Spindle condition assessment method and system based on multi-source information deep coupling network

Abstract

The present invention discloses a spindle state assessment method and system based on a multi-source information deep coupling network, which collects multi-source sensor information for characterizing electric spindle faults, extracts time domain features and frequency domain features of the multi-source sensor information, constructs a descending ridge membership model based on the acquired time domain features and frequency domain features, adaptively normalizes the time domain features and frequency domain features according to their ranges, and adopts a collaborative fusion method of multiple sensors to improve the accuracy of state recognition and avoid the error problem existing in a single sensor. The present invention adopts an entropy weight method to calculate the weight matrix corresponding to the feature matrix, avoiding the influence of feature indicators with smaller fluctuations on the final state recognition results. Taking into account the "short board effect" that may be produced by the sensor, the sensor characteristics with poor indicators are given special consideration. The electric spindle state recognition method proposed by the present invention has strong applicability and can be widely used in electric spindles in various states.

Description

Technical Field

[0001] The present invention belongs to the technical field of numerical control machining, and in particular relates to a spindle state evaluation method and system based on a multi-source information deep coupling network. Background Art

[0002] As a core component of machine tools, the performance and reliability of the spindle directly impact CNC machine tool machining accuracy, production efficiency, and product performance. As compact and powerful spindles, electric spindles are currently widely used in various fields. Advances in electric spindle condition assessment technology can facilitate the timely detection of potential faults and identify health degradation trends, thereby enabling maintenance based on the electric spindle's condition. This improves the maintenance and management of CNC systems, extends the life of the electric spindle and other components, and avoids the current mainstream problems of insufficient reliability and stability.

[0003] At present, most electric spindle monitoring algorithms perform simple time-domain analysis on basic sensor signals to obtain parameters such as peak values ​​and directly judge the current state. They have a low level of intelligence, poor accuracy in identifying actual states, and no advanced electric spindle state monitoring algorithms with strong adaptability. Yan Ruqiang et al. developed an automatic spindle state monitoring system based on OSACBM and bearing fault judgment based on the envelope spectrum algorithm of wavelet analysis. Fischer et al. proposed a bearing judgment index based on acoustic emission signals, which can determine the remaining life of bearings and perform preventive maintenance on them. Zhang Zhi et al. performed wavelet packet decomposition on the spindle vibration signal, and used its energy entropy value as an indicator to judge the state of milling processing, and judged the spindle state according to the occurrence of chatter. Li Qiang used the sliding SVDD model to establish a spindle health index calculation method based on multi-source information fusion and rough set theory.

[0004] Currently, spindle status monitoring for electric spindles often relies on a single vibration signal for status determination. However, as machine tools and electric spindle structures become increasingly complex, fault coupling is increasing, and faults in motors, bearings, and tools can affect each other. Therefore, the single sensor signal currently used cannot provide complete and accurate status information, making it difficult to effectively determine the status of the electric spindle. Therefore, the complementary fusion of information from the same sensor and the collaborative fusion of different sensors are key research challenges. Summary of the Invention

[0005] The purpose of the present invention is to provide a spindle state assessment method and system based on a multi-source information deep coupling network to overcome the shortcomings of the prior art.

[0006] A spindle condition assessment method based on a multi-source information deep coupling network includes the following steps:

[0007] S1, collects multi-source sensor information for characterizing electric spindle faults;

[0008] S2: Extract time domain features and frequency domain features of multi-source sensing information;

[0009] S3: constructing a descending ridge membership model based on the acquired time domain features and frequency domain features, and adaptively normalizing the time domain features and frequency domain features according to their ranges to obtain a normalized feature matrix that can be used to calculate weights;

[0010] S4: Calculate the characteristic information entropy according to the characteristic matrix, and use the entropy weight method to calculate the weight matrix corresponding to the characteristic matrix;

[0011] S5: using the equalization function to calculate the discreteness of the characteristic matrix, calculating the variable weight vector corresponding to the multi-source sensor information according to the discreteness of the characteristic matrix, and calculating the optimized weight matrix according to the variable weight vector;

[0012] S6: Perform feature matrix weighted fusion on the optimized weight matrix to obtain the spindle state index vector, and use the gray clustering whitening weight function to calculate the spindle state index vector to obtain the spindle state classification result.

[0013] Preferably, the multi-source sensor information includes a vibration signal, a current signal and a temperature signal for characterizing the electric spindle fault.

[0014] Preferably, the time domain features include peak value, root mean square, kurtosis index and pulse index, and the frequency domain features include the electric spindle rotation frequency energy and double rotation frequency energy in the high-frequency signal.

