Data management system and method of industrial equipment four-wire machine based on big data

Abstract

The invention relates to the technical field of industrial equipment data management, and particularly discloses a data management system and method for an industrial equipment four-wire machine based on big data, and the method comprises the steps: obtaining a high-frequency vibration time domain signal and a working condition parameter of the four-wire machine; looking up a table based on the working condition parameters to obtain initialization parameters, and performing variational mode decomposition to obtain a finite bandwidth mode component; performing multi-scale sample entropy calculation on each component, and constructing an entropy feature vector; inputting the feature vector into a support vector machine to obtain an attribute tag of the vibration event; associating the attribute tag and the feature vector to the production batch and the operation and maintenance record data, and dynamically correcting the discrimination threshold according to the data set; according to the method, the problem of false alarm and missing alarm caused by difficulty in distinguishing working condition disturbance and fault symptoms is solved, and accurate sensing of the running state of the four-wire machine and self-adaptive dynamic adjustment of the distinguishing threshold are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment data management technology, specifically to a data management system and method for a four-line machine for industrial equipment based on big data. Background Technology

[0002] Four-thread overlock sewing machines (hereinafter referred to as four-thread machines) are core sewing equipment in the garment manufacturing industry, and their operating status directly determines production efficiency and product quality. With the development of industrial internet and big data technology, more and more garment manufacturing companies are beginning to deploy various sensors on four-thread machines to collect operating data such as vibration, current, and speed to achieve equipment status monitoring and predictive maintenance. Currently, data management methods for four-thread machines mainly adopt single sensor threshold alarms or spectrum analysis technology based on Fourier transform.

[0003] The existing technology has the following problems: Existing data management methods for four-thread sewing machines generally employ static thresholding or single-spectrum analysis techniques when processing high-frequency vibration signals. However, in actual sewing workshop environments, the severe impacts generated when four-thread sewing machines sew thick layers or hard linings are highly similar to and overlap with fault symptoms such as early bearing pitting and needle breakage in the time and frequency domains. This makes it impossible for existing methods to effectively distinguish between the two, leading to a technical dilemma of "false alarms and missed alarms." At the same time, due to the lack of a feedback mechanism for the results of manual verification, the discrimination rules cannot be adaptively adjusted with the aging process of the equipment and changes in operating conditions. As a result, the false alarm rate and missed alarm rate continue to rise over time, ultimately causing predictive maintenance systems to be abandoned by field personnel due to frequent invalid alarms, making it impossible to achieve accurate perception and dynamic management of equipment health status. Summary of the Invention

[0004] The purpose of this invention is to provide a data management system and method for a four-line machine for industrial equipment based on big data, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: The data management method for industrial equipment four-line machine based on big data, S1, acquires the high-frequency vibration time domain signal collected in real time by the vibration sensing unit deployed at the key transmission components of the four-line machine, and acquires the working condition parameters related to the current speed of the spindle of the four-line machine. S2, based on the working condition parameters, look up the table to obtain the preset filter bank initialization parameters, and use the initialization parameters to perform variational mode decomposition on the high-frequency vibration time domain signal to obtain several finite bandwidth mode components with preset center frequencies; S3, calculate the sample entropy for each finite bandwidth modal component on multiple coarse-grained time scales to obtain the multi-scale entropy value corresponding to each modal component, and construct an entropy feature vector representing the current operating state from all multi-scale entropy values; S4. Input the entropy feature vector into the classification and discrimination layer based on the support vector machine structure. The classification and discrimination layer outputs the attribute label of the vibration event according to the preset decision logic. The attribute label is used to indicate whether the current vibration belongs to the working condition disturbance event or the fault symptom event. S5 associates the attribute labels and corresponding entropy feature vectors with the production batch data and operation and maintenance record data of the four-line machine to form a time-series associated data set, and dynamically corrects the discrimination threshold of subsequent vibration events based on the data set.

[0006] As a further aspect of the present invention: S2 specifically includes: Based on the spindle speed value in the working condition parameters, a pre-built speed-parameter mapping table is queried to obtain the modal component quantity K value, the initial value of the center frequency of each modal component, the quadratic penalty factor, and the fidelity balance parameter corresponding to the corresponding spindle speed value. The speed-parameter mapping table is obtained in advance by conducting modal experiments on the same model of four-line machine at different speeds. The high-frequency vibration time-domain signal is used as the signal to be decomposed, and the modal component quantity K, the initial value of the center frequency of each modal component, the quadratic penalty factor, and the fidelity balance parameter are substituted into the iterative solution process of variational mode decomposition. In the frequency domain, the center frequency and bandwidth of each modal component are updated cyclically using the alternating direction multiplier method until the preset convergence accuracy condition is met, at which point the iteration stops, resulting in K finite bandwidth modal components with preset center frequencies.

