for the case of limited network bandwidth, the transmission of high importance data is prioritized.

[0085] Through the above method, efficient collection and hierarchical transmission of equipment operation parameters are realized, laying a foundation for subsequent data processing and state analysis. The method effectively reduces the data transmission burden, improves the system response speed, and ensures real-time acquisition of key data through edge computing and hierarchical transmission strategy.

[0086] The data standardization preprocessing method in step S2 includes the following specific steps:

[0087] S201: Abnormal value identification is performed on the received equipment operation data. Specifically, the following method is used: Preferably, first, based on the 3σ criterion, identify the statistically significant abnormal values, calculate the mean and standard deviation of each parameter, and mark the data points deviating from the mean by more than 3 times the standard deviation as potential abnormal values; second, combined with the physical constraint conditions of the equipment operation parameters, such as the effective value range of temperature, pressure and other parameters, mark the data points exceeding the physical constraints as definite abnormal values; finally, by calculating the change rate of adjacent data points, identify the sudden change points that do not conform to the parameter change law.

[0088] S202: Processing is performed on the identified abnormal values. Preferably, for definite abnormal values such as data points exceeding physical constraints, they are directly marked as invalid data; for potential abnormal values, combined with equipment operating condition information, if the abnormal value occurs during equipment operating condition switching, it is retained, otherwise it is marked as a suspicious data point. The suspicious data point is given a lower weight in subsequent analysis.

[0089] S203: Perform missing data filling processing. Preferably, according to the time span of data missing and the parameter characteristics, different filling strategies are adopted: for short-time (not more than 3 sampling periods) data missing, linear interpolation method is used for filling; for long-time data missing, if the parameter has periodic characteristics, the filling is based on the periodic law of historical data; if the parameter has no obvious periodicity, a multivariate regression model is used to estimate and fill the missing data using related parameter information.

[0090] S204: Standardization conversion is performed on the preprocessed data. Preferably, the minimum-maximum normalization method is used to map the numerical range of each parameter to the [0, 1] interval: first, determine the effective value range of each parameter based on historical data; then, set the standardization reference value according to the physical meaning of the parameter; finally, convert the data to a unified dimension through linear transformation. For different types of parameters, such as temperature, pressure, vibration, etc., corresponding standardization conversion models are established.

[0091] S205: Construct a standardized device operating parameter data stream. Preferably, the standardized parameters are aligned according to the time stamp to form a unified data stream format; at the same time, the quality marks generated in the data preprocessing process are retained, including abnormal value marks, data filling marks and other information, for subsequent analysis of the reliability evaluation.

[0092] Through the above data preprocessing steps, the quality and availability of the device operating data are effectively improved, providing a reliable data basis for subsequent feature extraction and state diagnosis. The method comprehensively considers the physical characteristics and statistical characteristics of the device operating parameters, and the preprocessing result not only maintains the authenticity of the data, but also ensures the normativity of the analysis process.

[0093] Referring to Figure 2 , the feature extraction and sparse representation method in step S3 is as follows:

[0094] S301: Construct a training data set based on the standardized device operating parameter data stream. Preferably, typical operating data samples of the device under different operating conditions are selected from the historical data, including normal operation, load change, start-stop transition and other operating conditions; preferably, the selected data samples are divided into a training set and a validation set according to a ratio of 80%:20%; preferably, the training data is randomly shuffled to improve the robustness of the training.

[0095] S302: Determine the initial parameter configuration of the K-SVD algorithm. Preferably, the initial dictionary size is set to be 2 times the dimension of the training sample, the reconstruction error threshold is initially set to 0.01, and the maximum number of iterations is set to 100. An initial dictionary matrix is generated using a random initialization method, and the basis vectors of the dictionary are normalized.

