Electric hand drill wear state prediction and health management system

Abstract

The invention relates to an electric hand drill wear state prediction and health management system, which belongs to the technical field of intelligent fault diagnosis and predictive maintenance of industrial equipment, and comprises a data acquisition and preprocessing unit used for acquiring and processing a multi-modal physical signal to generate a standardized data frame; the multi-domain feature transformation unit is used for receiving the standardized data frame and transforming the standardized data frame into a health feature vector and a load feature vector; the dynamic health baseline construction unit is used for reconstructing and generating a dynamic health baseline through a depth generation model according to the time sequence of the health feature vector and the load feature vector; and the residual error sequence generation and statistical monitoring unit is used for calculating the distance between the health feature vector and the dynamic health baseline, generating a residual error sequence, and performing statistical processing on the residual error sequence to obtain a statistical magnitude. According to the invention, the interference of working condition change on health state assessment is eliminated, and pure and reliable data input is provided for subsequent accurate monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent fault diagnosis and predictive maintenance of industrial equipment, in particular to a hand drill wear state prediction and health management system. BACKGROUND

[0002] In the health monitoring of mechanical equipment such as hand drills, traditional analysis methods mainly rely on the analysis of physical signals such as vibration and current generated during operation. However, the characteristics of these signals are not only related to the wear and degradation state of the equipment itself, but also greatly affected by external working loads such as drilling material, feed speed, and torque size.

[0003] This situation leads to the following core problems that traditional methods are difficult to solve:

[0004] The working condition and fault characteristics are confused, and the traditional method is difficult to effectively distinguish whether the signal change is caused by normal working load change or by early wear and degradation of internal parts such as bearings and gears. This coupling and confusion of characteristics is the root cause of false positives or false negatives of the monitoring system.

[0005] Lack of dynamic adaptability, due to the lack of a health reference standard that can dynamically adjust to working conditions, i.e. health baseline, the accuracy and reliability of the state evaluation of the system will significantly decrease when facing complex and variable actual working environments; for example, normal operation under high load may be misjudged as a fault, while early weak fault signals under low load may be submerged in normal fluctuations.

[0006] Insufficient prediction accuracy, because the influence of working conditions cannot be accurately stripped to quantify the true degradation, the prediction of the remaining useful life of the equipment by traditional methods is often not accurate enough to effectively support predictive maintenance decisions.

[0007] The above problems lie in the fact that traditional data processing and analysis techniques cannot effectively decouple external working condition changes and internal health degradation, which are two different types of factors from the highly coupled original signals; as a result, when the working conditions of the hand drill change frequently, managers cannot obtain stable and reliable health status evaluation and accurate life prediction, which greatly limits the implementation of proactive maintenance strategies, making it difficult to effectively avoid sudden failures, prolong the service life of equipment, and ensure job safety. SUMMARY

[0009] The purpose of the present application is to provide a hand drill wear state prediction and health management system to solve the problems raised in the background art.

[0010] The technical solution of the present application is as follows:

[0011] a data acquisition and preprocessing unit configured to acquire multi-modal physical signals and process the acquired signals to generate standardized data frames;

[0012] a multi-domain feature transformation unit configured to receive the standardized data frames and transform the standardized data frames into a health feature vector and a load feature vector;

[0013] a dynamic health baseline construction unit configured to reconstruct a dynamic health baseline from a time series of the health feature vector and the load feature vector by using a deep generative model;

[0014] a residual sequence generation and statistical monitoring unit configured to calculate a distance between the health feature vector and the dynamic health baseline, generate a residual sequence, and perform statistical processing on the residual sequence to obtain a statistical quantity;

[0015] a parameter optimization and life prediction unit configured to determine a control limit parameter according to a preset optimization target;

[0016] an instruction generation and control execution unit configured to determine a control limit according to the control limit parameter, compare the statistical quantity with the control limit, generate a diagnosis instruction and perform a control operation if the statistical quantity exceeds the control limit, and determine that the system is in a normal state if the statistical quantity does not exceed the control limit.

[0017] Preferably, the process of transforming the standardized data frames by the multi-domain feature transformation unit comprises:

[0018] firstly, calculating time domain features, frequency domain features and time-frequency domain features of the standardized data frames in parallel;

[0019] then, combining the calculated time domain features, frequency domain features and time-frequency domain features to generate a high-dimensional original feature vector;

[0020] finally, applying a nonlinear manifold learning mapping to the high-dimensional original feature vector to decompose the high-dimensional original feature vector into the health feature vector and the load feature vector.

[0021] Preferably, the process of applying the nonlinear manifold learning mapping comprises:

[0022] finding K nearest neighbors for each high-dimensional original feature vector;

[0023] calculating optimal weights by which the high-dimensional original feature vector can be linearly reconstructed from the K nearest neighbors;

[0024] finding a set of coordinates in a low-dimensional space such that the reconstruction error is minimized while maintaining the same optimal weight relationship, thereby completing the decomposition.

[0025] Preferably, the deep generative model is a variational autoencoder model with a time convolution attention mechanism.

[0026] The training process of the model is performed by minimizing an objective function composed of reconstruction error and KL divergence between posterior distribution and prior distribution;

[0027] The data used for training is only healthy samples collected under normal working conditions of the electric drill.