[0007] As a further aspect of the present invention: S3 specifically includes: For each finite bandwidth mode component, coarse-graining is performed sequentially. The original data sequence is segmented and averaged in ascending order of scale factor from 1 to 20 to generate 20 coarse-grained sequences corresponding to each scale factor. For each coarse-grained sequence, the sample entropy is calculated, the embedding dimension is set to 2, and the similarity tolerance is set to 0.15 times the standard deviation of the corresponding coarse-grained sequence, so as to obtain the sample entropy values ​​of the corresponding modal components at 20 time scales. The sample entropy values ​​of each modal component at the smallest scale are removed to filter out high-frequency noise interference. The sample entropy values ​​at the remaining scales are then concatenated in ascending order of scale factor to form a multidimensional entropy feature vector representing the current operating state.

[0008] As a further aspect of the present invention: S4 specifically includes: Receive the entropy feature vector and map it to the high-dimensional feature space defined by the support vector machine with radial basis functions as kernel functions. Calculate the sign distance of the entropy feature vector relative to the classification hyperplane of the support vector machine. The symbolic distance is input into a preset softening decision function, which converts the symbolic distance into a probability output value ranging from 0 to 1. The probability output value is compared with the preset high threshold and low threshold. When the probability output value is higher than the high threshold, a fault symptom event label is output. When the probability output value is lower than the low threshold, a working condition disturbance event label is output. When the probability output value is between the high threshold and the low threshold, a pending label is output and a manual review instruction is triggered.

[0009] As a further aspect of the present invention: the signed distance of the entropy-valued feature vector relative to the support vector machine classification hyperplane specifically includes: Read the pre-stored set of support vectors, as well as the Lagrange multipliers and class labels corresponding to each support vector. The set of support vectors is obtained by offline training on historical entropy feature vector samples. Using the radial basis function as the kernel function, the inner product of the current entropy feature vector and each support vector in the support vector set in the high-dimensional feature space is calculated respectively. All inner product values ​​are weighted and summed according to the corresponding Lagrange multipliers and class labels. After adding the bias term, the output value of the decision function is obtained. The sign and absolute value of the output value of the decision function are used as the sign distance of the current entropy feature vector relative to the classification hyperplane, where the sign indicates the side to which it belongs and the absolute value indicates the degree of deviation from the hyperplane.

[0010] As a further aspect of the present invention: S5 specifically includes: Align the attribute tags and their corresponding entropy feature vectors with the current production batch data and operation and maintenance record data using timestamps to generate associated data records containing equipment status and production and operation and maintenance information, and store them in the historical database; Retrieve all relevant data records containing manual verification and confirmation tags within the most recent preset period from the historical database to form a threshold correction sample set. The manual verification and confirmation tags include fault symptom events and operating condition disturbance events that have been corrected after verification by on-site personnel. The statistical threshold correction method analyzes the distribution pattern of the sign distance between the entropy feature vector corresponding to fault symptom events in the sample set and the classification hyperplane. Multiple quantiles of the cumulative distribution function are used as new threshold boundaries, and the high and low thresholds in the softening decision function are updated using the new threshold boundaries.

[0011] As a further aspect of the present invention: the distribution pattern of the symbolic distance between the entropy feature vectors corresponding to fault symptom events in the statistical threshold correction sample set and the classification hyperplane specifically includes: Extract all associated data records marked as fault symptom events from the threshold-corrected sample set, obtain the symbolic distance value corresponding to each record, and form a fault symptom symbolic distance set; The kernel density estimation method is used to fit the probability density distribution of the fault symptom symbol distance set, and a smoothing function is used to approximate the true distribution of the symbol distance to generate a continuous probability density curve. The lower and upper quantiles corresponding to the preset confidence levels are located at the tails on both sides of the probability density curve, and the lower and upper quantiles are determined as the low and high threshold update values ​​of the softening decision function, respectively.

[0012] As a further aspect of the present invention: the generation of a continuous probability density curve specifically includes: Read all symbol distance values ​​from the fault symptom symbol distance set and determine that the kernel function used for kernel density estimation is a Gaussian kernel function; Based on the sample size and data dispersion of the fault symptom symbol distance set, the optimal smoothing bandwidth corresponding to the current set is calculated by solving the bandwidth selection criterion that minimizes the asymptotic integral mean square error. Centered on the distance value of each fault symptom symbol, probability density components are constructed according to the Gaussian kernel function and the optimal smoothing bandwidth. All components are superimposed and divided by the product of the sample size and the bandwidth to obtain a continuous and smooth probability density curve.