[0015] Preferably, a low-pass filter is used to filter the vibration signal and the current signal to obtain a signal including a low-frequency component;

[0016] The time domain and frequency domain features of the low-frequency component signals were calculated using MATLAB, and 12 feature quantities of the two vibration sensors and six feature quantities of the current sensor were obtained.

[0017] The temperature signal is smoothed to obtain the point value signal of the temperature, and the feature matrix is ​​constructed together with the feature quantity obtained in step (2.2). The feature matrix contains x samples, and each sample has a total of 19 parameters.

[0018] Preferably, the constructed descending ridge membership function is:

[0019]

[0020] Select the parameters a and b in the semi-ridge membership function. For each column of parameters, select 90% and 10% of its maximum value as parameters b and a respectively.

[0021] Preferably, the information entropy of each indicator is calculated, and the information entropy value corresponding to the j-th indicator is calculated according to the following formula:

[0022]

[0023] Among them, if z ij =0, that is, the current indicator score is 0, then define This satisfies the lower limit of entropy to 0.

[0024] Preferably, the discreteness of the feature matrix is ​​calculated:

[0025]

[0026] Construct state equilibrium function:

[0027]

[0028] In the formula, r is the discreteness of the factor state vector, X i is the state value of indicator i, m is the number of indicators in the indicator group, τ is the "incentive-penalty ratio", α and β are the indicator state critical values ​​of punishment and incentive respectively, (0, α) is the penalty zone, [α, β] is the "incentive-penalty balance interval", and (β, 1] is the incentive zone. Among them, ω and ξ are variable weight constraint parameters, which are set to 0.3 and 0.1 respectively.

[0029] Preferably, the parameters α and β in the balancing function are determined by the K-means algorithm. After obtaining the preliminary feature index weights using the entropy weight method, they are multiplied and added with the corresponding indicators to obtain the preliminary spindle state index. The spindle state index is K-means clustered and the cluster center values ​​at both ends are taken as the incentive parameters.

[0030] The state variable weight vector S(X) is calculated by the equilibrium function, and the weight vector W=[w1,w2,…,w n ]Calculate the weight matrix after optimizing the weights:

[0031]

[0032] Preferably, the spindle state index is calculated based on the weight matrix and the normalized feature matrix after the weights are optimized:

[0033] h=m*w T

[0034] Among them, h is the main axis state index, m is the characteristic index vector, and w is the characteristic index weight obtained by variable weight. Finally, the main axis state index h corresponding to each sample is calculated. i , and merge them into the final spindle state index vector H=[h1,h2,h3...h i ].

[0035] A spindle condition assessment system based on a multi-source information deep coupling network includes a preprocessing module, an optimization module, and a classification module:

[0036] A preprocessing module is used to collect multi-source sensor information used to characterize the electric spindle fault and extract the time domain characteristics and frequency domain characteristics of the multi-source sensor information;

[0037] An optimization module constructs a descending ridge membership model based on the acquired time domain features and frequency domain features, adaptively normalizes the time domain features and frequency domain features according to their ranges, and obtains a normalized feature matrix that can be used to calculate weights; calculates feature information entropy based on the feature matrix, calculates a weight matrix corresponding to the feature matrix using an entropy weight method, calculates the discreteness of the feature matrix using an equalization function, calculates a variable weight vector corresponding to the multi-source sensor information based on the discreteness of the feature matrix, and calculates an optimized weight matrix based on the variable weight vector;

[0038] The classification module is used to perform feature matrix weighted fusion on the optimized weight matrix to obtain the spindle state index vector, and the spindle state classification result is obtained by using the grey clustering whitening weight function to calculate the spindle state index vector.

[0039] Compared with the prior art, the present invention has the following beneficial technical effects:

[0040] The present invention discloses a spindle state assessment method based on a multi-source information deep coupling network. Aiming at the problem of electric spindle state assessment, the present invention adopts a collaborative fusion method of multiple sensors to collect multi-source sensor information for characterizing electric spindle faults; and extracts time domain features and frequency domain features of the multi-source sensor information, constructs a descending ridge membership model based on the acquired time domain features and frequency domain features, and adaptively normalizes the time domain features and frequency domain features according to their ranges to obtain a normalized feature matrix that can be used to calculate weights, thereby improving the accuracy of state recognition and avoiding the error problem existing in a single sensor. The present invention adopts an entropy weight method to calculate the weight matrix corresponding to the feature matrix, thereby avoiding the influence of feature indicators with smaller fluctuations on the final state recognition result. Taking into account the "short board effect" that may be produced by the sensor, the sensor characteristics with poor indicators are given special consideration. The electric spindle state recognition method proposed by the present invention has strong applicability and can be widely used in electric spindles in various states. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the electric spindle state assessment method based on multi-source information weighted fusion in an embodiment of the present invention.

