A warning method and system based on front-end perception data

Abstract

The present invention relates to the technical field of industrial equipment fault detection and early warning, and discloses an early warning method and system based on front-end perception data, wherein an early warning method based on front-end perception data includes: acoustic signal enhancement and feature extraction, construction of an acoustic feature dictionary; fault stage division and feature analysis, using an integrated change point detection algorithm to identify the key stages of fault evolution, and construction of a stage-by-stage fault feature library; time series evolution model construction, building an acoustic feature time series evolution model based on a recurrent neural network to identify fault evolution laws; chain progressive prediction, using a stage-specific prediction method to predict future fault states; multi-dimensional early warning decision-making, generating early warning information, and assisting maintenance decisions; improving fault detection sensitivity and increasing early warning lead time; realizing the prediction of fault development trends, improving prediction accuracy, and being able to predict fault development trends in advance.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault detection and early warning, and more specifically, to an early warning method and system based on front-end perception data. Background Art

[0002] In industrial production environments, the operating status of mechanical equipment is crucial to production efficiency and safety. Timely detection and prediction of fault development trends can effectively avoid equipment damage and production interruptions. Currently, industrial equipment fault monitoring and early warning mainly have the following technical problems:

[0003] Traditional mechanical fault monitoring mainly relies on vibration sensor data, which is not sensitive enough to early and weak fault characteristics and is easily interfered with in noisy environments. Although acoustic monitoring has shown good potential in early fault detection, existing acoustic monitoring systems mostly focus on immediate fault detection and lack a comprehensive understanding of the fault evolution process. At the same time, existing systems have difficulty capturing the stage-by-stage characteristic changes and inflection points in the fault evolution process, resulting in insufficient accuracy in predicting the fault development trajectory. In addition, the output information of traditional early warning systems is single and lacks specific decision-making guidance, making it impossible to provide maintenance personnel with comprehensive and intuitive decision support information.

[0004] Therefore, there is an urgent need for a new early warning method that can improve the detection sensitivity of early weak fault characteristics of mechanical equipment in industrial environments, predict the development trend of faults, and accurately grasp the fault evolution process. Summary of the Invention

[0005] The present invention provides an early warning method and system based on front-end perception data to solve the technical problems in related technologies such as insufficient sensitivity in industrial equipment fault detection, lack of fault trend prediction capability, and difficulty in capturing key features of fault evolution.

[0006] The present invention provides an early warning method based on front-end perception data, including: acoustic signal enhancement and feature extraction, modal separation and enhancement processing of the collected acoustic signals, application of improved Vaviley transform to extract weak fault features, construction of an acoustic feature dictionary, and use of improved Vaviley transform to enhance the ability to extract weak features:

[0007] ;

[0008] in represents the Varvey coefficient, is a scale parameter that controls the frequency resolution of the analysis, is the translation parameter, controlling the time positioning, is the wavelet basis function, is a window function used to enhance the ability to extract local features. represents the complex conjugate operator, represents the integral operation from negative infinity to positive infinity, Indicates about variables The differential of

[0009] According to the characteristics of different types of mechanical faults, adaptive wavelet basis functions are constructed:

[0010] ;

[0011] in Indicates that for Adaptive wavelet basis function of fault-like features, For the Modulation function for fault-like feature matching;

[0012] Fault stage division and feature analysis: Based on weak fault features, an integrated change point detection algorithm is used to identify the key stages of fault evolution and build a stage-by-stage fault feature library.

[0013] Construction of a time series evolution model: Based on the staged fault feature library, an acoustic feature time series evolution model is constructed based on a recurrent neural network to identify the fault evolution law;

[0014] Chain progressive prediction: Based on the acoustic feature time series evolution model and the stage-by-stage fault feature library, a stage-specific prediction method is used to predict future fault states.

[0015] Multi-dimensional early warning decision-making generates early warning information based on predicted future fault states to assist maintenance decisions.

[0016] Furthermore, the process of modal separation and enhancement includes:

[0017] Decompose the collected original acoustic signal into different frequency components;

[0018] The adaptive modal filter is used to process each modal component, enhance the signal characteristics related to mechanical equipment failure, and suppress irrelevant noise;

[0019] Reconstruct the enhanced acoustic signal.

[0020] Furthermore, the improved Waweele transform includes:

[0021] Introducing window function based on wavelet transform to enhance the ability to extract local features;

[0022] According to the characteristics of different types of mechanical faults, adaptive wavelet basis functions are constructed.

[0023] Furthermore, the process of constructing the acoustic feature dictionary includes:

[0024] Establish an original dictionary containing various typical fault characteristics;

[0025] Through online dictionary learning methods, the feature dictionary is continuously updated and optimized based on newly collected fault samples;

[0026] Use regularization parameters to control sparsity and improve the accuracy of feature representation.

[0027] Furthermore, the integrated change point detection algorithm includes:

[0028] Combining multiple change point detection methods based on entropy, density and statistical test;

[0029] Adaptively adjust the weight coefficient of each detection method based on its performance on historical data;

[0030] Dynamically adjust threshold parameters to balance detection sensitivity and false alarm rate.

[0031] Furthermore, the process of constructing the acoustic feature temporal evolution model based on the recurrent neural network includes:

[0032] Using long short-term memory networks as the basic architecture to enhance the ability to model long-term dependencies;

[0033] Introducing gated recurrent units and attention mechanisms to build a hybrid time series model;

[0034] A multi-scale fusion strategy is adopted to combine the prediction results of different time scales to improve the robustness and generalization ability of the model.