[0013] A data management system for four-line industrial equipment based on big data includes: The data sensing and acquisition module acquires high-frequency vibration time-domain signals in real time from the vibration sensing unit deployed at the key transmission components of the four-thread machine, as well as operating parameters related to the current speed of the spindle of the four-thread machine. The parameter configuration decomposition module obtains the preset filter bank initialization parameters by looking up the table based on the operating condition parameters, and uses the initialization parameters to perform variational mode decomposition on the high-frequency vibration time domain signal to obtain several finite bandwidth mode components with preset center frequencies. The multi-scale entropy feature extraction module calculates the sample entropy for each finite bandwidth modal component at multiple coarse-grained time scales, obtains the multi-scale entropy value corresponding to each modal component, and constructs an entropy feature vector representing the current running state from all multi-scale entropy values. The support vector classification and discrimination module inputs the entropy feature vector into the classification and discrimination layer based on the support vector structure. The classification and discrimination layer outputs the attribute label of the vibration event according to the preset decision logic. The attribute label is used to indicate whether the current vibration belongs to the working condition disturbance event or the fault symptom event. The closed-loop threshold optimization module associates attribute labels and corresponding entropy feature vectors with the production batch data and operation and maintenance record data of the four-line machine to form a time-series correlated data set, and dynamically corrects the discrimination threshold for subsequent vibration events based on the data set.

[0014] The beneficial effects of this invention are: (1) This invention introduces a feature extraction method that combines variational mode decomposition and multi-scale sample entropy, which can effectively separate normal operating disturbances from early fault symptoms in high-frequency vibration signals, solving the problem of high false alarm rate of traditional threshold method under complex operating conditions such as thick material sewing and hard lining impact. At the same time, based on the symbol distance and softened decision function output by support vector machine, the classification results are converted into probability output and high and low thresholds are set, which reduces random false alarms caused by signal fluctuations, reduces the frequency of invalid alarms, and improves the trust of on-site maintenance personnel in the monitoring system.

[0015] (2) This invention constructs a closed-loop feedback mechanism that includes manual verification labels. By estimating the kernel density of the symbol distance corresponding to fault symptom events, the high and low thresholds of the softening decision function are dynamically updated. As equipment operation data continues to accumulate, the discrimination boundary can be adaptively adjusted, avoiding the problem of fixed threshold failure caused by factors such as equipment aging and changes in operating conditions. This dynamic correction mechanism enables the system to maintain stable discrimination performance during long-term operation, reduces the frequency of manual intervention, and lowers the implementation cost of predictive maintenance. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1As shown, this invention is a data management method for a four-line industrial equipment based on big data, comprising the following steps: S1, acquire the high-frequency vibration time-domain signal collected in real time by the vibration sensing unit deployed at the key transmission components of the four-wire machine, and acquire the working condition parameters related to the current speed of the spindle of the four-wire machine; S2, based on the working condition parameters, look up the table to obtain the preset filter bank initialization parameters, and use the initialization parameters to perform variational mode decomposition on the high-frequency vibration time domain signal to obtain several finite bandwidth mode components with preset center frequencies; S3, calculate the sample entropy for each finite bandwidth modal component on multiple coarse-grained time scales to obtain the multi-scale entropy value corresponding to each modal component, and construct an entropy feature vector representing the current operating state from all multi-scale entropy values; S4. Input the entropy feature vector into the classification and discrimination layer based on the support vector machine structure. The classification and discrimination layer outputs the attribute label of the vibration event according to the preset decision logic. The attribute label is used to indicate whether the current vibration belongs to the working condition disturbance event or the fault symptom event. S5 associates the attribute labels and corresponding entropy feature vectors with the production batch data and operation and maintenance record data of the four-line machine to form a time-series associated data set, and dynamically corrects the discrimination threshold of subsequent vibration events based on the data set.

[0020] In S1, high-frequency vibration time-domain signals are acquired in real time from vibration sensing units deployed at key transmission components of the four-thread machine, and operating parameters related to the current spindle speed of the four-thread machine are also acquired, specifically including: First, vibration sensing units are deployed at key transmission components of the four-thread sewing machine. These key transmission components include the spindle bearing housing, the needle bar mechanism connection, and the feed dog drive shaft end cap. The vibration sensing unit uses a piezoelectric accelerometer, fixed to the outer surface of these components via a threaded connection or magnetic base, to sense the waveform changes of mechanical vibration in the time domain. The vibration sensing unit continuously acquires raw vibration signals. This sampling frequency is determined based on the fundamental frequency harmonics corresponding to the highest spindle speed of the four-thread sewing machine to ensure the capture of high-frequency resonant components corresponding to fault symptoms. The acquired data is transmitted in real-time to the edge computing node in the form of a continuous time-domain waveform, forming a high-frequency vibration time-domain signal.

[0021] Meanwhile, this step obtains the current spindle speed value through a rotary encoder deployed at the end of the four-thread sewing machine spindle. The rotary encoder outputs several pulses per revolution, and the edge computing node converts the number of pulses per unit time into spindle speed data in revolutions per minute. In addition, the instantaneous value of the motor load current is obtained through a current transformer connected in series with the power supply line of the four-thread sewing machine's main motor. This current value reflects the load fluctuation of the current sewing operation. The spindle speed value and the motor load current value together constitute the operating parameters related to the current operating status.

[0022] Edge computing nodes synchronize and bind high-frequency vibration time-domain signals with operating parameters according to a unified timestamp alignment rule, providing raw input data with time consistency for subsequent data processing steps.