[0028] Preferably, the process of residual sequence generation and statistical processing of the residual sequence by the residual sequence generation and statistical monitoring unit is as follows:

[0029] An exponential weighted moving average calculation is applied to the residual sequence;

[0030] The current statistical value is recursively combined with the current residual value through a smoothing factor to generate the current statistical value.

[0031] Preferably, the process of parameter optimization and life prediction by the parameter optimization and life prediction unit is as follows:

[0032] Setting the maximum working condition-residual separation degree as the optimization objective;

[0033] Setting the control limit parameter as a control limit width coefficient;

[0034] Using a Bayesian optimization process to find the control limit width coefficient that maximizes the working condition-residual separation degree.

[0035] Preferably, the process of calculating the working condition-residual separation degree includes:

[0036] First, the kernel density estimation method is applied within a sliding time window to obtain the residual probability density function under different working condition states;

[0037] Then, based on the discrete form of K-L divergence, the first separation degree between the probability distributions of the fault working condition and the healthy working condition is calculated;

[0038] Then, the second separation degree of the probability distributions between different healthy working conditions is calculated;

[0039] Finally, the working condition-residual separation degree is determined through the ratio relationship between the first separation degree and the second separation degree.

[0040] Preferably, the parameter optimization and life prediction unit is also used for predicting the remaining useful life;

[0041] The prediction process specifically includes:

[0042] A monotonically increasing degradation state variable is constructed by time accumulation of the statistical value generated by the residual sequence generation and statistical monitoring unit;

[0043] The evolution process of the degradation state variable is modeled as a Wiener process with drift;

[0044] The Wiener process is defined by a drift coefficient representing the average rate of degradation and a diffusion coefficient describing the uncertainty;

[0045] The remaining useful life is obtained by calculating the expected time for the degradation state variable to first reach the preset failure threshold from the current value.

[0046] The present application provides a handheld drill wear state prediction and health management system, which has the following improvements and advantages compared with the prior art:

[0047] 1. The significant progress of the technical solution is that the independent features corresponding to the equipment health state and external operating load can be accurately separated and refined from the complex original signal; the system generates a high-dimensional original feature vector containing time domain, frequency domain and time-frequency domain information by performing multi-domain transformation on the collected multi-modal physical signal; unlike the direct analysis of mixed signals in traditional methods, the present application innovatively applies nonlinear manifold learning mapping to the high-dimensional vector, successfully decomposing it into a load feature vector and a health feature vector; this orthogonal feature decomposition eliminates the interference of working condition changes on health state evaluation, providing pure and reliable data input for subsequent accurate monitoring, and achieving a qualitative improvement compared with the prior art;

[0048] 2. The present application uses a deep generative model to construct a health baseline dynamically adjusted according to the working condition for the health feature vector based on the real-time separated load feature vector; the baseline represents the feature performance that the equipment should have when it is in an ideal healthy state under the current specific load; by calculating the residual between the real health feature vector and the dynamic baseline, the system can quantify the deviation caused by the degradation of the equipment itself; this design makes the health evaluation standard no longer static, but can adapt to external operating conditions in real time, thereby greatly improving the accuracy and robustness of fault detection;

[0049] 3. The present application also has significant improvement in the sensitivity and reliability of anomaly detection; the system performs statistical processing on the generated residual sequence using an exponentially weighted moving average method, which can effectively smooth short-term random noise while amplifying long-term and weak degradation trends caused by gradual wear, achieving sensitive capture of early fault signals; more importantly, the system discards the reliance on empirical manual parameter setting and instead uses a Bayesian optimization process to adaptively determine the optimal control limit parameter with the goal of maximizing the working condition-residual separation degree; this data-driven optimization mechanism ensures that the fault discrimination boundary is always at the best position to distinguish between healthy and fault states, improving the overall diagnostic performance of the system from a global perspective;

[0050] 4. The scheme realizes scientific prediction of the remaining service life by modeling the degradation process of the device as a Wiener process with drift; a monotone increasing degradation state variable is constructed by time accumulation of the monitoring statistics, and the expected time of the first arrival at the failure threshold is calculated based on the random process theory, providing a strong quantitative basis for predictive maintenance decision; the technical scheme realizes a complete technical closed loop from signal decoupling, dynamic baseline construction to parameter self-optimization and life prediction through the cooperative work of each unit, and the health management and prediction results provided are more accurate and reliable, which has significant technical progress and practical value. BRIEF DESCRIPTION OF DRAWINGS

[0052] The application will be further explained in conjunction with the accompanying drawings and embodiments:

[0053] Figure 1 is a flow chart of a hand drill wear state prediction and health management system of the application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in conjunction with specific embodiments.

[0056] Embodiment 1

[0057] Please refer to Figure 1 , the application provides a hand drill wear state prediction and health management system, comprising:

[0058] A data acquisition and preprocessing unit is used for acquiring and processing multi-modal physical signals to generate standardized data frames;

[0059] A multi-domain feature transformation unit is used for receiving the standardized data frames and transforming them into health feature vectors and load feature vectors;

[0060] A dynamic health baseline construction unit is used for reconstructing a dynamic health baseline through a deep generative model according to the time series of the health feature vectors and the load feature vectors;

[0061] A residual sequence generation and statistical monitoring unit is used for calculating the distance between the health feature vectors and the dynamic health baseline, generating a residual sequence, and statistically processing the residual sequence to obtain statistics;

[0062] A parameter optimization and life prediction unit is used for determining the control limit parameters according to a preset optimization target;

[0063] An instruction generating and control executing unit is configured to determine a control limit according to the control limit parameter, compare the statistical quantity with the control limit, generate a diagnosis instruction and execute a control operation if the statistical quantity exceeds the control limit, and determine a normal state if the statistical quantity does not exceed the control limit.