[0035] Furthermore, the chain progressive prediction method includes:

[0036] Identify the fault stage based on the current state;

[0037] Use the specific prediction model of the corresponding stage to predict the next state;

[0038] Iteratively update the current state and gradually predict the future state.

[0039] Furthermore, the stage-specific prediction method includes:

[0040] A linear model with random noise terms was used in the early failure stage;

[0041] The speed function model is used in the development fault stage;

[0042] The accelerated development stage uses a nonlinear model with acceleration function;

[0043] The asymptotic model is used in the stable or declining period.

[0044] Furthermore, the multi-dimensional warning information generated by the multi-dimensional warning decision-making process intuitively presents the entire process of fault evolution through a visual interface, and adaptively adjusts the warning threshold and notification strategy according to the urgency of the fault, including:

[0045] Fault type and location information;

[0046] Development stage information;

[0047] Projected evolutionary trajectory;

[0048] Maintenance advisory information.

[0049] The present invention provides an early warning system based on front-end perception data, which is used to execute the above-mentioned early warning method based on front-end perception data, including:

[0050] The acoustic signal processing module is used to perform modal separation and enhancement processing on the collected acoustic signals, apply the improved Waville transform to extract weak fault features, and construct an acoustic feature dictionary;

[0051] The fault stage identification module is used to identify the key stages of fault evolution using an integrated change point detection algorithm and build a stage-by-stage fault feature library;

[0052] The time series evolution modeling module is used to build an acoustic feature time series evolution model based on a recurrent neural network to identify fault evolution patterns;

[0053] A chained progressive prediction module is used to predict future fault states based on the current fault stage using a stage-specific prediction method;

[0054] Multi-dimensional early warning decision module, used to generate early warning information.

[0055] The beneficial effects of the present invention are: improved fault detection sensitivity, increased warning lead time; realized the prediction of fault development trend, improved prediction accuracy, and can predict fault development trend in advance; accurately grasped the fault evolution process, improved the change point detection accuracy and fault development trajectory prediction accuracy; optimized maintenance decisions, reduced unplanned downtime, and reduced maintenance costs; effectively extended equipment service life and improved production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of an early warning method based on front-end perception data in the present invention;

[0057] Figure 2 is a flow chart of the acoustic signal enhancement and feature extraction steps of the present invention;

[0058] Figure 3 is a flow chart of the fault stage division and feature analysis steps of the present invention;

[0059] Figure 4 is a flow chart of the steps for constructing a temporal evolution model of the present invention;

[0060] Figure 5 is a flow chart of the chained progressive prediction steps of the present invention;

[0061] Figure 6 It is a flow chart of the multi-dimensional early warning decision-making steps of the present invention. DETAILED DESCRIPTION

[0062] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0063] At least one embodiment of the present invention discloses an early warning method based on front-end perception data, such as Figures 1 to 6 As shown, including:

[0064] Step 1: Acoustic signal enhancement and feature extraction: Perform modal separation and enhancement on the collected acoustic signal, apply the improved Vavilet transform to extract weak fault features, and construct an acoustic feature dictionary;

[0065] This step mainly addresses the problem that acoustic signals in industrial environments are easily interfered with by noise and that early weak fault characteristics are difficult to identify. It includes the following sub-steps:

[0066] Step 1.1, acoustic mode separation and signal enhancement;

[0067] The acoustic mode separation algorithm is used to enhance the original acoustic signal collected from the acoustic sensor to remove environmental noise interference. The specific implementation method is as follows:

[0068] First, the original acoustic signal collected Decompose into different frequency components:

[0069] ;

[0070] in Indicates time The original acoustic signal at Indicates the acoustic modal components, represents the noise component, represents the total number of modes, Indicates from arrive The summation operation.

[0071] Then, the modal components are processed by adaptive modal filters to enhance the signal characteristics related to mechanical equipment failure and suppress irrelevant noise:

[0072] ;

[0073] in Indicates the enhanced acoustic modal components, Indicates that the adaptive filter is at frequency The frequency response function at is dynamically adjusted according to the signal-to-noise ratio:

[0074] ;

[0075] in Indicates the Mode at frequency The signal-to-noise ratio at It is an adjustment parameter used to control the cutoff characteristics of the filter.

[0076] Finally, reconstruct the enhanced acoustic signal:

[0077] ;

[0078] in represents the reconstructed enhanced acoustic signal, Indicates from arrive The summation operation.

[0079] Through this sub-step, the background noise in the industrial environment can be effectively removed, and the signal quality of subsequent processing can be improved.

[0080] Step 1.2, improved Vaviley transform and weak feature extraction;

[0081] The improved Vaviley transform is applied to perform time-frequency analysis on the enhanced acoustic signal to extract weak fault features. Traditional wavelet transform has limitations in time-frequency localization. This implementation uses the improved Vaviley transform to enhance the ability to extract weak features:

[0082] ;

[0083] in represents the Varvey coefficient, is a scale parameter that controls the frequency resolution of the analysis, is the translation parameter, controlling the time positioning, is the wavelet basis function, is a window function used to enhance the ability to extract local features. represents the complex conjugate operator, represents the integral operation from negative infinity to positive infinity, Indicates about variables The differential of .

