A non-intrusive digital-analog linkage control system fault detection method

Abstract

The present invention discloses a non-invasive digital-analog linkage control system fault detection method, which uses sensors to collect the output time series of a multi-axis motion control system under a specified state, and after wavelet decomposition, calculates the corresponding relative wavelet energy entropy and other preset statistical features, and then constructs a composite health index through weighted combination, establishes respective Gaussian mixture hidden Markov models for different specified states, and uses the Gaussian mixture hidden Markov model to obtain the posterior probability of the composite health index. By minimizing the optimization objective function, the weight coefficient of the composite health index and the model parameters of the Gaussian mixture hidden Markov model are iteratively adjusted. The technical solution of the present invention has high fault classification accuracy and strong interpretability, and does not require the deployment of additional sensors.

Description

Technical Field

[0001] The present application belongs to the technical field of fault prediction of industrial control systems, and specifically relates to a non-invasive digital-analog linkage control system fault detection method. Background Art

[0002] A multi-axis motion control system is an industrial control system that uses multiple actuators to achieve precise control of the position, speed, and posture of mechanical moving parts. It is widely used in discrete manufacturing industries such as laser engraving, printing, and textiles. The components of a multi-axis motion control system usually include sensors, motion control and drivers, motors, transmission mechanisms, and mechanical loads. During the production and processing process, the failure or failure of any key component will affect the dynamic performance of the multi-axis motion control system, causing the motion trajectory to deviate from the expected control effect, resulting in a decrease in control accuracy and product scrapping. In severe cases, it will cause equipment downtime, equipment damage, and casualties. Therefore, studying the fault detection method of the multi-axis motion control system has important engineering practical significance for timely discovering potential faults, accurately classifying faults, and guiding the formulation of maintenance strategies, reducing enterprise maintenance costs, and improving production efficiency.

[0003] In order to ensure that the system has high-precision motion control performance, closed-loop control technology is widely used in multi-axis motion control systems, which increases the complexity of the system while improving the dynamic performance of the system. Specifically, even small faults will be further transmitted and amplified in the closed-loop control system, posing a huge challenge to the accurate fault diagnosis and isolation of the system. Existing multi-axis motion control system fault detection methods can be mainly divided into signal analysis-based, kinematic / dynamic model-based and data-driven methods.

[0004] Among them, the signal analysis-based method believes that the fault of the closed-loop control system will eventually be reflected in the output signal of the system. By monitoring the time series of the system output and using spectrum analysis, Hilbert-Huang transform, empirical mode decomposition, wavelet packet decomposition and other technologies, the frequency domain and time-frequency domain features sensitive to faults can be extracted to achieve fault diagnosis of multi-axis motion control systems. However, the signal analysis-based method has limitations in processing non-stationary and nonlinear signals, and the monitored sequence is easily disturbed by the complex working conditions and uncertain external environment of the industrial site, which makes the fault diagnosis accuracy based on the signal analysis method low, especially it is difficult to identify weak early fault signals. The model-based method is mainly based on residual analysis. According to the different residual generation methods, it can be divided into equivalent space-based, parameter estimation-based and state estimation-based methods. It obtains the kinematic / dynamic model of the closed-loop control system through mechanism analysis or experimental modeling, and calculates the residual signal of the model output estimate and the measured value to complete the fault diagnosis. However, multi-axis motion control systems have inherent nonlinearity, mechanical coupling between axes, and strong coupling of multi-axis synchronization control algorithms, which are often difficult to be accurately modeled in practical applications, thus limiting the further application and promotion of model-based methods.