[0042] Figure 2 Schematic diagram of spindle status indicator clustering in an embodiment of the present invention.

[0043] Figure 3Schematic diagram of the spindle health grading model in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0045] The present invention provides a spindle state assessment method based on a multi-source information deep coupling network, comprising the following steps:

[0046] S1, collects multi-source sensor information for characterizing electric spindle faults;

[0047] Specifically, multiple vibration sensors, multiple current sensors and multiple temperature sensors are installed on the spindle to be tested to obtain multi-source sensing information including vibration signals, current signals and temperature signals for characterizing the electric spindle fault, ensuring that the information content of the measured signals is rich enough to obtain accurate spindle status assessment results.

[0048] S2: Extract time domain and frequency domain features of multi-source sensor information; time domain features include peak value, root mean square, kurtosis index and pulse index; frequency domain features include the electric spindle rotation frequency energy and double rotation frequency energy in high-frequency signals;

[0049] S3: constructing a descending ridge membership model based on the acquired time domain features and frequency domain features, and adaptively normalizing the time domain features and frequency domain features according to their ranges to obtain a normalized feature matrix that can be used to calculate weights;

[0050] S4: Calculate the characteristic information entropy according to the characteristic matrix, and use the entropy weight method to calculate the weight matrix corresponding to the characteristic matrix;

[0051] S5: using the equalization function to calculate the discreteness of the characteristic matrix, calculating the variable weight vector corresponding to the multi-source sensor information according to the discreteness of the characteristic matrix, and calculating the optimized weight matrix according to the variable weight vector;

[0052] S6: Perform feature matrix weighted fusion on the optimized weight matrix to obtain the spindle state index vector, and calculate the spindle state classification result according to the gray clustering whitening weight function.

[0053] Combine Figure 1The present invention first completes the collection of the spindle vibration signal, current signal and temperature signal, and then denoises, filters and smoothes the collected vibration signal, current signal and temperature signal; then extracts the characteristic indicators of the vibration signal, current signal and temperature signal, calculates the weight after normalizing the acquired characteristic indicators, and calculates the optimized weight according to the variable weight function, and finally obtains the spindle status indicator by weighted fusion of the characteristic indicators, which is input into the spindle status health grading model function to classify the electric spindle status.

[0054] Combine Figure 2 When changing the weights in step 5, K-means clustering is used to obtain the required parameters in the function.

[0055] Combine Figure 3 ,When calculating the spindle status in step 6, the spindle status health grading model function is used to calculate the health grade category of the electric spindle.

[0056] The vibration sensors installed in S1 are located on the spindle housing on both sides of the electric spindle tool to monitor tool vibration; the spindle housing on both sides of the electric spindle front bearing monitors spindle bearing vibration. A current sensor is connected to the motor cable in the electrical cabinet to monitor current fluctuations in the electric spindle motor.

[0057] The specific process of extracting time domain features and frequency domain features in S2 is as follows:

[0058] (2.1) Using a low-pass filter to filter the vibration signal and the current signal to obtain a signal including low-frequency components;

[0059] (2.2) Using MATLAB, we calculated the time domain and frequency domain characteristics of the low-frequency component signals, obtaining 12 characteristic quantities of the two vibration sensors and six characteristic quantities of the current sensor.

[0060] (2.3) Smooth the temperature signal to obtain the point value signal of the temperature, and construct a feature matrix together with the feature quantity obtained in step (2.2). The feature matrix contains n samples, and each sample has 19 parameters.

[0061] The descending ridge membership function constructed in step S3 is:

[0062]

[0063] Select the parameters a and b in the semi-ridge membership function. For each column of parameters, select 90% and 10% of its maximum value as parameters b and a respectively.

[0064] The 19 columns of parameter vectors are respectively subjected to semi-ridge membership normalization to obtain the normalized feature matrix.

[0065] The process of calculating the weight matrix according to the entropy weight method in step S4 is:

[0066] Calculate the information entropy of each indicator and calculate the information entropy value corresponding to the jth indicator according to the following formula:

[0067]

[0068] Where n is the total number of samples and m is the number of indicator types. ij =0, that is, the current indicator score is 0, then define This satisfies the lower limit of entropy to 0.