[0064] The embodiment discloses a handheld power drill wear state prediction and health management system; the system builds a complete technical closed loop from bottom signal collection to top decision control, aiming to evaluate the health state of the handheld power drill in real time and accurately, separate the influence of working load change and internal wear degradation of the equipment, and predict the remaining service life;

[0065] A data acquisition and preprocessing unit aims to obtain original physical signals that can comprehensively reflect the running state of the handheld power drill, and convert them into a standardized format suitable for deep analysis; in the embodiment, the unit realizes synchronous acquisition of multi-modal physical signals through a sensor array built in the handheld power drill; the array specifically includes: a Hall sensor for measuring three-phase drive current, a three-axis micro-electromechanical acceleration sensor for capturing multi-point vibration information of the machine shell, and a thermocouple for monitoring the temperature of the motor winding; the sampling frequency Refers to the number of data points collected per unit time, which ensures that the key characteristics of load transients and early faults can be captured, and the source is the system parameters preset according to the Nyquist theorem and experimental verification; the discrete time sequence signals collected are represented as an instantaneous multi-modal signal vector:

[0066]

[0067] Wherein, : instantaneous multi-modal signal vector, : discrete time step, : total number of sensor channels; then, a signal preprocessing process is performed, which includes a digital band-pass filter that eliminates power frequency interference and high-frequency random noise; then, normalization processing is performed to map the signal amplitudes of different physical quantities to a unified interval, such as [0, 1], thereby generating a standardized data frame for subsequent unit analysis; : real set;

[0068] The normalization processing here usually adopts the minimum-maximum normalization method, which aims to eliminate the numerical range difference caused by different physical quantities of different sensor signals, and ensure that each feature dimension obtains equal initial weight in subsequent model training, and the calculation formula of the process is:

[0069]

[0070] Wherein, is the original data point, and respectively the minimum and maximum value of the signal in a sample period, is the data point after normalization located in the interval ;

[0071] a multi-domain feature transformation unit, which aims to convert the pre-processed time series information into numerical representations that can accurately distinguish between the working load and the health state; the unit receives the aforementioned standardized data frames and transforms them into health feature vectors and load feature vectors;

[0072] a dynamic health baseline construction unit, which aims to learn and construct a health state reference standard that can be dynamically adjusted according to the working load; in this embodiment, the unit reconstructs a dynamic health baseline through a deep generative model based on the time series of health feature vectors and load feature vectors generated by the multi-domain feature transformation unit; the baseline represents the state characteristics that a fully healthy electric drill should have under the current specific load;

[0073] a residual sequence generation and statistical monitoring unit, which aims to quantify the deviation of the current electric drill health state from the ideal health state and perform statistical monitoring to detect abnormalities; the unit first calculates the Euclidean distance between the health feature vector and the dynamic health baseline, which is defined as the residual , which measures the real-time health state deviation; the calculation process is as follows:

[0074]

[0075] wherein, : the real-time health feature vector output by the multi-domain feature transformation unit; : the health baseline vector reconstructed by the dynamic health baseline construction unit; : the Euclidean distance or L2 norm;

[0076] By continuous calculation, a residual sequence is generated; the unit performs statistical processing on the residual sequence to obtain a statistical quantity , which is used to smooth the residual sequence to enhance the sensitivity to weak and continuous degradation trends;

[0077] a parameter optimization and life prediction unit, which aims to adaptively determine the key monitoring parameters of the system and predict the remaining useful life of the tool; in this embodiment, the unit determines the control limit parameters for abnormality judgment based on a preset optimization objective, i.e., the maximum load-residual separation degree; at the same time, the unit also predicts the remaining useful life based on the statistical monitoring results;

[0078] The instruction generation and control execution unit aims to transform the results of analysis and prediction into specific device control or user warning actions; this unit first sets the control limit according to the control limit parameters determined by the parameter optimization and life prediction unit ; The threshold refers to the threshold for determining whether the statistical quantity is abnormal, and its role is to distinguish between normal device state fluctuations and potential fault signals. The source is the data collected based on the operation of the device under various health conditions, and the threshold is calculated by statistical methods. This unit compares the aforementioned statistical quantity with the control limit in real time; if the value of continuously exceeds , it is determined that the tool may have entered an early failure state, and the system generates a diagnostic instruction containing information such as fault type and confidence, and performs corresponding control operations, such as lighting a specific color LED indicator light, providing tactile feedback through a built-in vibrator, or actively limiting the maximum output torque of the motor in severe cases; if does not exceed , it is determined to be in a normal state, and the system continues to monitor;

[0079] Through the cooperative work of the above-mentioned units, this embodiment realizes a prediction and management system that can adapt to external operating condition changes and accurately monitor internal health states. It overcomes the technical difficulties of confusing operating condition changes and fault characteristics in traditional methods, and can provide real-time and dynamic health baselines as a reference, thereby significantly improving the accuracy and reliability of early fault detection. Ultimately, through accurate life prediction and timely warning control, it effectively supports predictive maintenance decisions, prolongs the life of the device, and ensures operation safety.