[0096] S303: Perform a dictionary size optimization process. Preferably, dictionaries of different sizes are optimized and verified: first, the dictionary size is gradually increased from the initial value, with a step size of 0.5 times the dimension of the training sample; then, for each size value, the K-SVD algorithm is executed to optimize the dictionary until the reconstruction error threshold is met or the maximum number of iterations is reached; next, the reconstruction error is calculated on the validation set; finally, a curve of the reconstruction error versus the dictionary size is plotted, and the size value corresponding to the stable error is selected as the final dictionary size.

[0097] S304: Based on the optimized dictionary size, perform a complete optimization process of the K-SVD algorithm. Preferably, in each iteration: first, calculate the sparse coefficients using the Orthogonal Matching Pursuit (OMP) algorithm; then, update the dictionary matrix column by column, and perform singular value decomposition on each basis vector and its corresponding non-zero sparse coefficient to update the basis vector and the corresponding sparse coefficient; finally, calculate the reconstruction error, and stop the iteration if the termination condition is met.

[0098] S305: Sparse representation conversion using optimized dictionary. Preferably, the input standardized equipment operation parameter data stream is subjected to sparse decomposition using the OMP algorithm: first, set the sparsity threshold to control the number of non-zero coefficients; then, through the iterative projection process, select the most relevant basis vectors and calculate the sparse coefficients; finally, record the non-zero sparse coefficients and their corresponding basis vector indexes.

[0099] S306: Extract key features and build feature storage structure. Extract the following key features from the sparse representation: amplitude and position information of sparse coefficients, usage frequency statistics of basis vectors, and similarity of sparse representations at adjacent time points. At the same time, associate the extracted key features with metadata (timestamp, sensor identifier, etc.) to build a structured feature storage format.

[0100] S307: Store the extracted key features, metadata, and dictionary information in a distributed database. Use a distributed storage strategy to store large-scale feature data in a database cluster, establish feature indexes to support efficient querying, and ensure data consistency through database transaction mechanisms.

[0101] Through the above method, efficient dimensionality reduction and feature extraction of equipment operation data are achieved. This method optimizes the dictionary size and sparse representation strategy adaptively, not only maintaining the key information of the original data, but also significantly reducing the complexity of data storage and processing. The extracted features contain both dynamic change features and basic state features of the equipment, providing comprehensive feature representation for subsequent state diagnosis.

[0102] In step S4, the reference Figure 3 The method for constructing a double-flow sparse perception LSTM model comprises the following specific steps:

[0103] S401: Initialize the basic structure of the double-flow network. According to the key feature dimensions stored in the distributed database, configure the input layer specifications of the main flow branch and the auxiliary flow branch. The input dimension of the main flow branch matches the dimension of the sparse representation, and the input dimension of the auxiliary flow branch corresponds to the number of dictionary basis vectors. Preferably, set the size of the network's hidden layer, with an initial value of 1.5 times the input dimension.

[0104] S402: Construct an improved LSTM unit structure. Introduce a sparsity adjustment factor in the input gate to adjust feature importance by calculating the proportion of non-zero elements in the current input; introduce a memory adjustment factor in the forget gate to control the degree of historical information preservation based on the similarity of sparse representations at adjacent time points; introduce an output adjustment factor in the output gate to balance the output weight of features in combination with sparsity and memory factors.

[0105] The specific improvement structure is as follows:

[0106] The input gate introduces a sparsity adjustment factor a on the basis of the input gate calculation formula of the standard LSTM t The calculation formula is:

[0107] i t = σ (W i x t + U i h t-1 + a t ⊙ w i + b i )

[0108] Wherein, w i is the input feature weight vector, which is used to adjust the importance of different position features in sparse representation, ⊙ represents element-level multiplication, W i is the input weight matrix, U i is the hidden state weight matrix of the input gate, h t-1 is the hidden state vector at t-1 time, b i is the bias vector of the input gate, and the calculation formula of the sparsity adjustment factor a t is:

[0109]

[0110] Wherein, ε is a preset reconstruction error threshold, preferably, the reconstruction error threshold is set to 0.01, which can effectively distinguish significant sparse coefficients and noise coefficients. count (|x t |> ε) statistics the number of sparse coefficients greater than the reconstruction error threshold in the current input, dim (x t ) is the dimension of sparse representation;