[0069] According to the entropy weight calculation formula, the preliminary weight corresponding to the characteristic index is calculated using the following formula:

[0070]

[0071] Entropy weight can reduce the impact of indicators with smaller fluctuations and avoid misjudgment problems caused by the insensitivity of the sensor itself.

[0072] The calculated feature index weights are combined to obtain the weight matrix W = [w1,w2,…,w m ].

[0073] In step S5, the balancing function is used to calculate the discreteness of the characteristic matrix, and the variable weight vector corresponding to the multi-source sensor information is calculated according to the discreteness of the characteristic matrix. The process of calculating the optimized weight matrix according to the variable weight vector is as follows:

[0074] (5.1) Calculate the discreteness of the indicator eigenvector:

[0075]

[0076] Where j is the number of indicator types, x j A specific single feature.

[0077] (5.2) Construct a state balance function, consider the “short board effect” of sensor parameters, and assign higher weights to the sensor features with poor performance. The balance function is as follows:

[0078]

[0079] In the formula, r is the discreteness of the factor state vector, X i is the state value of indicator i, m is the number of indicators in the indicator group, α and β are the critical values ​​of the indicator state for punishment and incentive respectively, (0, α) is the penalty zone, [α, β] is the "incentive-punishment balance interval", and (β, 1] is the incentive zone. Among them, ω and ξ are variable weight constraint parameters, which are set to 0.3 and 0.1 respectively.

[0080] (5.3) The parameters α and β in the equilibrium function are determined by the K-means algorithm. After obtaining the preliminary feature index weights using the entropy weight method, they are multiplied and added with the corresponding index to obtain the preliminary spindle state index. The spindle state index is clustered by K-means and the cluster center values ​​at both ends are taken as the incentive parameters.

[0081] (5.4) The state variable weight vector S(X) is calculated by the equilibrium function, and the weight vector W = [w1,w2,…,w m ]Calculate the weight matrix after optimizing the weights:

[0082]

[0083] The process of obtaining the spindle status classification in step S6 is as follows:

[0084] (6.1) According to the weight matrix and normalized characteristic matrix after weight optimization, the spindle state index is calculated:

[0085] h=m*w T

[0086] Among them, h is the main axis state index, m is the characteristic index vector, and w is the characteristic index weight obtained by weighting. Finally, the main axis state index h corresponding to each sample is calculated. i , and merge them into the final spindle state index vector H=[h1,h2,h3...h n ].

[0087] (6.2) The spindle status indicators are clustered by the grayscale whitening weight function. The spindle status health grading model function based on the whitening weight function is as follows:

[0088] The function indicating the optimal spindle state is:

[0089]

[0090] The function indicating that the spindle is in good condition is

[0091]

[0092] The function representing the spindle state difference is:

[0093]

[0094] In the above spindle health status classification model function, parameters a, b, c, and d are model parameters, which are given according to the spindle status parameters themselves and are corrected by considering the distribution of the collected spindle data.

[0095] (6.3) Input the indicators into the health status grading model to obtain the main axis status category corresponding to the sample and realize the main axis status classification.

[0096] The present invention aims to solve the problem of electric spindle state assessment by adopting a collaborative fusion method of multiple sensors, thereby improving the accuracy of state recognition and avoiding the error problem existing in a single sensor.

[0097] The present invention proposes a new entropy weight calculation method to avoid the influence of characteristic indicators with smaller fluctuations on the final state recognition results; the present invention takes into account the "short board effect" that may be produced by the sensor and focuses on the sensing characteristics with poor indicators.

[0098] The electric spindle state identification method proposed in the present invention has strong applicability and can be widely applied to electric spindles in various states.

Claims

1. A spindle condition assessment method based on a multi-source information deep coupling network, characterized in that: The following steps are involved: S1, collects multi-source sensor information for characterizing electric spindle faults; S2: Extract time domain features and frequency domain features of multi-source sensing information; S3: constructing a descending ridge membership model based on the acquired time domain features and frequency domain features, and adaptively normalizing the time domain features and frequency domain features according to their ranges to obtain a normalized feature matrix that can be used to calculate weights; S4: Calculate the characteristic information entropy according to the characteristic matrix, and use the entropy weight method to calculate the weight matrix corresponding to the characteristic matrix; S5: using the equalization function to calculate the discreteness of the characteristic matrix, calculating the variable weight vector corresponding to the multi-source sensor information according to the discreteness of the characteristic matrix, and calculating the optimized weight matrix according to the variable weight vector; S6: Perform feature matrix weighted fusion on the optimized weight matrix to obtain the spindle state index vector, and use the gray clustering whitening weight function to calculate the spindle state index vector to obtain the spindle state classification result.