[0080] The process of transforming the standardized data frame by the multi-domain feature transformation unit includes:

[0081] First, the time domain features, frequency domain features, and time-frequency domain features of the standardized data frame are calculated in parallel;

[0082] Then, the calculated time domain features, frequency domain features, and time-frequency domain features are combined to generate a high-dimensional original feature vector;

[0083] Finally, the high-dimensional original feature vector is applied to a nonlinear manifold learning mapping to decompose into a health feature vector and a load feature vector;

[0084] In this embodiment, the process of transforming the standardized data frame by the multi-domain feature transformation unit is further limited, aiming to extract information from multiple dimensions and perspectives to construct a comprehensive feature space that is highly sensitive to load and health state changes;

[0085] To achieve this, the unit performs parallel computation of time-domain, frequency-domain, and time-frequency-domain features on standardized data frames. The time-domain features are calculated based on statistical moments to quantify the macroscopic statistical properties of the signal; for example, the root mean square (RMS) is used to quantify the overall energy of the signal, while kurtosis is used to describe the impulsiveness of the signal waveform. The frequency-domain features are calculated by applying a fast Fourier transform to the signal frames to reveal the frequency distribution of signal energy; specifically, the power spectral density is calculated. To analyze signal energy at different frequency points The signal is distributed across multiple frequency domains; the extraction of time-frequency features relies on wavelet packet transform, which aims to simultaneously capture the local features of the signal in both time and frequency. This transform decomposes the signal into a series of orthogonal frequency sub-bands and calculates the energy entropy of each sub-band. It is used to characterize the uncertainty of signal energy in the time-frequency plane; the energy entropy is calculated as follows:

[0086]

[0087] in, Energy entropy Number of decomposition layers Frequency band index Normalized energy percentage : Number of coefficients; Logarithmic function;

[0088] Normalized energy percentage in the formula , refers to the first Layer In the first frequency band, the second The energy of each small wave packet coefficient This accounts for the total energy of the frequency band The proportion; if the wavelet packet coefficients within this frequency band are ,but:

[0089] and

[0090] in, : The energy of the coefficients of the i-th wavelet packet; : The coefficients of the i-th wavelet packet; Total energy of the frequency band; : Number of coefficients;

[0091] therefore;

[0092]

[0093] in, Normalized energy percentage; : The coefficients of the i-th wavelet packet; : 1st wavelet packet coefficient, dummy variable for summation; : number of coefficients;

[0094] energy entropy It measures the uniformity of signal energy distribution in time-frequency domain. The more uneven the energy distribution, i.e. concentrated in a few coefficients, the smaller the entropy value, and vice versa;

[0095] Based on the above calculation, all time-domain features, frequency-domain features and time-frequency domain features are combined to generate a high-dimensional original feature vector ; It refers to a comprehensive numerical vector containing all extractable features of the electric drill at the current moment, which provides a comprehensive information basis for subsequent load and health state separation. Its source is the splicing of the aforementioned multi-domain feature calculation results. To separate the coupling effects of external working condition changes such as workpiece material and feed speed and internal health state degradation such as bearing wear and gear fatigue in the high-dimensional original feature vector , this unit applies a nonlinear manifold learning mapping ; ; It refers to a pre-trained dimensionality reduction model that maps data points in a high-dimensional feature space to a low-dimensional manifold that preserves their local neighborhood geometry. It is trained based on a sample data set containing various working conditions and health states. Through this mapping, is decomposed into two approximately orthogonal low-dimensional subspace vectors: load feature vector and health feature vector ; mainly captures signal changes caused by external factors, while mainly reflects signal changes caused by internal factors;

[0096] Through this combination of multi-domain parallel feature extraction and nonlinear manifold learning mapping, this embodiment can extract more informative and discriminative feature representations from the original signal. Compared with methods that only use single-domain features, it can more comprehensively capture subtle changes in device state. Most importantly, by successfully decoupling the mixed feature vector into independent load features and health features, it provides high-quality input for subsequent construction of dynamic health baselines that are not affected by working conditions, greatly improving the robustness and accuracy of health state assessment.

[0097] The application process of nonlinear manifold learning mapping includes:

[0098] Find K nearest neighbors for each high-dimensional original feature vector;

[0099] The optimal weights that can linearly reconstruct the high-dimensional original feature vector through K nearest neighbors are calculated;

[0100] A set of coordinates in the low-dimensional space is found, which minimizes the reconstruction error while maintaining the same optimal weight relationship, thus completing the decomposition;

[0101] In this embodiment, the specific application process of the nonlinear manifold learning mapping is further illustrated; this embodiment uses the local linear embedding algorithm to realize the mapping, the core idea of which is to assume that high-dimensional data points are distributed on a low-dimensional manifold, and to find a low-dimensional embedding that can maintain the local neighborhood geometry of data points; this process ensures that the conversion from high-dimensional to low-dimensional does not lose critical topological information;