[0084] According to the characteristics of different types of mechanical faults, adaptive wavelet basis functions are constructed:

[0085] ;

[0086] in Indicates that for Adaptive wavelet basis function of fault-like features, For the Modulation function for fault-like signature matching.

[0087] Step 1.3, acoustic feature dictionary construction;

[0088] An acoustic feature dictionary is constructed based on the sparse coding principle to efficiently represent and identify various fault features:

[0089] ;

[0090] in Represents the minimize operation, Represents the acoustic feature dictionary matrix, which contains atoms of various typical fault features. represents a sparse coefficient vector, is the regularization parameter, controlling the sparsity, represents the square of the L2 norm (square of the Euclidean distance), Represents the L1 norm (the sum of the absolute values ​​of each element).

[0091] Update and optimize the feature dictionary through online dictionary learning methods:

[0092] ;

[0093] in Indicates the The dictionary of iterations, Indicates the The dictionary of iterations, is the learning rate, which controls the step size of dictionary update, About dictionary The gradient operator, is the objective function.

[0094] Step 1.4: Matching pursuit and feature classification

[0095] Detection and classification of small acoustic feature changes using the matching pursuit algorithm:

[0096] ;

[0097] ;

[0098] ;

[0099] in represents the initial residual, which is equal to the Vavilet transformation coefficient; represents the Varvelet coefficient; and Respectively represent Second and The residual error of the iteration, Representation dictionary The atoms (basic eigenvectors), represents the inner product operation, Indicates the The atom index selected by the iteration, Indicates finding the index that makes the following expression take the maximum value .

[0100] This algorithm can decompose complex acoustic signals into a combination of basic features, achieve high-precision detection of small acoustic feature changes, and output fault feature descriptors:

[0101] ;

[0102] in Represents a set of fault feature descriptors, 、 、 Respectively represent the first, second, and Fault characteristic components, Represents the total number of feature components.

[0103] These features are used for subsequent fault stage classification and feature analysis.

[0104] Through this step, the system obtains a high-quality set of fault feature descriptors, which accurately describes the acoustic feature changes of mechanical equipment and provides basic data for subsequent fault stage division and feature analysis. In addition, the constructed acoustic feature dictionary It will also serve as an important knowledge base for subsequent analysis.

[0105] Step 2: Fault stage division and feature analysis: Based on weak fault features, an integrated change point detection algorithm is used to identify the key stages of fault evolution and build a stage-by-stage fault feature library.

[0106] This step uses the high-quality fault feature descriptors extracted in step 1 , to solve the problem that the existing system is difficult to capture the stage-by-stage characteristic changes and inflection points in the fault evolution process, and to provide a basis for subsequent time series evolution modeling and prediction by finely dividing and analyzing the characteristics of the fault development process.

[0107] Step 2.1, integrating the change point detection algorithm;

[0108] An integrated change point detection algorithm is used to identify key change points in the fault evolution process. Change point detection refers to identifying the time points in a time series when statistical characteristics change, which is crucial for accurately grasping the key turning points in the fault development process. This implementation method integrates multiple change point detection methods to improve the accuracy and robustness of detection:

[0109] ;

[0110] in represents the set of detected change points, Indicates a point in time, Indicates the A distribution difference measurement function used to measure the time window and The difference in data distribution between is the corresponding weight coefficient, is the threshold parameter, represents the total number of distribution difference measurement functions used, Represents a time window The acoustic signal data within Represents a time window The acoustic signal data within Indicates the time window length, Represents the summation symbol.

[0111] Commonly used distribution difference measurement functions include:

[0112] Entropy-based metrics:

[0113] ;

[0114] in It represents an entropy-based metric that detects distribution changes by comparing the information entropy difference between two data sets. It is suitable for capturing mutations in data complexity. Represents the information entropy function, which calculates the uncertainty of data distribution; and Represents the data sets in two time windows respectively; Represents the absolute value operator;

[0115] Density-based metrics:

[0116] ;

[0117] in It represents a density-based metric that uses KL divergence to measure the difference in probability distribution and is sensitive to changes in the shape of the distribution. represents KL divergence (relative entropy), which measures the difference between two probability distributions; and Represents the data sets and The probability density function of

[0118] Metrics based on statistical tests:

[0119] ;

[0120] in It represents a metric based on a statistical test, which is essentially a t-test statistic and is suitable for detecting changes in the mean. and Represents the data sets and The mean of and Represents the data sets and variance; and Represents the data sets and The number of samples.

[0121] By adaptively adjusting the weight and threshold , achieving accurate detection of change points under different fault types and noise conditions:

[0122] ;

[0123] in Indicates the In the iteration The weight of the method, Indicates the In the iteration The weight of the method, Indicates the The method is The error rate in the test, is the adaptive learning rate, which controls the magnitude of weight adjustment. Represents the exponential function.

[0124] Step 2.2, extracting stage fault features;

[0125] Based on the detected change points, the fault development process is divided into multiple stages:

[0126] ;

[0127] in Indicates the fault stage, including all acoustic signal data within this stage, and Represent the starting and ending change points of this stage respectively, Indicates time The acoustic signal at .

[0128] Extract and analyze the acoustic features of each stage and construct a stage-by-stage feature representation:

[0129] ;

[0130] in Indicates the The feature vector of each stage, 、 、 Respectively represent the first, second, and Feature extraction function, Represents the total number of feature extraction functions.