[0111] The forgetting gate introduces a memory adjustment factor β on the basis of the forgetting gate calculation formula of the standard LSTM t The calculation formula is:

[0112] f t = σ (W f x t + U f h t-1 + β t ⊙ w f + b f )

[0113] Wherein, w f is the forgetting feature weight vector, which is used to adjust the retention degree of historical information in sparse representation, W f is the input weight matrix, U f is the hidden state weight matrix of the forgetting gate, b f is the bias vector of the forgetting gate, and the memory adjustment factor βt The calculation formula is:

[0114] β t =sim(s t s t-1 )

[0115] Among them, s t and s t-1 These are the sparse representations of the state vectors at the current time and the previous time, respectively.

[0116] The output gate is calculated using the standard LSTM output gate formula, and an output adjustment factor γ is introduced. t The calculation formula is as follows:

[0117] o t =σ(W o x t +U o h t-1 +γ t ⊙w o +b o )

[0118] Among them, w o W is the output feature weight vector, used to adjust the output importance of features of different dimensions in the sparse representation. o For the input weight matrix, U o Let b be the hidden state weight matrix of the output gate. o The output gate's bias vector, the output adjustment factor γ t The calculation formula is:

[0119] γ t =λ·α t +(1-λ)·β t

[0120] Wherein, λ is a weighting parameter, preferably set to 0.6, so as to achieve an appropriate balance between sparsity adjustment and memory adjustment, and to balance the influence of sparsity adjustment factor and memory adjustment factor.

[0121] Preferably, the input feature weight vector w i Forgotten feature weight vector w f and output feature weight vector w o The initial values ​​are all randomly initialized using a standard normal distribution to ensure the stability of the training.

[0122] Preferably, the similarity function sim(s) t s t-1 Using the cosine similarity calculation method, the degree of change in sparse representation at adjacent time points can be effectively measured.

[0123] Preferably, the activation function adopts a sigmoid function, with an output range of [0, 1], suitable as a gating signal.

[0124] S403: Implement the recurrent neural network layer of the main stream branch. Design a multi-layer LSTM structure to process the sparse coefficient sequence, and use the improved gating mechanism described above for each LSTM unit. Add a fully connected layer after the LSTM layer to map the high-dimensional features to a lower-dimensional dynamic feature space. Preferably, to prevent overfitting, introduce a Dropout mechanism between the LSTM layers, with an inactivation probability of 0.2.

[0125] S404: Construct the statistical analysis layer of the auxiliary stream branch. Design multiple parallel feature extraction units to calculate the activation frequency, combination probability, and activation sequence pattern of the basis vector, respectively. Convert the statistical analysis results to a fixed-dimensional basis state feature vector through a fully connected layer. Preferably, add L2 regularization during feature conversion to improve the generalization ability of the model.

[0126] S405: Implement the feature fusion layer. Combine the dynamic feature vector and the basis state feature vector using feature concatenation to obtain the initial fusion features. Then perform dimension reduction on the fusion features using linear discriminant analysis, and the target dimension after dimension reduction is determined by the feature variance contribution rate. Preferably, retain more than 90% of the variance information.

[0127] S406: Establish a model training and optimization mechanism. Preferably, use the mini-batch stochastic gradient descent method to train the model, with a batch size of 64 and an initial learning rate of 0.001. Preferably, introduce a learning rate decay strategy, and when the validation set loss does not improve for 5 consecutive epochs, reduce the learning rate to 0.1 times the original. Preferably, implement an early stopping mechanism, and when the validation set loss does not significantly improve for 10 consecutive epochs, stop training.

[0128] S407: Implement the online updating mechanism of the model. Preferably, periodically use newly collected data to incrementally train the model, with an update period determined by the frequency of device operating condition changes, usually 1-2 weeks. During model updating, use a sliding window mechanism to manage training data to maintain the timeliness of the training set.