[0023] In S2, preset filter bank initialization parameters are obtained by looking up a table based on operating condition parameters. These initialization parameters are then used to perform variational mode decomposition on the high-frequency vibration time-domain signal, yielding several finite-bandwidth modal components with preset center frequencies. Specifically, these include: This step first uses the spindle speed value obtained in step S1 to query a pre-constructed speed-parameter mapping table. The speed-parameter mapping table is constructed as follows: a prototype of the same model as the four-wire machine to be monitored is selected and installed on a dedicated test bench. Under no-load conditions, the spindle speed is gradually increased from the lowest operating speed to the highest operating speed, with a test speed point set at intervals of 100 revolutions per minute. At each test speed point, vibration signals are collected using vibration sensors, and the signals are subjected to spectral analysis to determine the main frequency band range where vibration energy is concentrated at that speed. Based on the spectral analysis results, the number of modal components corresponding to that speed point is determined. This number is taken as the smallest integer that can cover the main frequency band range, typically between 4 and 8. Simultaneously, the center frequency of each main frequency band is recorded as the initial value of the center frequency of each modal component corresponding to that speed point. Furthermore, for each test speed point, a grid search method is used to optimize and determine the secondary penalty factor and fidelity balance parameters. The optimization objective is to minimize the sum of the bandwidths of the decomposed modal components and the degree of aliasing between the components. All test speed points and their corresponding modal component numbers, initial values ​​of center frequencies of each modal component, quadratic penalty factors, and fidelity balance parameters are stored to form a speed-parameter mapping table.

[0024] When entering the online processing flow, this step receives the high-frequency vibration time-domain signal and operating parameters output from step S1, and extracts the current spindle speed value from the operating parameters. This spindle speed value is compared with the test speed points in the speed-parameter mapping table to find the test speed point closest to the current spindle speed value. The number of modal components, the initial value of the center frequency of each modal component, the quadratic penalty factor, and the fidelity balance parameters corresponding to this test speed point are read from the mapping table.

[0025] This step then uses the high-frequency vibration time-domain signal as the signal to be decomposed, and uses the number of modal components, the initial values ​​of the center frequencies of each modal component, the quadratic penalty factor, and the fidelity balance parameters obtained from the above table lookup as the initial input conditions for variational mode decomposition. The mathematical essence of variational mode decomposition is to solve a constrained variational problem, the goal of which is to minimize the sum of the estimated bandwidths of each modal component under the constraint that the sum of the modal components equals the original signal.

[0026] This step employs the alternating direction multiplier method to solve the variational problem described above. The solution process begins by setting the initial center frequencies of each modal component to values ​​read from the mapping table, and initializing the frequency domain representation of each modal component to zero. Then, iterative iterations are performed: First, the center frequencies of all modal components are fixed, and the frequency domain representation of each modal component is updated using Wiener filtering. Second, the frequency domain representation of all modal components is fixed, and the center frequencies of each modal component are updated by moving the center frequencies to the centroid position of the power spectrum in the corresponding modal component's frequency domain representation. Third, the Lagrange multipliers are updated to strengthen the constraints. After each round of these three steps, the change in the frequency domain representation of each modal component between the current iteration and the previous iteration is calculated. Iteration stops when the changes in all modal components are less than a preset convergence accuracy threshold. The convergence accuracy threshold is typically set to... After the iteration is completed, the frequency domain representation of each modal component is converted back to the time domain through inverse Fourier transform to obtain a number of finite bandwidth modal components. Each modal component has a center frequency that is finally determined after iterative updates from the initial value, and each modal component is separated from each other in the frequency domain, together forming a complete decomposition of the original high-frequency vibration time domain signal.

[0027] In S3, sample entropy is calculated for each finite bandwidth modal component at multiple coarse-grained time scales to obtain multi-scale entropy values ​​corresponding to each modal component. An entropy feature vector characterizing the current operating state is then constructed from all multi-scale entropy values, specifically including: This step first receives several finite bandwidth modal components output from step S2. Each modal component is a discrete time sequence of length N. The value of N is determined according to the sampling frequency and the analysis window duration, and is usually 1024 or 2048 data points.

[0028] This step performs coarse-graining processing sequentially for each finite-bandwidth modal component. The purpose of coarse-graining is to examine the complexity of the time series across multiple time scales. The specific operation is as follows: The scale factor is set to an integer range from 1 to 20. For each scale factor value, the original time series is segmented according to that scale factor, with each segment containing the same number of data points as the scale factor value. The arithmetic mean of all data points within each segment is calculated, and this average is used as the coarse-grained data point for that segment. The averages calculated from all segments are arranged sequentially in chronological order, forming a coarse-grained sequence whose length is shortened to the original sequence length divided by the scale factor. This process is repeated for scale factors 1 to 20, ultimately generating 20 coarse-grained sequences of different lengths from the same finite-bandwidth modal component. When the scale factor is 1, the coarse-grained sequence is the original sequence itself.