2. The spindle state assessment method based on a multi-source information deep coupling network according to claim 1 is characterized in that: The multi-source sensing information includes vibration signals, current signals and temperature signals used to characterize the fault of the electric spindle.

3. The spindle state assessment method based on a multi-source information deep coupling network according to claim 1 is characterized in that: The time domain features include peak value, root mean square, kurtosis index and pulse index, and the frequency domain features include the electric spindle rotation frequency energy and double rotation frequency energy in the high-frequency signal.

4. The spindle state assessment method based on a multi-source information deep coupling network according to claim 1 is characterized in that: The vibration signal and the current signal are filtered using a low-pass filter to obtain a signal including low-frequency components; The time domain and frequency domain features of the low-frequency component signals were calculated using MATLAB, and 12 feature quantities of the two vibration sensors and six feature quantities of the current sensor were obtained. The temperature signal is smoothed to obtain the point value signal of the temperature, and the feature matrix is ​​constructed together with the feature quantity obtained in step (2.2). The feature matrix contains x samples, and each sample has a total of 19 parameters.

5. The spindle state assessment method based on a multi-source information deep coupling network according to claim 4 is characterized in that: The constructed descending ridge membership function is: Select the parameters a and b in the semi-ridge membership function. For each column of parameters, select 90% and 10% of its maximum value as parameters b and a respectively.

6. The spindle state assessment method based on a multi-source information deep coupling network according to claim 1 is characterized in that: Calculate the information entropy of each indicator and calculate the information entropy value corresponding to the jth indicator according to the following formula: Among them, if z ij =0, that is, the current indicator score is 0, then define This satisfies the lower limit of entropy to 0.

7. The spindle state assessment method based on a multi-source information deep coupling network according to claim 1 is characterized in that: Compute the discreteness of the characteristic matrix: Construct state equilibrium function: In the formula, r is the discreteness of the factor state vector, X i is the state value of indicator i, m is the number of indicators in the indicator group, τ is the "incentive-penalty ratio", α and β are the indicator state critical values ​​of punishment and incentive respectively, (0, α) is the penalty zone, [α, β] is the "incentive-penalty balance interval", and (β, 1] is the incentive zone; ω and ξ are variable weight constraint parameters, which are set to 0.3 and 0.1 respectively.

8. The spindle state assessment method based on a multi-source information deep coupling network according to claim 7 is characterized in that: The parameters α and β in the equilibrium function are determined by the K-means algorithm. After obtaining the preliminary characteristic index weights using the entropy weight method, they are multiplied and added with the corresponding index to obtain the preliminary spindle state index. The spindle state index is clustered by K-means and the cluster center values ​​at both ends are taken as the incentive and penalty parameters. The state variable weight vector S(X) is calculated by the equilibrium function, and the weight vector W=[w1,w2,…,w n ]Calculate the weight matrix after optimizing the weights:

9. The spindle state assessment method based on a multi-source information deep coupling network according to claim 1 is characterized in that: According to the weight matrix after optimizing the weights and the normalized characteristic matrix, the spindle status index is calculated: h=m*w T Among them, h is the main axis state index, m is the characteristic index vector, and w is the characteristic index weight obtained by variable weight. Finally, the main axis state index h corresponding to each sample is calculated. i , and merge them into the final spindle state index vector H=[h1,h2,h3...h i ].

10. A spindle state assessment system based on a multi-source information deep coupling network, characterized in that: Including preprocessing module, optimization module and classification module: A preprocessing module is used to collect multi-source sensor information used to characterize the electric spindle fault and extract the time domain characteristics and frequency domain characteristics of the multi-source sensor information; The optimization module constructs a descending ridge membership model based on the acquired time domain features and frequency domain features, and adaptively normalizes the time domain features and frequency domain features according to their ranges to obtain a normalized feature matrix that can be used to calculate weights; Calculate the characteristic information entropy according to the characteristic matrix, calculate the weight matrix corresponding to the characteristic matrix using the entropy weight method, calculate the discreteness of the characteristic matrix using the balance function, calculate the variable weight vector corresponding to the multi-source sensor information according to the discreteness of the characteristic matrix, and calculate the optimized weight matrix according to the variable weight vector; The classification module is used to perform feature matrix weighted fusion on the optimized weight matrix to obtain the spindle state index vector, and the spindle state classification result is obtained by using the grey clustering whitening weight function to calculate the spindle state index vector.

Images