[0129] Through the above steps, a dual-stream neural network model capable of effectively processing sparse data features is constructed. The model enhances the perception of sparse features through an improved LSTM gating structure, effectively fuses dynamic features and basic features through a dual-stream architecture, and provides a reliable model foundation for accurate diagnosis and prediction of device status. At the same time, the online updating mechanism of the model ensures its adaptability and timeliness.

[0130] The state diagnosis and prediction method in step S5 is as follows:

[0131] S501: Initialize the sliding window parameters. Preferably, set the size T of the time window according to the time-series correlation characteristics of the device state, generally taking a value of 24-48 time steps to cover the complete evolution period of the device state. Preferably, set the window sliding step size to 1 to ensure continuous monitoring of state changes. The key features in the window include the sparse coefficient sequence and its corresponding basis vector index information.

[0132] S502: Organize the time-series feature sequence. For the data in each time window: first, arrange the sparse coefficients in chronological order to form a T x M-dimensional coefficient matrix, where M is the dimension of the sparse representation; then, construct the basis vector usage sequence to record the index of the activated basis vector at each time point; finally, correlate the feature sequences of adjacent windows through the overlapping part to ensure the continuity of state analysis.

[0133] S503: Perform current state diagnosis. Input the organized feature sequence into the dual-flow sparse-aware LSTM model: the main branch analyzes the change pattern of the sparse coefficient sequence to extract state dynamic features; the auxiliary branch analyzes the activation law of the basis vector to extract state basis features. Through the feature fusion layer, obtain the state representation vector at the current time, and output the specific state diagnosis result based on the pre-trained state recognizer.

[0134] S504: Configure the autoregressive prediction parameters. Preferably, set the prediction step P, generally 12-24 future time steps; determine the error accumulation threshold for iterative prediction to control the reliability of long-term prediction; establish a confidence interval estimation mechanism for prediction results, preferably calculate the 95% confidence interval for each prediction point.

[0135] S505: Perform the multi-step prediction process. Use the iterative prediction strategy: first, input the state representation vector at time t into the model to predict the state at time t+1; then, use the prediction result as the new input to predict the state at time t+2, and recursively complete P-step prediction; during the prediction process, the main branch predicts the trend of sparse feature changes, and the auxiliary branch predicts the activation pattern of the basis vector; the prediction results of the two branches are combined in the feature fusion layer to obtain the final state prediction.

[0136] S506: Implement the reliability evaluation of the prediction results. Calculate the uncertainty indicators of the prediction results: first, estimate the confidence interval of the prediction results based on the historical prediction errors; then, analyze the cumulative trend of the prediction errors with the prediction step; finally, when the cumulative error exceeds the set threshold, mark the reliability of the prediction results to provide a reference for maintenance decisions.

[0137] Through the above method, real-time diagnosis and reliable prediction of the equipment state are realized. The method guarantees the continuity of state analysis through a sliding window mechanism, provides comprehensive state representation through a double-flow network architecture, and realizes multi-step prediction through an autoregressive mechanism, thereby providing timely and reliable basis for equipment maintenance decision-making. Meanwhile, the reliability evaluation mechanism of the prediction result ensures the practicability of the prediction.

[0138] The method for maintaining the triggering condition judgment and generating the maintenance instruction in the step S6 comprises the following specific steps:

[0139] S601: classifying the diagnosis result based on the evaluation standard of the equipment state. The equipment state is divided into three categories of normal, early warning and abnormal, and a clear judgment standard is set for each category. The normal state indicates that the equipment operation parameter is within the allowable range; the early warning state indicates that some parameters are close to but not beyond the limit value; and the abnormal state indicates that the key parameter is beyond the safety range.

[0140] S602: analyzing the change trend of the state prediction result. According to the future state sequence obtained through the autoregressive prediction, the change rate of the key performance index is calculated, and compared with the pre-set degradation rate threshold to judge whether there is a significant performance deterioration trend.