[0029] This step then calculates the sample entropy for each coarse-grained sequence to quantify the regularity and complexity of each sequence at the corresponding time scale. The specific process of sample entropy calculation is as follows: For a given coarse-grained sequence of length M, the embedding dimension is first set to 2. Two consecutive data points in the sequence are grouped into a set of template vectors, resulting in a total of M-2+1 template vectors, i.e., M-1 template vectors. Then, the similarity tolerance is set to 0.15 times the standard deviation of the coarse-grained sequence. For each template vector, the distance between it and all other template vectors is counted. The distance is defined as the maximum absolute value of the difference between corresponding elements of the two vectors. The number of template vectors with distances less than the similarity tolerance is counted, and the ratio of this number to the total number of template vectors M-1 is calculated, denoted as the first-class counting ratio. Next, the embedding dimension is increased to 3, and three consecutive data points in the sequence are grouped into a set of template vectors, resulting in a total of M-3+1 template vectors, i.e., M-2 template vectors. Similarly, using a similarity tolerance of 0.15 times the standard deviation of the coarse-grained sequence, the number of template vectors with a distance less than the similarity tolerance is counted, and the ratio of this number to the total number of template vectors M minus 2 is calculated, denoted as the second-type counting ratio. Finally, the sample entropy value is calculated as the natural logarithm of the ratio of the first-type counting ratio to the second-type counting ratio. Following the above calculation process, the sample entropy values ​​of the current modal component are calculated for each of the 20 coarse-grained sequences corresponding to the current modal component at 20 time scales.

[0030] This step involves constructing feature vectors after obtaining the entropy values ​​of all samples for all modal components. Considering that the coarse-grained sequence corresponding to the smallest scale factor 1 is the original sequence, the sample entropy values ​​at this scale are easily affected by high-frequency noise components in the original signal, leading to decreased feature stability. Therefore, this step first removes the sample entropy values ​​of each modal component at scale factor 1, retaining the sample entropy values ​​at 19 time scales from scale factor 2 to 20. Subsequently, the sample entropy values ​​of all modal components at scale factors 2 to 20 are concatenated sequentially according to the modal component order, and within each modal component, in ascending order of scale factor, to form a multidimensional entropy feature vector. The dimension of this feature vector is equal to the number of modal components multiplied by 19, and it is used to characterize the complexity distribution characteristics of the current four-wire machine's operating state across multiple time scales.

[0031] In S4, the entropy feature vector is input into a classification and discrimination layer based on a support vector machine. The classification and discrimination layer outputs attribute labels for the vibration event according to a preset decision logic. These attribute labels indicate whether the current vibration belongs to a working condition disturbance event or a fault symptom event, specifically including: This step first receives the entropy feature vector output from step S3. This vector is a multi-dimensional numerical vector, and its dimension is equal to the number of modal components multiplied by 19. The goal of this step is to determine whether the current vibration event belongs to normal operating condition disturbance or equipment failure symptom based on this feature vector.

[0032] To achieve the above discrimination, a support vector machine (SVM) classifier needs to be pre-constructed. The construction process of this classifier is as follows: A large number of entropy feature vector samples generated during historical operation are collected, each corresponding to a known vibration event type. Each sample is labeled with a class label through manual verification, assigning negative labels to samples corresponding to operational disturbance events and positive labels to samples corresponding to fault symptom events. All labeled samples are used as the training set, and a radial basis function (RBF) is used as the kernel function for training. The RBF is a natural exponential function whose exponent is the negative of the square of the difference between two vectors divided by the square of the kernel function parameter. The training process involves solving a convex quadratic programming problem to find an optimal classification hyperplane that can separate the two classes of samples as much as possible in the high-dimensional feature space while maximizing the classification margin. After training, all sample points located on the classification margin boundary are retained; these sample points are the support vectors. Each support vector corresponds to a Lagrange multiplier and a class label. Simultaneously, a bias term is also obtained during the training process. The set of support vectors, the Lagrange multipliers corresponding to each support vector, the class label, and the bias term are stored as parameters of the classifier for use in online discrimination.

[0033] When entering the online discrimination process, this step first reads the pre-stored set of support vectors, the Lagrange multipliers corresponding to each support vector, the class labels corresponding to each support vector, and the bias term. For the currently input entropy feature vector, a kernel function is calculated between it and each support vector. The specific method for calculating the kernel function is as follows: calculate the difference vector between the current entropy feature vector and the support vectors, and obtain the square of the Euclidean distance between the two vectors by summing the squares of the elements in the difference vector. Divide the square of the Euclidean distance by the square of the kernel function parameter, take the negative value, and then calculate the exponent with the natural constant e as the base to obtain the kernel function output value. The kernel function output value between the current entropy feature vector and each support vector is calculated in the same way. The kernel function output value corresponding to each support vector is multiplied by the Lagrange multiplier of that support vector, and then multiplied by the class label of that support vector to obtain the weighted contribution value corresponding to each support vector. The weighted contribution values ​​corresponding to all support vectors are summed, and then the bias term is added to obtain the decision function output value. The output value is a real number, whose sign indicates which side of the classification hyperplane the current entropy feature vector lies on, and its absolute value indicates the degree to which the current feature vector deviates from the classification hyperplane. This step uses the output value of the decision function as the signed distance of the current entropy feature vector relative to the classification hyperplane.