[0005] In addition, the development of sensor technology and computer technology has provided a large amount of labeled state monitoring data for data-driven fault detection methods, bringing new development opportunities to deep neural network methods with learning capabilities. Although data-driven methods have demonstrated their great advantages in many industrial personalized cases, such as high-precision fault detection performance, independence from precise system modeling, and relatively low learning thresholds, their inherent "black box model" brings unexplainability, model parameters are difficult to transplant, and require large-scale labeled training data, making it difficult to obtain large-scale applications in the industrial field. Summary of the invention

[0006] The purpose of this application is to provide a non-intrusive digital-analog linked control system fault detection method, which obtains the built-in sensor data of the control system, combines the respective advantages of data-driven methods in processing multi-source sensor information with non-stationary and nonlinear characteristics and probability-based generation models in quantifying uncertainty and revealing the complex relationship between hidden variables, so as to overcome the lack of interpretability of the existing single data-driven fault diagnosis method and the poor diagnostic accuracy of the single system model-driven fault diagnosis method due to the difficulty in accurate modeling.

[0007] In order to achieve the above purpose, the technical solution of this application is as follows:

[0008] A non-intrusive digital-analog linkage control system fault detection method, comprising:

[0009] The sensor is used to collect the output time series of the multi-axis motion control system under the specified state, and after wavelet decomposition, the corresponding relative wavelet energy entropy and other preset statistical characteristics are calculated;

[0010] The relative wavelet energy entropy and other preset statistical features are normalized to obtain normalized features, and then the normalized features are weighted combined to construct a composite health index;

[0011] Establish respective Gaussian mixture hidden Markov models for different specified states, and use Gaussian mixture hidden Markov models to obtain the posterior probability of composite health indicators;

[0012] The weight coefficients of the composite health index and the model parameters of the Gaussian mixture hidden Markov model were iteratively adjusted by minimizing the optimization objective function.

[0013] Furthermore, the other preset statistical features include a pulse index, a rectified mean value, a standard deviation, a margin index, a kurtosis factor and a peak-to-peak value.

[0014] Furthermore, the specified state includes a normal state, a drift deviation state, a fixed deviation state, a precision degradation state, a partial failure state and a gear wear fault state.

[0015] Furthermore, the sensor includes a torque sensor, a displacement sensor and a speed sensor.

[0016] Furthermore, the composite health index is constructed by weighted combination, using the following formula:

[0017]

[0018] Among them, wg,r represents the weight coefficient corresponding to the g-th feature of the r-th class state, h g,r (k) represents the g-th normalized feature of the r-th class state at sampling point k, and G represents the number of normalized features.

[0019] Furthermore, the method of iteratively adjusting the weight coefficient of the composite health index and the model parameters of the Gaussian mixture hidden Markov model by minimizing the optimization objective function includes:

[0020] Construct the optimization objective function J(W r ):

[0021]

[0022] in, represents the actual label at time τ, p r (τ) represents the posterior probability, is the sum of weight coefficients, η>0 means parameter The upper bound of is the total number of samples after sliding window;

[0023] Introducing a penalty term into the objective function transforms the original constrained problem into an unconstrained problem, namely:

[0024]

[0025] Among them, σ1, σ2 are hyperparameters, and R represents the total number of specified states;

[0026] The gradient descent method is used to minimize the optimization objective function, and the weight coefficients of the composite health indicators and the model parameters of the Gaussian mixture hidden Markov model are obtained.

[0027] Compared with the prior art, the advantages of the present application are: a fault detection method based on data-model interaction is proposed, and by designing a global optimization objective function with a multi-feature fusion weight range penalty function, the fault-sensitive features extracted by the data-driven method and the probability generation model established by the model-driven method are efficiently coordinated under the fault diagnosis task. Different from the existing fault diagnosis methods based on external sensor signals such as vibration signals, the fault classifier is trained from the perspective of non-invasive data acquisition, that is, only the built-in sensor signals of the control system itself are used to achieve accurate classification of system faults, reducing the dependence of the system fault diagnosis task on the deployment of additional sensors. The technical solution of the present application has high fault classification accuracy and strong interpretability, and does not require the deployment of additional sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of the control system fault detection method of non-intrusive digital-analog linkage of this application;

[0029] Figure 2 This is a framework diagram of the control system fault detection method of non-intrusive digital-analog linkage in this application;

[0030] Figure 3 This is the online fault detection flow chart for this application;

[0031] Figure 4 This is a schematic diagram of the detection results of this application under different system states. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0033] An embodiment of the present application, such as Figure 1 , Figure 2As shown, a non-intrusive digital-analog linkage control system fault detection method is proposed, including:

[0034] Step S1, using a sensor to collect the output time series of the multi-axis motion control system under a specified state, and after wavelet decomposition, calculating the corresponding relative wavelet energy entropy and other preset statistical features.