[0131] Feature extraction functions include time domain features (mean, variance, kurtosis, skewness, etc.), frequency domain features (power spectrum density, frequency band energy ratio, etc.) and time-frequency features (wavelet coefficients, instantaneous frequency, etc.).

[0132] Step 2.3, stage failure pattern identification;

[0133] Based on the extracted stage features, a multi-class classifier is applied to identify the failure mode of each stage:

[0134] ;

[0135] in Indicates the Fault type labels for each stage, Represents a given feature The fault type under the condition is The posterior probability of Indicates the type of fault that makes the following expression take the maximum value .

[0136] This implementation uses an ensemble learning method to improve the accuracy of fault mode recognition:

[0137] ;

[0138] in Indicates the The basic classifier predicts the fault type as The probability of represents the corresponding weight coefficient, represents the total number of base classifiers, Represents the summation symbol.

[0139] Step 2.4, constructing the phased fault feature library;

[0140] Organize the identified fault features and pattern information at each stage into a structured feature library to provide a reference for subsequent time series evolution modeling and prediction:

[0141] ;

[0142] in represents the fault signature database, Indicates the The feature vector of each stage, Indicates the corresponding fault type label, Indicates the The duration of each phase, represents the remaining time from this stage to the final failure, Indicates the total number of failure stages.

[0143] Through this step, the system establishes a structured feature database containing the feature vectors of each fault stage, fault type label, duration and remaining time, and at the same time determines the stage information of the current fault. These outputs will serve as key inputs for the construction and prediction of the time series evolution model.

[0144] Step 3: Build a time series evolution model. Based on the staged fault feature library, build an acoustic feature time series evolution model based on a recurrent neural network to identify the fault evolution law.

[0145] This step is based on the fault feature database built in step 2 and identification of stage fault characteristics , to solve the problem that the existing system lacks the ability to predict the development trend of faults, and by constructing an acoustic feature time series evolution model, dynamic modeling and characterization of the fault evolution law can be achieved.

[0146] Step 3.1, time series data preprocessing;

[0147] Preprocess the staged fault feature sequence obtained from step 2 to prepare for the construction of the time series evolution model:

[0148] ;

[0149] in represents a time series feature dataset, 、 、 Represents the first, second, and The feature vector of each time point, 、 、 Respectively represent the first, second, and The timestamp corresponding to the time point, Indicates the total number of time points.

[0150] Perform data standardization to eliminate the impact of differences in feature dimensions:

[0151] ;

[0152] in Indicates the The normalized feature vector of each time point is represents the mean vector of the feature; The standard deviation vector representing the feature. The numerator and denominator operations are element-wise operations.

[0153] The model input sequence is constructed using sliding window technology:

[0154] ;

[0155] in represents the training dataset, Indicates that from Time point to The feature sequence of time points is used as the model input. Indicates the The feature vector of each time point is used as the prediction target. Indicates the sliding window size.

[0156] Step 3.2, recurrent neural network model construction;

[0157] Based on recurrent neural networks (RNNs) and their variants, we build an acoustic feature temporal evolution model to capture the dynamic evolution of fault features over time. This implementation uses a long short-term memory network (LSTM) as the underlying architecture to enhance the modeling capabilities of long-term dependencies:

[0158] The forward calculation process of the LSTM unit is as follows:

[0159] ;

[0160] ;

[0161] ;

[0162] ;

[0163] ;

[0164] ;

[0165] in Represents the state of the input gate, which controls the extent to which new information enters the cell; Indicates the state of the forget gate, which controls the degree of retention of historical information; Represents the state of the output gate, which controls the degree of cell state output; represents the candidate cell state; represents the cell state at the current time step, represents the cell state at the previous time step; represents the hidden state output, represents the hidden state of the previous time step; Represents the input features of the current time step; 、 、 、 Represent the weight matrices of the input gate, forget gate, output gate, and candidate cell states respectively; 、 、 、 Represent the bias vectors of the input gate, forget gate, output gate, and candidate cell state respectively; Represents the sigmoid activation function; represents the hyperbolic tangent activation function; Represents a vector concatenation operation.

[0166] To enhance the model's adaptability to the evolution of fault characteristics at different stages, this implementation introduces a gated recurrent unit (GRU) and an attention mechanism to construct a hybrid time series model:

[0167] ;

[0168] ;

[0169] ;

[0170] ;

[0171] in represents the hidden state at the current time step, represents the hidden state of the previous time step; Indicates the update gate state, which controls the fusion ratio of historical information and new information; Indicates the reset gate status, which controls the degree of reset of historical information; represents the candidate hidden state; 、 、 Represent the weight matrices of the update gate, candidate hidden state, and reset gate, respectively. 、 、 Represent the bias vectors of the update gate, candidate hidden state, and reset gate respectively; Represents the sigmoid activation function; represents the hyperbolic tangent activation function.

[0172] The calculation process of the attention mechanism is as follows:

[0173] ;

[0174] ;

[0175] ;

[0176] in Represents the time step Hide state from history The attention score, represents the attention vector, Represents a vector The transpose of and Represent the weight matrices of the historical hidden state and the current hidden state in the attention calculation, represents the normalized attention weight, represents the context vector, represents the exponential function, Indicates from 1 to The summation operation, Represents the time step Hide state from history Attention score.