[0034] This step then inputs the aforementioned symbolic distance into a preset softening decision function, converting the symbolic distance into a probability output value ranging from 0 to 1. The softening decision function is constructed as follows: a slope parameter is set, which controls the smoothness of the mapping from the symbolic distance to the probability value. Using the natural constant e as the base, the slope parameter is multiplied by the negative value of the symbolic distance to calculate an exponent, yielding a first intermediate value. 1 is added to the first intermediate value to obtain a second intermediate value. 1 is divided by the second intermediate value to obtain the final probability output value. According to the above functional relationship, when the symbolic distance is a large positive value, the probability output value approaches 1; when the symbolic distance is a large negative value, the probability output value approaches 0; and when the symbolic distance is 0, the probability output value is 0.5. The slope parameter is determined during the offline training phase using cross-validation, typically taking a value of 1 or 2.

[0035] This step concludes by comparing the probability output value with preset high and low thresholds. The initial values ​​of the high and low thresholds are determined based on the statistical distribution of historical data; for example, the high threshold is set to 0.8 and the low threshold to 0.2. The comparison logic is as follows: When the probability output value is higher than the high threshold, it indicates that the current vibration signal highly matches the characteristics of historical fault symptom samples. This step outputs a fault symptom event label and triggers subsequent alarm or shutdown procedures. When the probability output value is lower than the low threshold, it indicates that the current vibration signal highly matches the characteristics of historical operating condition disturbance samples. This step outputs an operating condition disturbance event label without triggering an alarm. When the probability output value is between the high and low thresholds, it indicates that the current feature is in a region of ambiguous boundaries, and the classifier is unlikely to provide a high-confidence judgment. This step outputs an undetermined label and simultaneously pushes the current entropy feature vector and the corresponding vibration data to the manual verification terminal. On-site technicians then conduct final confirmation by observing the actual operating status of the equipment and using sound localization, and feed the confirmed true label back to the system for subsequent threshold optimization and classifier updates.

[0036] In S5, attribute labels and their corresponding entropy feature vectors are associated with the production batch data and maintenance record data of the four-line machine, forming a time-series correlated data set. The threshold for subsequent vibration events is then dynamically adjusted based on this data set, specifically including: This step first receives the attribute labels and their corresponding entropy feature vectors output from step S4, and simultaneously acquires the production batch data for the current production task and the current equipment maintenance record data. The production batch data includes the type of fabric being sewn, order number, operator information, and process parameter settings. The maintenance record data includes the equipment's most recent maintenance time, recorded faults, and records of replaced parts. This step aligns the attribute labels, entropy feature vectors, production batch data, and maintenance record data according to a unified timestamp, generating a linked data record containing equipment status and production maintenance information. The timestamp in this record is in milliseconds to ensure precise correspondence with the acquisition time of the original vibration data. All generated linked data records are then stored one by one in a historical database. This historical database adopts a time-series database architecture, supporting fast retrieval by time range and massive data storage.

[0037] When dynamic threshold correction is required, this step retrieves all relevant data records from the historical database within the most recent preset period. The preset period can be set to the past 7 days, 30 days, or determined based on the cumulative operating time of the equipment, for example, a threshold correction every 200 hours of operation. From the retrieved data records, those containing manually verified confirmation tags are selected to form the threshold correction sample set. The manually verified confirmation tag refers to the final confirmation of the vibration event type by field technicians after step S4 outputs a pending tag and triggers a manual verification command. This confirmation is achieved through methods such as observing the actual operating status of the equipment, listening for sound localization, or stopping the machine for inspection. The confirmed true tag (fault symptom event or operating condition disturbance event) is then fed back to the system, overwriting the original pending tag. These manually verified and corrected records have the highest confidence level and are suitable as the basis for threshold optimization.

[0038] The next step involves statistical analysis of the data in the threshold correction sample set to determine the new discrimination threshold. Specifically, the following steps are performed: First, extract all associated data records labeled as fault symptom events from the threshold correction sample set and obtain the symbolic distance value for each record. The symbolic distance is the output value of the decision function calculated in step S4; its sign indicates the side of the classification hyperplane, and its absolute value indicates the degree of deviation. These symbolic distance values ​​are then aggregated to form a one-dimensional dataset called the fault symptom symbolic distance set, denoted as set S, containing n sample points, each a real number. .