[0141] S603: performing the judgment of the maintenance triggering condition. Three basic triggering conditions are set:

[0142] the maintenance instruction is triggered when the diagnosis result shows the abnormal state; the preventive maintenance suggestion is triggered when the key characteristic value exceeds the pre-set threshold; and the inspection notification is triggered when the prediction result shows continuous performance deterioration.

[0143] S604: generating the corresponding maintenance instruction according to the triggering condition. The maintenance instruction contains basic information such as equipment ID, triggering reason, suggested measure, and is sent to the maintenance personnel through the pre-set notification channel. For the maintenance instruction triggered by the abnormal state, the equipment management personnel is also notified.

[0144] As shown in Figure 4 Corresponding to the above method, the embodiment of the present application also discloses a remote equipment state diagnosis and maintenance optimization system based on big data, and the overall architecture of the system comprises:

[0145] A data acquisition module: this module acquires the operation parameters and system configuration information of the equipment through a distributed sensor network. Preferably, the layout of the sensor network follows the key monitoring point coverage principle, and each sensor node is connected to the nearest edge computing node. The edge node is responsible for data compression and hierarchical transmission, thereby effectively reducing the data transmission burden.

[0146] Data preprocessing module: This module is responsible for standardizing the received raw data. The module includes an outlier identification unit and a missing data imputation unit. The outlier identification unit performs anomaly detection based on statistical analysis and physical constraints, while the missing data imputation unit selects appropriate imputation methods based on data characteristics.

[0147] Feature extraction module: This module consists of three core units. The dictionary construction unit is responsible for constructing and optimizing the dictionary based on the K-SVD algorithm; the sparse representation unit uses the orthogonal matching pursuit algorithm to convert data into low-dimensional sparse representation; and the feature storage unit stores the extracted key features, metadata, and dictionary information into a distributed database.

[0148] State diagnosis module: This module is the core processing unit of the system, including a dual-flow neural network unit, a time series processing unit, and a prediction analysis unit. The dual-flow neural network unit implements an LSTM model with a sparse feature encoding layer; the time series processing unit organizes feature sequences through a sliding window mechanism; and the prediction analysis unit performs autoregressive operations to output diagnosis and prediction results.

[0149] Maintenance decision module: This module receives state diagnosis and prediction results, determines whether the maintenance trigger conditions are met, and generates corresponding maintenance instructions when the conditions are met. The module uses a hierarchical triggering mechanism to ensure the timeliness and accuracy of maintenance decisions.

[0150] Each functional module exchanges data and coordinates tasks through a distributed message middleware, ensuring the efficiency and reliability of system operation. The system uses modular design, with standardized interfaces between modules, facilitating system expansion and maintenance.

[0151] The implementation of this system provides a complete technical solution for equipment state monitoring and maintenance optimization. Through the collaborative work of each module, it realizes the intelligent management of the whole process from data acquisition to maintenance decision-making.