[0177] Finally, the output layer of the model is predicted through a fully connected network:

[0178] ;

[0179] in Represents the predicted feature vector for the next moment, represents the output layer weight matrix, represents the output layer bias vector, Represents the concatenation of the hidden state and context vector.

[0180] Step 3.3, time-frequency analysis and parameter estimation;

[0181] Based on the time series evolution model, the time-frequency analysis method is applied to identify the characteristic parameters of fault evolution, further improving the model's ability to express the laws of fault development:

[0182] First, perform time-frequency analysis on historical fault data to extract key parameters of fault evolution:

[0183] ;

[0184] in represents the set of fault evolution parameters, 、 、 Respectively represent the first, second, and parameters, Indicates the total number of parameters.

[0185] Then, recursive least squares (RLS) is applied for online parameter estimation:

[0186] ;

[0187] in Represents the time step The parameter estimates of represents the forgetting factor ( ), control the influence weight of historical data; represents a parameterized model, represents the actual observed value, represents the input features, Indicates the parameter that makes the following expression take the minimum value .

[0188] The recursive formula for parameter update is:

[0189] ;

[0190] ;

[0191] ;

[0192] in represents the gain vector, which controls the step size of parameter update. represents the covariance matrix, reflecting the uncertainty of parameter estimation, represents the covariance matrix of the previous time step, represents the parameter estimate at the previous time step, Represents a vector The transpose of .

[0193] Step 3.4, multi-scale fusion and model optimization;

[0194] In order to improve the robustness and generalization ability of the time series evolution model, this implementation adopts a multi-scale fusion strategy to combine the prediction results of different time scales:

[0195] ;

[0196] in represents the final fusion prediction result, Indicates the The prediction results of the time scale model are represents the corresponding fusion weight, Represents the total number of time-scale models.

[0197] The fusion weights are dynamically adjusted through an adaptive algorithm:

[0198] ;

[0199] in Indicates the In the iteration The fusion weight of the model, Indicates the In the iteration The fusion weight of the model, Indicates the The model at time step The prediction error of is the learning rate, which controls the magnitude of weight adjustment.

[0200] Through the above steps, the constructed time series evolution model can dynamically capture the evolution law of fault characteristics and provide a basis for the next step of chain progressive prediction.

[0201] Step 4: Chain progressive prediction: Based on the acoustic feature temporal evolution model and the stage-specific fault feature library, a stage-specific prediction method is used to predict future fault states.

[0202] This step comprehensively utilizes the current fault stage information identified in step 2 and the time-series evolution model constructed in step 3 to solve the problem that the existing system has difficulty in accurately predicting the fault development trajectory. By establishing a chain-type progressive prediction algorithm, it can achieve accurate prediction of future fault states and provide forward-looking guidance for maintenance decisions.

[0203] Step 4.1, progressive prediction method based on the current stage;

[0204] Based on the current fault stage characteristics identified in step 2, a chain progressive prediction method is used to predict the future fault state. This implementation method first identifies the stage of the current fault:

[0205] ;

[0206] in represents the current fault stage discriminant function, Indicates the current fault status. Indicates a given current state Under the conditions, The probability of each stage, Indicates the parameter that makes the following expression take the maximum value .

[0207] Then, the corresponding prediction model is constructed based on the characteristics of different fault stages:

[0208] ;

[0209] in Indicates the predicted future The fault status at the moment, Indicates the current fault stage The corresponding prediction function is, represents the time step of the prediction, Indicates the current fault status. and Represent the predicted next moment and future respectively Fault status at the moment.

[0210] The specific progressive prediction process is as follows:

[0211] Initialize and set the current state and prediction step length ;

[0212] Progressive prediction, for time steps From 1 to ;

[0213] According to the current status Identify the fault stage ;

[0214] Use the prediction model for the corresponding stage Predict the next state ;

[0215] Update current status ;

[0216] Output the prediction results, ;

[0217] By adopting this progressive approach, the error of short-term prediction can be controlled at a low level, and accurate prediction of long-term status can be achieved through multi-step iteration.

[0218] Step 4.2, stage-specific prediction model construction;

[0219] Build specialized prediction models based on the characteristics of different fault stages to improve prediction accuracy and adaptability:

[0220] For the early failure stage (weak characteristic stage):

[0221] ;

[0222] in represents the prediction function for the early failure stage, represents the weight matrix at the early stage, Indicates the current fault status. represents the bias vector, represents the noise term, which is used to simulate the randomness and fluctuation of early fault development.

[0223] For the developmental failure phase (linear growth phase):

[0224] ;

[0225] in represents the prediction function of the failure stage during the development period, Indicates the current fault status. represents the time step, Represents the fault evolution speed function, which changes dynamically with the fault status.

[0226] For the accelerated development phase (non-linear growth phase):

[0227] ;

[0228] in represents the prediction function of the accelerated development stage, Indicates the current fault status. represents the time step, represents the fault evolution speed function, represents the fault evolution acceleration function, represents the square of the time step, represents the coefficient of the second-order term.

[0229] For stable or declining periods:

[0230] ;

[0231] in represents the prediction function of the stable period or recession period, Represents the attenuation coefficient (range is 0 to 1), Indicates the current fault status. represents the limit state, The weight coefficient representing the limit state.