[0039] To characterize the distribution of symbolic distances corresponding to fault symptom events under real-world conditions, this step employs kernel density estimation to fit the probability density distribution of the dataset. Kernel density estimation is a non-parametric estimation method that does not presuppose that the data follows a specific distribution. Instead, it approximates the true probability density function by weighting each data point using a smoothing kernel function. In this step, a Gaussian function is chosen as the kernel function. The Gaussian kernel function is a natural exponential function whose exponent is the square of the difference between the independent variable and the data point divided by the negative value of twice the square of the bandwidth. Let the bandwidth be h. For any real number x, the kernel density estimate f(x) is calculated as follows: ; Where n is the sample size in the fault symptom symbol distance set. Let be the symbolic distance value for the i-th sample, h be the bandwidth parameter, and e be the natural constant. This formula represents the symbolic distance for each sample point. Centered on x, with bandwidth h controlling the smoothing width, the probability density estimate at x is obtained by superimposing the various Gaussian kernel functions and taking the average.

[0040] The value of bandwidth h directly affects the smoothness of kernel density estimation: too small a value of h will result in too many spikes in the curve, while too large a value will lead to over-smoothing and mask the true distribution characteristics. This step uses the criterion of minimizing the asymptotic integral mean square error to solve for the optimal bandwidth. Specifically, the standard deviation of the distance set of fault symptom symbols is calculated first, denoted as σ; then the interquartile range of the sample is calculated, denoted as R; the smaller of σ and R divided by 1.34 is taken as the reference scale, denoted as λ; based on the sample size n, an approximate value of the optimal bandwidth is calculated using the formula: bandwidth equals 0.9 multiplied by λ divided by the power of one-fifth of n. The bandwidth calculated in this way achieves a balance between estimation bias and variance, allowing the probability density curve to reflect the true distribution shape as closely as possible.

[0041] After obtaining the optimal bandwidth h, all sample points in the fault symptom symbol distance set are substituted into the kernel density estimation formula described above. For any real number x, the corresponding probability density value can be calculated. This step requires the cumulative distribution function of the probability density curve, i.e., the area under the integral from negative infinity to a certain value. Since fault symptom events correspond to the positive class, their symbol distances should theoretically be mostly positive, but in practice, a small number of negative values ​​may exist. This step focuses on the interval within which the symbol distances of fault symptom events should fall at a given confidence level.

[0042] Therefore, this step locates the lower and upper quantiles corresponding to the preset confidence level at the two tails of the probability density curve. The confidence level is set to 90%, meaning that the symbol distance of the fault symptom event is expected to fall between the lower and upper quantiles with a 90% probability. Specifically, the probability density function estimated by kernel density is numerically integrated, and the x-value when the cumulative distribution function value equals 0.05 is taken as the lower quantile, denoted as L; the x-value when the cumulative distribution function value equals 0.95 is taken as the upper quantile, denoted as U. Numerical integration can be performed using the trapezoidal rule or Simpson's rule, and the integration result is approximated by discrete sampling points in the computer program. L and U are the desired threshold boundaries.

[0043] Finally, this step determines the lower quantile L as the low threshold update value of the softening decision function and the upper quantile U as the high threshold update value. The updated high and low thresholds are written to a configuration file for use in subsequent online discrimination in step S4. In this way, the discrimination thresholds can be dynamically adjusted as equipment operating data accumulates and feedback from manual review results are received, ensuring that the classification boundary always adapts to the current health status and operating condition changes of the equipment. After multiple iterative optimizations, the high and low thresholds will gradually converge to stable values, effectively balancing the risks of false alarms and false negatives.

[0044] Please see Figure 2 As shown, the data management system for a four-line industrial equipment based on big data includes: The data sensing and acquisition module acquires high-frequency vibration time-domain signals in real time from the vibration sensing unit deployed at the key transmission components of the four-thread machine, as well as operating parameters related to the current speed of the spindle of the four-thread machine. The parameter configuration decomposition module obtains the preset filter bank initialization parameters by looking up the table based on the operating condition parameters, and uses the initialization parameters to perform variational mode decomposition on the high-frequency vibration time domain signal to obtain several finite bandwidth mode components with preset center frequencies. The multi-scale entropy feature extraction module calculates the sample entropy for each finite bandwidth modal component at multiple coarse-grained time scales, obtains the multi-scale entropy value corresponding to each modal component, and constructs an entropy feature vector representing the current running state from all multi-scale entropy values. The support vector classification and discrimination module inputs the entropy feature vector into the classification and discrimination layer based on the support vector structure. The classification and discrimination layer outputs the attribute label of the vibration event according to the preset decision logic. The attribute label is used to indicate whether the current vibration belongs to the working condition disturbance event or the fault symptom event. The closed-loop threshold optimization module associates attribute labels and corresponding entropy feature vectors with the production batch data and operation and maintenance record data of the four-line machine to form a time-series correlated data set, and dynamically corrects the discrimination threshold for subsequent vibration events based on the data set.