[0035] The sensors used in this embodiment include torque sensors, displacement sensors and speed sensors, which respectively collect torque, displacement and speed of the multi-axis motion control system. The specified states include normal state, drift deviation state, fixed deviation state, precision degradation state, partial failure state and gear wear fault state.

[0036] Taking the displacement sensor as an example, the collected output time series is defined as y(t), where t represents the sampling time. In this embodiment, the collected output time series y(t) is decomposed by wavelet to decompose it into low-frequency parts {a l,t} and high frequency part {d l,t}, 1≤l≤L, L is the number of decomposition scales, where {a l,t} can be decomposed into {a l+1,t} and {d l+1,t}.

[0037] After wavelet decomposition, the output time series y(t) is expressed as:

[0038] y(t)=a L,t +d L,t +d L-1,t +…+d 1,t (1);

[0039] Calculate the wavelet energy E for the output time series y(t) tot :

[0040]

[0041] Among them, the wavelet energy of the high-frequency part is l len represents {d l,t} length, the energy of the low-frequency part is l len ′ represents {a l,t}Length, E L+1 =E(a L ), E L =E(d L ), E L-1 =E(d L-1 ), until E1=E(d1).

[0042] Then the relative wavelet energy entropy H l′ is defined as:

[0043]

[0044] in And l′=1,2,…,L,L+1.

[0045] Relative wavelet energy entropy is a part of the time-frequency statistical features. Other preset statistical features include pulse index, rectified mean, standard deviation, margin index, kurtosis factor, and peak-to-peak value. Their calculation formulas are:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] z6=max[y(t)]-min[y(t)] (9);

[0052] Among them, z1 is the pulse index, z2 is the rectified mean, z3 is the standard deviation, z4 is the margin index, z5 is the kurtosis factor, z6 is the peak-to-peak value, and N represents the number of samples. represents the mean of y(t).

[0053] Step S2: normalize the relative wavelet energy entropy and other preset statistical features to obtain normalized features, and then construct a composite health index by weighted combination of the normalized features.

[0054] Relative wavelet energy entropy and other preset statistical features, for any of these statistical features, e g (k) is used as an example, and is transformed to the same range through normalization, namely:

[0055]

[0056] Among them, g represents the g-th statistical feature, and k represents the k-th sampling data. max and e min Corresponding to e g The maximum and minimum values ​​of (k), h g (k) is the normalized feature obtained after normalization.

[0057] This embodiment uses wavelet transform combined with sliding window technology, and sets the window length to L1=N, that is, the number of samples in each window is N, and the step length is S1. The data of the sampling point, T is the final sampling time.

[0058] Finally, a composite health index is constructed through weighted combination. r (k) namely:

[0059]

[0060] Among them, w g,r represents the weight coefficient corresponding to the g-th feature for establishing the r-th (1≤r≤R)-th state, the r-th state refers to one of the states mentioned in step S1, and the total number of states is represented by R. g,r (k) represents the g-th normalized feature of the r-th class state at sampling point k, and G represents the number of normalized features.

[0061] Step S3: establish respective Gaussian mixture hidden Markov models for different specified states, and use the Gaussian mixture hidden Markov model to obtain the posterior probability of the composite health indicator.