Claims

1. A big data based remote equipment state diagnosis and maintenance optimization method, characterized in that, The method comprises the following steps: S1. Collecting key operation parameters and necessary system configuration information of equipment through a distributed sensor acquisition device, performing data compression through an edge node, and transmitting the data according to importance levels; S2. Standardizing and preprocessing the received equipment operation data, including identifying outliers and filling in missing data, to obtain a standardized equipment operation parameter data stream; S3. Based on the standardized equipment operation parameter data stream, constructing and optimizing a dictionary using a K-SVD algorithm, and converting the high-dimensional standardized equipment operation parameter data stream into a low-dimensional sparse representation using the dictionary through a tracking algorithm, extracting key features representing the equipment state from the sparse representation, and storing the key features, corresponding metadata and dictionary information in a distributed database; S4. Using the key features in the distributed database to construct a dual-flow sparse-aware LSTM model with a sparse feature coding layer, wherein the main flow branch processes the time sequence variation of the sparse representation, and the auxiliary flow branch analyzes the usage mode of the basis vectors of the dictionary, and the LSTM unit of the dual-flow sparse-aware LSTM model comprises an improved input gate according to a sparsity adjustment factor, an improved forgetting gate according to a memory adjustment factor, and an improved output gate according to an output adjustment factor; S5. Using a sliding window mechanism to organize the key features in time sequence, inputting the organized feature sequence into the constructed dual-flow sparse-aware LSTM model, and obtaining the diagnosis result of the current state of the equipment and the prediction result of the future state of the equipment through autoregressive operation; S6. Setting a maintenance trigger condition according to the diagnosis result and the prediction result, and issuing a corresponding maintenance instruction when the trigger condition is met; Wherein, the metadata includes sampling time stamp, sensor identifier, data acquisition period basic attribute information, which is used to ensure the traceability and time sequence correlation of the equipment operation parameter data; Wherein, the dictionary information includes basis vector matrix, dictionary size parameter, reconstruction error threshold, which is used to ensure the accuracy of the sparse representation and the effectiveness of the dictionary optimization; Wherein, the time sequence variation of the sparse representation refers to the features obtained by analyzing the sparse representations of multiple consecutive time points, including sparse coefficient amplitude variation, non-zero coefficient position shift, and similarity between adjacent time points, which is used to depict the dynamic evolution process of the equipment state; Wherein, the usage mode of the basis vectors of the dictionary includes basis vector activation frequency statistics, basis vector combination probability distribution, and basis vector activation sequence mode, which is used to represent the basic state characteristics and typical working condition mode of the equipment.

2. The method of claim 1, wherein, The process of constructing and optimizing the dictionary using the K-SVD algorithm comprises: Determining the dictionary size through reconstruction error analysis, wherein the K-SVD algorithm is iteratively executed until a preset termination condition is met, and the reconstruction error of the optimized dictionary for the training data set is calculated based on the training data set for optimization verification of dictionaries of different sizes; Wherein, the training data set is composed of the standardized and preprocessed equipment operation parameter data stream, and contains typical operation data samples of the equipment under different working conditions; The final dictionary size is selected by analyzing the change trend of the reconstruction error with the dictionary size, and the size corresponding to the stable reconstruction error is selected as the final dictionary size; Based on the dictionary size, an optimized dictionary is constructed by using the K-SVD algorithm until the reconstruction error is lower than the preset reconstruction error threshold or the element variation of the dictionary is less than the preset variation threshold, and the dictionary generated in the K-SVD algorithm optimization process constitutes the basis vector matrix.

3. The method of claim 1, wherein, The tracking algorithm adopts an orthogonal matching pursuit algorithm, which projects the standardized device operation parameter data stream to a dictionary space through an iterative solving process to obtain the sparse representation, and controls the accuracy of the sparse representation by setting a reconstruction error threshold; The key features extracted from the sparse representation include the sparse coefficients and the basis vector index information of the corresponding dictionary, and the timestamp information in the metadata is associated.

4. The method of claim 1, wherein, The double-flow sparse-aware LSTM model is improved based on the gating structure of the standard LSTM model, and sparse degree adjustment factors, memory adjustment factors and output adjustment factors are introduced into the input gate, the forgetting gate and the output gate respectively, and the improved structure is specifically: The input gate introduces a sparsity adjustment factor on the basis of an input gate calculation formula of a standard LSTM The calculation formula is: wherein, is an input feature weight vector, used to adjust the importance of features at different positions in the sparse representation, and represents element-wise multiplication, is an input weight matrix, is an input gate hidden state weight matrix, is a hidden state vector at time t-1, is an input gate bias vector, and the sparsity adjustment factor is calculated as follows: wherein, is a pre-set reconstruction error threshold, counting the number of sparse coefficients in the current input that are greater than the reconstruction error threshold, is the dimension of the sparse representation; The forgetting gate introduces a memory adjustment factor on the basis of a forgetting gate calculation formula of a standard LSTM The calculation formula is: wherein, is a forgetting feature weight vector, used to adjust the degree of preservation of historical information in the sparse representation, is an input weight matrix, is a hidden state weight matrix of the forgetting gate, is a bias vector of the forgetting gate, and the memory adjustment factor The calculation formula is: wherein, and are the state vectors sparsely represented at the current time instant and at the previous time instant, respectively; The output gate introduces an output adjustment factor on the basis of an output gate calculation formula of a standard LSTM The calculation formula is: wherein, is an output feature weight vector, used to adjust the output importance of different dimensional features in the sparse representation, is an input weight matrix, is an output gate hidden state weight matrix, is an output gate bias vector, and the output adjustment factor The calculation formula is: wherein, is a weight parameter balancing the influence of the sparsity adjustment factor and the memory adjustment factor.