[0102] The internal logic of this application process is that for each high-dimensional original feature vector , the system finds nearest neighbors for it in its feature space; , which refers to the preset number of neighborhood points, which defines the local geometric range for reconstructing the data point, and its source is a hyperparameter that is determined in advance according to the density and distribution characteristics of the data set through cross-validation and other methods; based on the discovered neighbors, the system then calculates a set of optimal weights, which can linearly reconstruct the high-dimensional original feature vector through the above-mentioned nearest neighbors; the goal of this step is to find a set of weights that minimize the reconstruction error: , where is a neighbor of , and : high-dimensional original feature vector; : neighbor, neighboring point; : optimal weights that can linearly reconstruct through K nearest neighbors; : preset number of neighborhood points;

[0103] When solving this set of optimal weights , two constraint conditions are usually also met: weight normalization constraint, i.e., for each data point , the sum of reconstruction weights is 1, denoted as

[0104]

[0105] , where : reconstruction weight; : number of neighborhood points;

[0106] The second is translation invariance, i.e., if and its neighbors, the reconstruction relationship still holds. These constraints ensure that the weights only reflect the local geometry between data points;

[0107] In the low-dimensional space, the system solves a new set of coordinates, i.e. the load eigenvectors and the health eigenvectors , such that the new set of coordinates minimizes its own reconstruction error while maintaining the same optimal weight relationship as calculated in the previous step; this process essentially unfolds the high-dimensional manifold in the low-dimensional space while keeping the linear reconstruction relationship between each data point and its neighbors unchanged, thus completing the decomposition from to and ;

[0108] By employing the locally linear embedding algorithm, the embodiment realizes a highly efficient and structure-preserving nonlinear dimensionality reduction method; compared with linear dimensionality reduction methods, it can better handle complex and nonlinear data distribution, thus more effectively separating the coupled working condition and health status information; this ability to maintain the local neighborhood geometry ensures that early fault features caused by minor wear will not be discarded as noise during dimensionality reduction, thus providing more sensitive and reliable health eigenvectors for subsequent anomaly detection.

[0109] The deep generative model is a variational autoencoder model with time convolution attention mechanism;

[0110] The training process of the model is performed by minimizing the objective function composed of the reconstruction error and the KL divergence between the posterior distribution and the prior distribution;

[0111] The data used for training is only the health samples collected under normal working conditions of the electric drill;

[0112] In the embodiment, the specific type and training method of the deep generative model are further limited to construct a dynamic health baseline that can quickly adapt to changes in working conditions;

[0113] In the embodiment, the deep generative model is determined as a variational autoencoder model with time convolution attention mechanism; the variational autoencoder refers to a deep generative model that learns the probability distribution of the health status feature under the condition of the current working load , rather than just a single deterministic mapping; the time convolution network is used in the encoder part of the model, which can effectively capture the health feature time series long-term dependencies in the data, and is computationally efficient compared to traditional recurrent neural networks; the attention mechanism is embedded in the decoder, which allows the model to dynamically assign different weights to different time steps of the input sequence when reconstructing the health baseline; when the load characteristics change significantly, the attention weights quickly adjust, allowing the model to focus more on recent data and thus adapt quickly to new operating conditions;

[0114] The training process of the model is performed end-to-end by minimizing an objective function composed of two parts; the first part is the reconstruction error, which ensures that the decoder can accurately reconstruct the original health feature vector from the latent variable and the load condition; the second part is the KL divergence between the posterior distribution and the prior distribution, which acts as a regularization term that encourages the latent variable distribution produced by the encoder to approach the standard normal distribution, making the model's latent space structure more regular and conducive to generation and generalization;

[0115] This objective function is mathematically referred to as the evidence lower bound, and its minimization is equivalent to maximizing the log-likelihood of the data, which can be expressed as:

[0116]

[0117] where, : objective function, also known as evidence lower bound; : model parameters of the decoder and encoder, respectively; : original health feature vector; : load feature vector; : latent variable; : encoder, outputs the posterior distribution of the latent variable; : decoder, used to reconstruct the health features; : standard normal prior distribution of the latent variable; : expectation function, this term is the reconstruction error term; : KL divergence (Kullback-Leibler divergence), this term is the regularization term;

[0118] where the first term is the reconstruction error term, i.e., the expected log-likelihood of the original health feature reconstructed by the decoder under the posterior distribution of the latent variable generated by the encoder ; the second term is the KL divergence regularization term, which measures the difference between the posterior distribution output by the encoder and the standard normal prior distribution ;

[0119] To ensure the sensitivity of the model to unknown faults, the data used to train the model is only healthy samples collected under various normal working conditions of the electric screwdriver; this means that the model learns the concept of health, that is, under any given load condition, what feature distribution should a healthy device exhibit;

[0120] The variational autoencoder model with time convolution attention mechanism brings three major gain effects; the application of time convolution network enhances the model's ability to capture long-term dependencies of time series data, enabling the construction of a health baseline to consider the evolution of historical states; the introduction of the attention mechanism enables the model to respond quickly to working condition changes, dynamically adjust the baseline, and greatly reduce the false alarm rate caused by working condition switching; the strategy of training only with healthy samples makes the model very sensitive to any unknown fault mode deviating from the healthy state, thereby achieving strong anomaly detection capability without the need for a large amount of fault data in advance.