[0232] Step 4.3, similar case matching and reference;

[0233] Combined with similar failure cases in the historical database, the prediction results are optimized and corrected:

[0234] First, retrieve cases with similar fault characteristics to the current one from the historical database:

[0235] ;

[0236] in Indicates the current status and historical cases The similarity of and Represent the feature vectors of the current fault state and historical fault cases respectively, is the scale parameter for similarity calculation, represents the square of the Euclidean distance between two eigenvectors, Represents the exponential function.

[0237] Then, based on the weighted average of similarity, the prediction results are optimized:

[0238] ;

[0239] in represents the optimized prediction result, Indicates the impact weight of historical cases (range is 0 to 1), represents the original prediction result, represents the total number of historical cases, Indicates the current state and The similarity of historical cases, represents the sum of all similarities (for normalization), Indicates historical cases In the future The state of the moment, Indicates from 1 to The summation operation.

[0240] Step 4.4, fault severity assessment and remaining life prediction;

[0241] Based on the predicted future fault status, fault severity assessment and remaining service life prediction are performed:

[0242] Fault severity evaluation function:

[0243] ;

[0244] in represents the fault severity evaluation function, represents the total number of evaluation indicators, Indicates the The weight coefficient of the evaluation index, Indicates the An evaluation indicator function, Indicates from 1 to The summation operation.

[0245] Remaining useful life forecast:

[0246] ;

[0247] in Indicates that from the current state Remaining useful life at the start, Indicates the minimum operation. represents the time step, Indicates the fault severity of the predicted state, Indicates the fault severity threshold, exceeding which the device cannot continue to operate safely.

[0248] Through this step, an accurate prediction of the future development status of the fault is achieved, providing a scientific basis for fault handling and maintenance decisions.

[0249] Step 5: Multi-dimensional early warning decision-making, based on the predicted future fault status, generates early warning information to assist maintenance decision-making;

[0250] This step is based on the future fault status, fault severity, and remaining service life predicted in step 4, combined with the fault type and stage characteristics identified in step 2. It solves the problem that traditional early warning systems output single information and lack specific guidance. By building a multi-dimensional early warning decision-making system, it provides maintenance personnel with comprehensive and intuitive decision-making support information.

[0251] Step 5.1, multi-dimensional construction of early warning information;

[0252] Build a multi-dimensional early warning information system based on fault type, development speed and potential impact:

[0253] First, determine the warning level based on the predicted fault status and development trajectory:

[0254] ;

[0255] in represents the warning level function, Indicates the current fault status. represents the predicted future fault state, represents the fault severity function, represents the remaining useful life function, 、 、 Indicates the severity thresholds for minor faults, moderate faults, and major faults. 、 、 Represent the service life thresholds of long-term surplus, medium-term surplus, and short-term surplus, respectively. 、 、 、 Indicates the 1st, 2nd, 3rd, and 4th warning levels, Represents the logical AND operator, Represents the logical OR operator.

[0256] Then, a multi-dimensional early warning information package is constructed, including:

[0257] ;

[0258] ;

[0259] ;

[0260] ;

[0261] in Indicates a collection of fault information. Indicates the fault type. Indicates the fault location, Indicates the affected component; Represents a set of development stage information, Indicates the current stage, Indicates the next stage, Indicates the phase transition time; represents the set of expected evolutionary trajectories, 、 、 Respectively represent the future 、 、 Predicted fault status at the moment, Indicates the total number of predicted time points; Indicates a collection of maintenance suggestion information. Indicates the recommended maintenance time window. Indicates the maintenance priority, Indicates the required resources, Indicates a maintenance procedure.

[0262] Finally, a comprehensive early warning report is generated:

[0263] ;

[0264] in Indicates comprehensive early warning report, Indicates the warning level. Indicates a collection of fault information. Represents a set of development stage information, represents the set of expected evolutionary trajectories, Indicates a collection of maintenance suggestion information.

[0265] Step 5.2, visualization interface and interactive system;

[0266] The generated multi-dimensional warning information is presented intuitively through a visual interface to assist maintenance decision-making:

[0267] Fault evolution trajectory visualization: Uses multiple visualization methods such as time series graphs and phase space trajectory graphs to display the complete evolution process of the fault from the current state to the predicted future state.

[0268] Dynamic monitoring of key parameters: Real-time display of changing trends of key fault parameters, including characteristic frequency, energy distribution, pattern similarity and other indicators, and marking of key change points and threshold lines.

[0269] Maintenance decision-making support interface: Based on the predicted failure development trend, it provides a variety of maintenance options and displays the cost, risk and benefit analysis of each option.

[0270] The system design meets the following interaction requirements:

[0271] Multi-level information display: supports multi-level information exploration from overview to details

[0272] Real-time update and historical backtracking: taking into account both real-time data monitoring and historical evolution trajectory backtracking

[0273] Customized views and alerts: Provide customized information display based on user roles and concerns

[0274] Step 5.3: Adaptive optimization of early warning strategy

[0275] To adapt to the needs of different industrial environments and equipment types, this implementation introduces an adaptive optimization mechanism for early warning strategies:

[0276] ;

[0277] in express A constant early warning strategy, express A constant early warning strategy, Indicates user feedback and system performance evaluation results, represents the policy update function.