[0062] Specifically, this embodiment uses a Gaussian Mixed HiddenMarkov Model (GMHMM) to obtain the posterior probability of the composite health index. It includes:

[0063] Step 3.1, construct a Gaussian mixture hidden Markov model GMHMM, GMHMM includes the hidden state transfer matrix A, the observation probability matrix B and the initial state probability vector π, that is:

[0064] λ=(A,B,π) (12);

[0065] where A=(a i,j ) θ×θ Represents the hidden state s i Transfer to j The probability of s i 、s j From the hidden state sequence S = {s1, s2, ..., s θ}. i,j =P(q(k)=s j |q(k-1)=s i), 1≤i, j≤θ. θ represents the number of hidden states. The hidden state sequence is the set of states implied in the GMHMM; the observation sequence is the set that can be observed in the GMHMM. This step is to further operate on the result o(k) of step 2 and construct a Gaussian mixture hidden Markov model). q(k)∈S is the hidden state at time k, and o(k)∈V is the composite health indicator at time k. B=(b i (o(k))) θ×κ Describes the hidden state s i The probability of observing o(k) can be expressed by the multivariate Gaussian mixture probability density function, that is:

[0066]

[0067] Where κ represents the number of observation sequences, M is the number of Gaussian elements in the Gaussian mixture model; c i,m is the state i The mth mixing weight under i,m and Σ i,m Respectively represent the state s i The mean vector and variance matrix of the m-th Gaussian distribution under N(o(k)|μ i,m ,Σ i,m ) represents the multivariate Gaussian density function.

[0068] Step 3.2: Use the Baum-Welch algorithm to train the GMHMM, update the five sets of parameters π, a, c, μ, and Σ, and then select the optimal set of model parameters, that is, The proposed method requires training R GMHMMs for R operating states, collectively referred to as

[0069] The Baum-Welch algorithm is an iterative algorithm for training hidden Markov models. It is an unsupervised learning algorithm for estimating the parameters of hidden Markov models so that the model can better fit the observed data. This embodiment uses the Baum-Welch algorithm to train the GMHMM model to obtain the optimal parameters of the GMHMM model.

[0070] Step 3.3: Calculate the number of Composite health index o r The posterior probability p of (τ) r (τ), that is:

[0071]

[0072] When calculating the posterior probability of the composite health indicator, the sliding window technique is used again, which is called the second sliding window. is the total number of samples after sliding the window. For the sliding window technique, an appropriate observation sequence length L2 is selected, and the step size is set to S2 = 1, in order to avoid the disadvantage of insufficient information contained in a single sample and to facilitate accurate identification of online samples. The setting of the observation sequence length L2 can be specified through experiments and human experience, and will not be repeated here.

[0073] In this step, the Gaussian mixture hidden Markov model is used to obtain the posterior probability of the composite health index, in preparation for the subsequent construction of the optimization objective function.

[0074] Step S4: iteratively adjust the weight coefficients of the composite health index and the model parameters of the Gaussian mixture hidden Markov model by minimizing the optimization objective function.

[0075] Specifically, this embodiment iteratively adjusts the weight coefficient of the composite health index and the model parameters of the Gaussian mixture hidden Markov model by minimizing the optimization objective function, including:

[0076] Step 4.1: Construct the optimization objective function of the rth specified state.

[0077] This step is done by using the posterior probability p of step 3.3 r (τ) and the actual label Construct the local optimization objective function of the rth training model:

[0078]

[0079] in, represents the actual label at time τ. is the sum of weight coefficients, η>0 means parameter The upper bound of .

[0080] Step 4.2: Introduce a penalty term into the objective function to transform the original constrained problem into an unconstrained problem, namely:

[0081]

[0082] Among them, σ1 and σ2 are hyperparameters. In this step, multiple local optimization objective functions are summed to obtain the global optimization objective function.

[0083] Step 4.3: Use the gradient descent method to minimize the optimization objective function and solve the weight coefficient of the composite health index and the model parameters of the Gaussian mixture hidden Markov model.

[0084] In this step, the gradient descent method is used to minimize the optimization objective function, J(W r ) to obtain the linear weighted fusion weight w g,r Optimal solution for Gaussian mixture hidden Markov model parameters Finally, the feedback between composite health indicators and modeling was realized, thus forming a closed-loop feedback mechanism between composite health indicator construction and GMHMM training.

[0085] In another specific embodiment, Figure 3 As shown, online fault detection is performed, including the following steps:

[0086] Step 5.1: For each new sample, slide the window on the sensor data and perform wavelet transform in the window. The wavelet transform method is the same as before.

[0087] Step 5.2: Calculate the corresponding relative wavelet energy entropy and other preset statistical features.