[0121] Embodiment 5

[0122] The process of statistical processing of the residual sequence by the residual sequence generation and statistical monitoring unit is as follows:

[0123] Exponential weighted moving average calculation is applied to the residual sequence;

[0124] The current statistical value is recursively combined with the previous statistical value through a smoothing factor to generate the current statistical value;

[0125] In this embodiment, the process of statistical processing of the residual sequence by the residual sequence generation and statistical monitoring unit is further illustrated, aiming to smooth short-term noise and amplify long-term, weak degradation trends;

[0126] This process uses the exponential weighted moving average control chart algorithm to calculate the residual sequence ; EWMA is a recursive calculation method that combines the current residual value with the previous statistical value through a smoothing factor to generate the current statistical value ; It refers to a preset parameter between 0 and 1, which determines the weight of historical data and current data in calculating the statistical value, and is an empirical value set according to the requirement of system sensitivity to changes. A smaller value makes the control chart more sensitive to small and continuous drifts;

[0127] The recursive calculation process of this statistical value is as follows:

[0128] Statistical smoothing factor current residual previous statistical value:

[0129]

[0130] wherein, EWMA statistic, : smoothing factor, : current residual, : previous time statistic;

[0131] This recursive relationship gives higher weights to recent residual values while preserving the influence of historical information, making it effective in filtering out random noise and highlighting the persistent upward or drifting trend of the residual sequence, which is a typical characteristic of progressive equipment wear;

[0132] By applying the exponentially weighted moving average calculation, the embodiment significantly enhances the system's ability to detect early weak fault signals; compared to directly using the original residual sequence for threshold judgment, the EWMA statistic can effectively smooth out benign spikes caused by measurement noise or transient operating condition fluctuations, thereby avoiding false alarms; at the same time, its memory effect on historical data enables the statistic to stably accumulate and eventually exceed the control limit when the equipment experiences slow and continuous degradation, achieving earlier and more reliable fault warning than the direct threshold method.

[0133] The process of determining the control limit parameter by the parameter optimization and life prediction unit is:

[0134] Setting the maximum condition-residual separation degree as the optimization objective;

[0135] Setting the control limit parameter as the control limit width coefficient;

[0136] Using the Bayesian optimization process to find the control limit width coefficient that maximizes the condition-residual separation degree;

[0137] In this embodiment, the process of determining the control limit parameter by the parameter optimization and life prediction unit is further limited to achieve adaptive and optimal fault discrimination threshold;

[0138] This process sets the maximum condition-residual separation degree as the optimization objective; the condition-residual separation degree refers to a quantitative indicator that measures the degree of distinction between true fault states and different health condition states in the residual space, as well as the distinction between different health condition states; a high SWR value means that fault signals can be clearly separated from various normal condition background noise;

[0139] This process sets the control limit width coefficient as the control limit parameter to be optimized; Refers to the multiplier factor used to calculate the control limit , which determines the balance between the sensitivity and specificity of the warning:

[0140]

[0141] wherein, : mean value of the healthy state : standard deviation of the healthy state : mean value of the healthy state : standard deviation of the healthy state The greater the value, the wider the control limit, the system is less likely to be false positive, but may sacrifice the sensitivity to early failure; vice versa

[0142] In order to find the optimal value that maximizes the SWR, the embodiment adopts a Bayesian optimization process; Bayesian optimization refers to an efficient global optimization algorithm, which is to find the optimal solution of the target function, SWR in this case, with the least number of evaluations without exhaustive search; The process builds a proxy model of SWR about , such as Gaussian process, and uses acquisition functions such as expected improvement to intelligently select the next point to be evaluated, so as to balance between exploration, trying to areas with high uncertainty and exploitation, fine searching in areas with known good performance;

[0143] By introducing the Bayesian optimization process with the goal of maximizing SWR, the embodiment changes the setting of control limit parameter from the experience-dependent manual setting process to the data-driven adaptive optimization process; This ensures that the control limit can always be set at the best position to distinguish between healthy and failure states, thereby optimizing the detection performance of the system globally; Compared with using fixed parameters, this method can automatically find the best sensitivity and specificity trade-off point according to the data characteristics of specific equipment and working conditions, significantly improving the overall accuracy of fault diagnosis.

[0144] The calculation process of working condition-residual separation degree includes:

[0145] First, apply kernel density estimation method in the sliding time window to obtain the residual probability density function under different working condition states;

[0146] Then, based on the discrete form of K-L divergence, calculate the first separation degree between the probability distribution of fault working condition and healthy working condition;

[0147] Then calculate the second separation degree of the probability distribution between different healthy working conditions;

[0148] Finally, determine the working condition-residual separation degree through the ratio relationship between the first separation degree and the second separation degree;

[0149] In this embodiment, the calculation process of working condition-residual separation degree is further disclosed to ensure the calculability of the optimization target and the explicitness of the physical meaning;

[0150] To calculate the separation, the system applies a kernel density estimation method to the collected residual samples within a sliding time window of length Kernel density estimation is a non-parametric statistical method that estimates the unknown probability density function from observed data samples. The calculation is as follows:

[0151] Probability density function of the state class Kernel function bandwidth parameter Window length:

[0152]

[0153] ; Residual variable Window length of the sliding time window, The value of is a parameter determined in advance through experiments based on the stationarity requirement and computational efficiency of the signal, aiming to ensure that the data in the window can stably reflect the statistical characteristics of the current working condition. Residual samples in the time window

[0154] Through this method, the residual probability distribution under different states can be obtained, such as under healthy condition H, i, under healthy condition H, j, and under fault condition F.