5. The method of claim 1, wherein, The specific structure of the main flow branch and the auxiliary flow branch of the double-flow sparse-aware LSTM model is: The main flow branch includes a recurrent neural network layer for processing sparse coefficient sequences, and the output of the recurrent neural network layer is converted into a dynamic feature vector through a fully connected layer; The auxiliary flow branch includes a statistical analysis layer for analyzing the basis vector usage mode, and the statistical analysis layer calculates the activation frequency and combination relationship of the basis vector, and the output thereof is converted into a basic state feature vector through a fully connected layer; The LSTM model is provided with a feature fusion layer at the end, which combines the dynamic feature vector and the basic state feature vector through feature splicing, and adopts a linear discriminant analysis method for dimension reduction to obtain a final state representation vector.

6. The method of claim 1, wherein, The specific implementation of the sliding window mechanism and the autoregressive operation includes: A time window with a fixed size of T time steps is used to organize the key feature sequence, and the key features include sparse coefficients and corresponding basis vector indexes, and the window is moved one step at a time when new data is received; Based on the key feature sequence in the current time window, the main flow branch analyzes the sparse coefficient sequence, and the auxiliary flow branch analyzes the basis vector index sequence, and the current state representation vector is obtained through the feature fusion layer; A multi-step prediction is performed in an autoregressive manner, and the state representation vector prediction result at time t is used as the input at time t+1, and the state prediction at future time steps is completed by recursion; In the autoregressive process, the main flow branch predicts the future trend of the sparse feature, and the auxiliary flow branch predicts the future activation mode of the basis vector of the dictionary, and the prediction results of the main flow branch and the auxiliary flow branch are fused at the end of the network to output the final device state prediction result and the current state diagnosis result.

7. The method of claim 1, wherein, The maintenance trigger condition includes: When the diagnostic result shows that the equipment is in an abnormal state, the system issues a maintenance instruction, which includes a description of the currently detected abnormal state; When the sparse coefficient value in the key feature exceeds the preset feature threshold, the system issues a preventive maintenance suggestion; When the prediction result shows that the equipment state has a deterioration trend, the system issues an inspection notice.

8. A big data based remote device condition diagnosis and maintenance optimization system for implementing the method as claimed in claim 1, characterized by, Comprise: A data acquisition module for acquiring equipment operating parameters and system configuration information through a distributed sensor network, the data acquisition module including edge nodes for performing data compression and hierarchical transmission; A data preprocessing module for standardizing the received equipment operating data, including an outlier identification unit and a missing data imputation unit, and outputting standardized equipment operating parameter data streams; A feature extraction module including a dictionary construction unit for constructing and optimizing a dictionary based on the K-SVD algorithm, a sparse representation unit for converting standardized equipment operating parameter data streams into low-dimensional sparse representations through an orthogonal matching pursuit algorithm, and a feature storage unit for storing extracted key features, metadata, and dictionary information into a distributed database; A state diagnosis module including a dual-flow neural network unit for constructing a dual-flow sparse perception LSTM model with a sparse feature encoding layer, a time series processing unit for organizing key feature sequences through a sliding window mechanism, and a prediction analysis unit for performing autoregressive operations and outputting equipment state diagnosis results and prediction results; A maintenance decision module for determining whether maintenance trigger conditions are met based on the state diagnosis results and prediction results, and issuing corresponding maintenance instructions when the conditions are met; Wherein, the system modules exchange data and coordinate tasks through a distributed message middleware.

Images