[0278] Optimization content includes:

[0279] Dynamic adjustment of warning thresholds: Based on equipment operation history and failure cases, adaptively adjust the warning thresholds of fault severity and remaining life

[0280] ;

[0281] in Indicates the The threshold level is The value of the moment, Indicates the The threshold level is The value of the moment, To adjust the step size, express The false alarm rate at each moment, express The detection rate at the time, Indicates the threshold adjustment factor.

[0282] Optimize the frequency and method of early warning: Optimize the frequency and notification method of early warning according to the urgency of the fault and the user response mode

[0283] ;

[0284] in represents the warning frequency function, Indicates the warning level. 、 、 、 Respectively represent the corresponding warning levels 、 、 、 The notification frequency will be adjusted as the system runs smoothly.

[0285] Step 5.4, early warning decision system integration and interface;

[0286] To achieve seamless integration with existing industrial systems, this implementation has designed a standardized early warning decision interface:

[0287] Data interface, supporting receiving device status data from multiple sources, including acoustic sensors, vibration sensors, temperature sensors, etc.

[0288] ;

[0289] in Represents a collection of data interfaces. Represents sensor data, Represents process data, Represents historical data.

[0290] Control interface, providing the ability to interact with the upper control system and the lower execution unit, supporting automatic control and manual intervention

[0291] ;

[0292] in Represents a collection of control interfaces. Indicates automatic control function, Indicates the manual intervention function, Indicates emergency response function.

[0293] Information interface to achieve data sharing and collaborative decision-making with enterprise resource planning, manufacturing execution system and other systems

[0294] ;

[0295] in Represents a collection of information interfaces. Indicates the maintenance plan, Indicates resource allocation, Represents production scheduling.

[0296] Through this step, multi-dimensional construction and intuitive display of fault warning information are achieved, providing a comprehensive scientific basis for maintenance decisions and improving the practicality and decision-making support capabilities of the warning system.

[0297] An early warning system based on front-end perception data, used to execute the above-mentioned early warning method based on front-end perception data, comprising:

[0298] The acoustic signal processing module is used to perform modal separation and enhancement processing on the collected acoustic signals, apply the improved Waville transform to extract weak fault features, and construct an acoustic feature dictionary;

[0299] The fault stage identification module is used to identify the key stages of fault evolution using an integrated change point detection algorithm and build a stage-by-stage fault feature library;

[0300] The time series evolution modeling module is used to build an acoustic feature time series evolution model based on a recurrent neural network to identify fault evolution patterns;

[0301] A chained progressive prediction module is used to predict future fault states based on the current fault stage using a stage-specific prediction method;

[0302] Multi-dimensional early warning decision module, used to generate early warning information.

[0303] Here, the present invention provides an implementation example:

[0304] A bearing fault early warning system was implemented in the rolling mill production line of a large steel mill. The production line operates in a complex environment with diverse noise sources, including mechanical collisions, fluid noise, and electromagnetic interference. Traditional vibration sensors lacked sensitivity for early bearing fault detection in this environment and were unable to effectively predict failure development trends.

[0305] The steel mill rolling production line faces the following major problems: the complex noise environment seriously interferes with the extraction of fault characteristics; the bearing fault types are diverse (inner ring cracks, rolling element wear, cage damage, etc.); the unplanned downtime is costly (losses of approximately 150,000 yuan per hour); maintenance resources are limited, and accurate prediction of fault development is required to rationally plan maintenance.

[0306] Based on this implementation, an acoustic sensor network was deployed on key equipment in the steel rolling production line. The sensor installation locations were carefully designed to ensure accurate and comprehensive sound collection. The acoustic sensors worked in parallel with existing vibration sensors to verify system performance and provide complementary data. The acoustic sensor network deployment and key implementation steps are shown in Table 1.

[0307] Table 1: Overview of acoustic sensor networks and implementations

[0308]

[0309] In monitoring a bearing in a finishing mill, the original acoustic signal was subject to strong ambient noise interference, resulting in a low signal-to-noise ratio. Applying an acoustic mode separation algorithm and an improved Vavilet transform, we successfully extracted subtle fault features. Table 2 shows a comparison of the signal processing results.

[0310] Table 2: Comparison of signal processing effects

[0311]

[0312] The acoustic feature dictionary constructed based on sparse coding contains various typical fault characteristics of bearings. The matching pursuit algorithm is used to successfully detect early tiny cracks in the inner ring of the bearing. This fault is completely concealed by traditional vibration monitoring.

[0313] The system uses an integrated change point detection algorithm to divide the development process of the bearing inner ring crack fault into four key stages. It then uses a stage-specific prediction model, combined with similar cases in the historical database, to generate an accurate fault development trajectory and remaining service life prediction. Table 3 shows the fault stage division and prediction results, as shown in Table 3:

[0314] Table 3: Fault stage classification and prediction results

[0315]

[0316] Based on the prediction results, the system generates multi-dimensional warning decision information to provide a scientific basis for maintenance personnel. The warning levels and maintenance recommendations for different fault stages are shown in Table 4:

[0317] Table 4: Warning levels and maintenance recommendations

[0318]

[0319] In this case, the system detected a tiny crack in the bearing inner ring 25 days before failure, enabling the maintenance department to replace the faulty bearing in advance during a planned downtime, avoiding significant losses caused by unplanned downtime.