[0155] Based on this, the first separation between the probability distribution of the fault condition and the healthy condition is calculated through the discrete form of K-L divergence. K-L divergence is used to measure the difference between two probability distributions, which is used here to quantify the average difference between the fault distribution and all healthy distributions At the same time, the second separation of the probability distribution between different healthy conditions is calculated. This step calculates the K-L divergence between any two different healthy conditions, such as and , and takes the maximum value to quantify the difference in residual distribution caused by normal condition fluctuations. Through the ratio relationship between the first separation and the second separation, the condition-residual separation SWR is determined. The calculation formula is as follows:

[0156] Condition-residual separation

[0157]

[0158] ; Residual probability distribution under fault condition F Average residual probability distribution under all healthy conditions Residual probability distribution under different healthy conditions H, i and H, j : take the maximum value between different health conditions on (i, j);

[0159] The numerator of the ratio represents the difference between the fault signal and the normal signal, and the denominator represents the difference between the normal signals; Therefore, the larger the SWR value, the higher the signal-to-noise ratio of the fault signal relative to the normal condition fluctuation;

[0160] The SWR calculation method defined in this embodiment provides a target function with clear physical meaning and deep insight into the actual problem for the optimization process; Not only considering the difference between fault and health, the numerator, but more importantly, the fluctuation difference between normal conditions as a benchmark, the denominator is normalized; This design makes the optimization goal not simply to pull away the distance between fault and health, but to pursue an ideal state that effectively distinguishes faults while tolerating normal condition switching to the greatest extent, so that the control limit parameters obtained by optimization have stronger robustness and practical application value

[0161] In order to express more clearly, the SWR formula can be rewritten in the standard form of K-L divergence in information theory. K-L divergence Measures the difference between probability distributions and ; Therefore, the SWR formula can be equivalently understood as:

[0162]

[0163] Where the numerator represents the divergence between the fault residual distribution and the average health residual distribution, and the denominator represents the maximum value of the residual distribution difference between different health conditions; The definition aims to find a control limit that maximizes the signal-to-noise ratio between the fault signal, the numerator, and the normal condition fluctuation, the denominator; : K-L divergence used to measure the difference between probability distributions P and Q; : fault residual distribution; : average health residual distribution; : residual distribution of different health conditions;

[0164] The parameter optimization and life prediction unit is also used to predict the remaining useful life;

[0165] The prediction process specifically includes:

[0166] By time accumulating the statistics generated by the residual sequence generation and statistical monitoring unit, a monotonically increasing degradation state variable is constructed;

[0167] The evolution process of the degradation state variable is modeled as a Wiener process with drift;

[0168] The Wiener process is defined by a drift coefficient representing the average degradation rate and a diffusion coefficient describing the uncertainty;

[0169] The remaining useful life is obtained by calculating the expected time for the degradation state variable to reach a pre-defined failure threshold for the first time from the current value;

[0170] In this embodiment, the implementation of the parameter optimization and the other function of the life prediction unit, i.e. the prediction of the remaining useful life, is further disclosed;

[0171] The implementation of this prediction process starts from the construction of a monotonically increasing degradation state variable by time accumulation of the residual sequence generated by the residual generation unit and the statistical quantities generated by the statistical monitoring unit ; which refers to an index quantifying the accumulated degradation degree of the whole process from health to failure, and the role is to transform the non-trended and fluctuating health monitoring statistical quantities into a monotonically increasing process that can represent the cumulative effect of wear and tear;

[0172] The evolution process of this degradation state variable is then modeled as a Wiener process with drift; the Wiener process refers to a continuous-time stochastic process that describes the random cumulative effect similar to Brownian motion, which is very suitable for simulating mechanical wear and tear and other degradation processes with randomness and accumulation; the mathematical expression of this model is as follows:

[0173] Drift coefficient of degradation state variable Diffusion coefficient of degradation state variable Standard Wiener process:

[0174]

[0175] : degradation state variable, : drift coefficient, : diffusion coefficient, : standard Wiener process; : infinitesimal increments of degradation state, time and standard Wiener process, respectively;

[0176] In this model, the drift coefficient represents the average degradation rate, and its dimension is the unit of degradation per unit of time; the diffusion coefficient describes the uncertainty of the degradation process, and its dimension is the square root of the unit of degradation per unit of time, which describes the uncertainty or random fluctuation in the degradation process; these two parameters are calibrated online using historical health and failure data sequences through maximum likelihood estimation and other methods;

[0177] The prediction value of the remaining useful life is obtained by calculating the expected time for the degradation state variable to reach a pre-defined failure threshold for the first time from the current value ​ ; The upper limit of the cumulative degradation amount when the tool cannot meet the performance requirements, which is set in advance according to industry standards, experimental data or expert experience; based on the first arrival time theory of Wiener process, the prediction formula of RUL is determined as:

[0178] Remaining useful life Failure threshold Current degradation amount Degradation rate:

[0179]

[0180] Wherein, : remaining useful life, : failure threshold, : current degradation amount, : dimensionless drift coefficient;