[0088] Step 5.3: Use the optimal fusion coefficient w g,r Calculate composite health indicators (multi-feature fusion).

[0089] Step 5.3: Use the optimal parameter model And the observation sequence length L2 is used to calculate the posterior probability.

[0090] Step 5.4: Finally, output the fault category corresponding to the maximum value of the posterior probability, thereby completing the fault detection.

[0091] In a specific embodiment, the present application is also verified by constructing an engraving machine platform;

[0092] Set the circular tracking reference trajectory of the platform cx(t) = 10sin(2πt / 1600) + 20, cy(t) = 10cos(2πt / 1600) + 20. Then the contour trajectory radius is R = 10, the center is a circle (20,20), the system operation cycle is 16s, the sampling cycle is 5ms, and the sampling time is 480s. Collect data of 6 types of labels on the engraving machine platform, including drift deviation, fixed offset, accuracy degradation, partial failure, gear failure and normal state, and then use the forward-backward algorithm to calculate the maximum probability of the test data based on the optimal fusion coefficient and GMHMM parameters. Figure 4The recognition results of every 10 observed test samples in the GMHMM modeled in each state are shown, and the ordinate represents the posterior likelihood probability. The state in which the posterior likelihood probability output after modeling is the largest is the state. It can be seen from the first 5 sub-figures that the diagnosis is accurate for fixed bias, accuracy degradation, partial failure, gear fault and normal state; it can be seen from the 6th sub-figure that the first 59 minority samples of drift deviation fault are not correctly diagnosed. This confusion is mainly because the drift deviation fault is an early fault characterized by slow changes in parameters over time and relatively small magnitudes of early faults. Overall, the average classification accuracy of the method of the present application in all 6 states reaches 95.37%, and the fault detection accuracy is high.

[0093] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A non-intrusive digital-analog linkage control system fault detection method, characterized in that: The non-intrusive digital-analog linkage control system fault detection method comprises: The sensor is used to collect the output time series of the multi-axis motion control system under the specified state, and after wavelet decomposition, the corresponding relative wavelet energy entropy and other preset statistical characteristics are calculated; The relative wavelet energy entropy and other preset statistical features are normalized to obtain normalized features, and then the normalized features are weighted combined to construct a composite health index; Establish respective Gaussian mixture hidden Markov models for different specified states, and use Gaussian mixture hidden Markov models to obtain the posterior probability of composite health indicators; By minimizing the optimization objective function, the weight coefficients of the composite health index and the model parameters of the Gaussian mixture hidden Markov model are iteratively adjusted; The composite health index is constructed by weighted combination using the following formula: ； in, Indicates r Class status The weight coefficient corresponding to the feature, Indicates sampling point Next Class status Normalized features, G Represents the number of normalized features; The method of iteratively adjusting the weight coefficient of the composite health index and the model parameters of the Gaussian mixture hidden Markov model by minimizing the optimization objective function includes: Construction r Optimization objective function for class-specified states : ； in, Indicates at time The actual label at represents the posterior probability, is the sum of weight coefficients, Representation parameters The upper bound of is the total number of samples after sliding window; Introducing a penalty term into the objective function transforms the original constrained problem into an unconstrained problem, namely: ； in, , is a hyperparameter, R Indicates the total number of specified states; The gradient descent method is used to minimize the optimization objective function, and the weight coefficients of the composite health indicators and the model parameters of the Gaussian mixture hidden Markov model are obtained.

2. The non-intrusive digital-analog linkage control system fault detection method according to claim 1 is characterized in that: The other preset statistical features include pulse index, rectified mean, standard deviation, margin index, kurtosis factor and peak-to-peak value.

3. The non-intrusive digital-analog linkage control system fault detection method according to claim 1 is characterized in that: The specified states include a normal state, a drift deviation state, a fixed deviation state, a precision degradation state, a partial failure state, and a gear wear fault state.

4. The non-intrusive digital-analog linkage control system fault detection method according to claim 1 is characterized in that: The sensors include a torque sensor, a displacement sensor and a speed sensor.

Images