[0320] After 12 months of actual operation, this implementation method has achieved technical results. The following is the verification data of the two most critical technical results:

[0321] The comparison of fault detection performance is shown in Table 5:

[0322] Table 5: Fault detection performance comparison

[0323]

[0324] The statistics of maintenance decision optimization effect are shown in Table 6:

[0325] Table 6: Statistics on maintenance decision optimization effect

[0326]

[0327] The above data demonstrates that this implementation improves fault detection sensitivity, shortening the average fault warning time from 7.6 days to 22.1 days, a 190.8% improvement. In terms of optimizing maintenance decisions, the system reduced unplanned downtime by 54.0%, lowered maintenance costs by 37.6%, and increased equipment availability to 97.6%.

[0328] These technical effects fully verify the effectiveness and practical value of this implementation in actual industrial environments, bringing economic benefits to enterprises.

[0329] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. An early warning method based on front-end perception data, characterized in that: include: Acoustic signal enhancement and feature extraction: Perform modal separation and enhancement processing on the collected acoustic signals, apply the improved Vaviley transform to extract weak fault features, build an acoustic feature dictionary, and use the improved Vaviley transform to enhance the ability to extract weak features: ； in represents the Varvelet coefficient, is a scale parameter that controls the frequency resolution of the analysis, represents the reconstructed enhanced acoustic signal, is the translation parameter, controlling the time positioning, is the wavelet basis function, is a window function used to enhance the ability to extract local features. represents the complex conjugate operator, represents the integral operation from negative infinity to positive infinity, Indicates about variables The differential of According to the characteristics of different types of mechanical faults, adaptive wavelet basis functions are constructed: ； in Indicates that for Adaptive wavelet basis function of fault-like features, For the Modulation function for fault-like feature matching; Fault stage division and feature analysis: Based on weak fault features, an integrated change point detection algorithm is used to identify the key stages of fault evolution and build a stage-by-stage fault feature library. The integrated change point detection algorithm includes: Combining multiple change point detection methods based on entropy, density and statistical test; Adaptively adjust the weight coefficient of each detection method based on its performance on historical data; Dynamically adjust threshold parameters to balance detection sensitivity and false alarm rate; The time series evolution model is constructed. For the staged fault feature library, an acoustic feature time series evolution model is constructed based on a recurrent neural network to identify the fault evolution law. The process of constructing the acoustic feature time series evolution model based on a recurrent neural network includes: Using long short-term memory networks as the basic architecture to enhance the ability to model long-term dependencies; Introducing gated recurrent units and attention mechanisms to build a hybrid time series model; Adopting a multi-scale fusion strategy to combine prediction results at different time scales to improve the robustness and generalization ability of the model; Chain progressive prediction uses a stage-specific prediction method to predict future fault states based on the acoustic feature time-series evolution model and the stage-by-stage fault feature library. The chain progressive prediction method includes: Identify the fault stage based on the current state; Use the specific prediction model of the corresponding stage to predict the next state; Iteratively update the current state and gradually predict the future state; The stage-specific prediction method includes: A linear model with random noise terms was used in the early failure stage; The speed function model is used in the development fault stage; The accelerated development stage uses a nonlinear model with acceleration function; The asymptotic model is used for the stable or declining period; Multi-dimensional early warning decision-making generates early warning information based on predicted future fault states to assist maintenance decisions.

2. The early warning method based on front-end perception data according to claim 1 is characterized in that: The process of modal separation and enhancement includes: Decompose the collected original acoustic signal into different frequency components; The adaptive modal filter is used to process each modal component, enhance the signal characteristics related to mechanical equipment failure, and suppress irrelevant noise; Reconstruct the enhanced acoustic signal.

3. The early warning method based on front-end perception data according to claim 1 is characterized in that: The improved Waweeley transform includes: Introducing window function based on wavelet transform to enhance the ability to extract local features; According to the characteristics of different types of mechanical faults, adaptive wavelet basis functions are constructed.

4. The early warning method based on front-end perception data according to claim 1 is characterized in that: The process of constructing the acoustic feature dictionary includes: Establish an original dictionary containing various typical fault characteristics; Through online dictionary learning methods, the feature dictionary is continuously updated and optimized based on newly collected fault samples; Use regularization parameters to control sparsity and improve the accuracy of feature representation.

5. The early warning method based on front-end sensing data according to claim 1 is characterized in that: The multi-dimensional warning information generated by the multi-dimensional warning decision-making process intuitively presents the entire process of fault evolution through a visual interface, and adaptively adjusts the warning threshold and notification strategy according to the urgency of the fault, including: Fault type and location information; Development stage information; Projected evolutionary trajectory; Maintenance advisory information.

6. An early warning system based on front-end perception data, characterized in that: A method for performing an early warning method based on front-end perception data according to any one of claims 1 to 5, comprising: The acoustic signal processing module is used to perform modal separation and enhancement processing on the collected acoustic signals, apply the improved Waville transform to extract weak fault features, and construct an acoustic feature dictionary; The fault stage identification module is used to identify the key stages of fault evolution using an integrated change point detection algorithm and build a stage-by-stage fault feature library; The time series evolution modeling module is used to build an acoustic feature time series evolution model based on a recurrent neural network to identify fault evolution patterns; A chained progressive prediction module is used to predict future fault states based on the current fault stage using a stage-specific prediction method; Multi-dimensional early warning decision module, used to generate early warning information.

Images