[0181] The physical meaning of this formula is that the remaining available degradation amount, the numerator divided by the average degradation rate, the denominator, that is, the remaining useful life is obtained;

[0182] The remaining useful life here is the expected value obtained based on the first arrival time theory of Wiener process; it is mainly determined by the drift coefficient , which represents the average trend of degradation; and the diffusion coefficient in the model describes the randomness of the degradation process, which can be used to quantify the uncertainty of RUL prediction, for example, by calculating the confidence interval of the predicted value; a larger value means that the volatility of the degradation path is stronger, resulting in a wider confidence interval of RUL prediction, which provides more comprehensive information for risk assessment and maintenance decision-making;

[0183] By modeling the degradation process of the device as a Wiener process with drift, the embodiment provides a powerful mathematical framework for RUL prediction that can reflect both the average degradation trend and the randomness of the process; compared with the simple linear extrapolation method, this method can provide more physically realistic prediction results, and can quantify the uncertainty of prediction through the diffusion coefficient ; this makes the prediction result no longer a single deterministic value, but a probability distribution, which provides key data support for managers to conduct risk assessment and develop more refined maintenance strategies.

[0184] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A drill grinder wear state prediction and health management system, characterized in that, The method comprises the following steps: a data acquisition and preprocessing unit is used to acquire multi-modal physical signals and process them to generate standardized data frames; a multi-domain feature transformation unit is used to receive the standardized data frames and transform them into health feature vectors and load feature vectors; a dynamic health baseline construction unit is used to reconstruct a dynamic health baseline from the time series of the health feature vectors and the load feature vectors by using a deep generative model; a residual sequence generation and statistical monitoring unit is used to calculate the distance between the health feature vectors and the dynamic health baseline, generate a residual sequence, and statistically process the residual sequence to obtain a statistical quantity; a parameter optimization and life prediction unit is used to determine a control limit parameter according to a preset optimization target; an instruction generation and control execution unit is used to determine a control limit according to the control limit parameter, compare the statistical quantity with the control limit, generate a diagnosis instruction and execute a control operation if the statistical quantity exceeds the control limit, and determine a normal state if the statistical quantity does not exceed the control limit.

2. A wear state prediction and health management system for a hand drill according to claim 1, characterized in that, The process of transforming the standardized data frames by the multi-domain feature transformation unit comprises the following steps: firstly, the time domain features, the frequency domain features and the time-frequency domain features of the standardized data frames are calculated in parallel; then, the calculated time domain features, the frequency domain features and the time-frequency domain features are combined to generate high-dimensional original feature vectors; finally, a nonlinear manifold learning mapping is applied to the high-dimensional original feature vectors to decompose them into health feature vectors and load feature vectors.

3. A wear state prediction and health management system for a hand drill according to claim 2, characterized in that, The application process of the nonlinear manifold learning mapping comprises the following steps: K nearest neighbor points are found for each high-dimensional original feature vector; the optimal weight capable of linearly reconstructing the high-dimensional original feature vector from the K nearest neighbor points is calculated; a set of coordinates is found in a low-dimensional space, which minimizes the reconstruction error under the same optimal weight relationship, thereby completing the decomposition.

4. The wear state prediction and health management system of a hand drill according to claim 1, wherein, The deep generative model is a variational autoencoder model with a time convolution attention mechanism; The training process of the model is performed by minimizing the objective function composed of the reconstruction error and the KL divergence between the posterior distribution and the prior distribution; The data used for training is only the health samples collected by the electric drill under normal working conditions.

5. The wear state prediction and health management system of a hand drill according to claim 1, wherein, The process of statistically processing the residual sequence by the residual sequence generation and statistical monitoring unit comprises the following steps: an exponential weighted moving average calculation is applied to the residual sequence; the statistical quantity is generated by recursively combining the current residual value and the statistical quantity at the previous time through a smoothing factor.

6. A wear state prediction and health management system for a hand drill according to claim 1, characterized in that, The process of determining the control limit parameter by the parameter optimization and life prediction unit comprises the following steps: the maximization of the working condition-residual separation degree is set as the optimization target; the control limit parameter is set as the control limit width coefficient; the Bayesian optimization process is used to find the control limit width coefficient that maximizes the working condition-residual separation degree.

7. A wear state prediction and health management system for a hand drill according to claim 6, characterized in that, The calculation process of the working condition-residual separation degree comprises the following steps: firstly, the kernel density estimation method is applied within a sliding time window to obtain the residual probability density function under different working condition states; then, the first separation degree between the probability distributions of the fault working condition and the health working condition is calculated based on the discrete form of the K-L divergence; the second separation degree of the probability distributions between different health working conditions is calculated; finally, the working condition-residual separation degree is determined through the ratio relationship between the first separation degree and the second separation degree.

8. The wear state prediction and health management system of a hand drill according to claim 1, wherein, The parameter optimization and life prediction unit is further configured to predict the remaining useful life; The prediction process specifically comprises: constructing a monotonically increasing degradation state variable by time accumulating the statistics generated by the residual sequence generation and statistical monitoring unit; modeling the evolution of the degradation state variable as a Wiener process with drift; the Wiener process is defined by a drift coefficient representing the average degradation rate and a diffusion coefficient describing the uncertainty; the remaining useful life is obtained by calculating the expected time for the degradation state variable to first reach a pre-set failure threshold from the current value.

Images