[0045] The working principle of this invention is as follows: High-frequency vibration time-domain signals and operating parameters related to the current spindle speed are acquired in real-time from vibration sensing units deployed at key transmission components of the four-line machine. Preset filter bank initialization parameters are obtained by looking up the operating parameters in a table, and variational mode decomposition is performed on the high-frequency vibration time-domain signals using these parameters to obtain several finite-bandwidth modal components with preset center frequencies. Sample entropy is calculated for each finite-bandwidth modal component on multiple coarse-grained time scales to obtain multi-scale entropy values ​​corresponding to each modal component. An entropy feature vector characterizing the current operating state is constructed from all multi-scale entropy values. The entropy feature vector is input to a classification and discrimination layer based on a support vector machine architecture. The classification and discrimination layer outputs attribute labels indicating whether the current vibration belongs to an operating condition disturbance event or a fault symptom event, according to a preset decision logic. The attribute labels and their corresponding entropy feature vectors are associated with the production batch data and maintenance record data of the four-line machine to form a time-series associated data set. The discrimination threshold for subsequent vibration events is dynamically adjusted based on this data set. This invention enables accurate perception and intelligent judgment of the operating status of a four-line machine. Through a closed-loop feedback mechanism, the judgment threshold is dynamically optimized based on equipment operating data and manual review results, effectively reducing the false alarm rate caused by operating condition disturbances and the missed alarm rate of early faults.

[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A data management method for a four-line machine in industrial equipment based on big data, characterized in that, Includes the following steps: S1, acquire the high-frequency vibration time-domain signal collected in real time by the vibration sensing unit deployed at the key transmission components of the four-wire machine, and acquire the working condition parameters related to the current speed of the spindle of the four-wire machine; S2, based on the working condition parameters, look up the table to obtain the preset filter bank initialization parameters, and use the initialization parameters to perform variational mode decomposition on the high-frequency vibration time domain signal to obtain several finite bandwidth mode components with preset center frequencies; S3, calculate the sample entropy for each finite bandwidth modal component on multiple coarse-grained time scales to obtain the multi-scale entropy value corresponding to each modal component, and construct an entropy feature vector representing the current operating state from all multi-scale entropy values; S4. Input the entropy feature vector into the classification and discrimination layer based on the support vector machine structure. The classification and discrimination layer outputs the attribute label of the vibration event according to the preset decision logic. The attribute label is used to indicate whether the current vibration belongs to the working condition disturbance event or the fault symptom event. S5 associates the attribute labels and corresponding entropy feature vectors with the production batch data and operation and maintenance record data of the four-line machine to form a time-series associated data set, and dynamically corrects the discrimination threshold of subsequent vibration events based on the data set.

2. The data management method for a four-line industrial equipment based on big data according to claim 1, characterized in that, S2 specifically includes: Based on the spindle speed value in the working condition parameters, a pre-built speed-parameter mapping table is queried to obtain the modal component quantity K value, the initial value of the center frequency of each modal component, the quadratic penalty factor, and the fidelity balance parameter corresponding to the corresponding spindle speed value. The speed-parameter mapping table is obtained in advance by conducting modal experiments on the same model of four-line machine at different speeds. The high-frequency vibration time-domain signal is used as the signal to be decomposed, and the modal component quantity K, the initial value of the center frequency of each modal component, the quadratic penalty factor, and the fidelity balance parameter are substituted into the iterative solution process of variational mode decomposition. In the frequency domain, the center frequency and bandwidth of each modal component are updated cyclically using the alternating direction multiplier method until the preset convergence accuracy condition is met, at which point the iteration stops, resulting in K finite bandwidth modal components with preset center frequencies.

3. The data management method for a four-line industrial equipment based on big data according to claim 1, characterized in that, S3 specifically includes: For each finite bandwidth mode component, coarse-graining is performed sequentially. The original data sequence is segmented and averaged in ascending order of scale factor from 1 to 20 to generate 20 coarse-grained sequences corresponding to each scale factor. For each coarse-grained sequence, the sample entropy is calculated, the embedding dimension is set to 2, and the similarity tolerance is set to 0.15 times the standard deviation of the corresponding coarse-grained sequence, so as to obtain the sample entropy values ​​of the corresponding modal components at 20 time scales. The sample entropy values ​​of each modal component at the smallest scale are removed to filter out high-frequency noise interference. The sample entropy values ​​at the remaining scales are then concatenated in ascending order of scale factor to form a multidimensional entropy feature vector representing the current operating state.

4. The data management method for a four-line industrial equipment based on big data according to claim 1, characterized in that, S4 specifically includes: Receive the entropy feature vector and map it to the high-dimensional feature space defined by the support vector machine with radial basis functions as kernel functions. Calculate the sign distance of the entropy feature vector relative to the classification hyperplane of the support vector machine. The symbolic distance is input into a preset softening decision function, which converts the symbolic distance into a probability output value ranging from 0 to 1. The probability output value is compared with the preset high threshold and low threshold. When the probability output value is higher than the high threshold, a fault symptom event label is output. When the probability output value is lower than the low threshold, a working condition disturbance event label is output. When the probability output value is between the high threshold and the low threshold, a pending label is output and a manual review instruction is triggered.

5. The data management method for a four-line industrial equipment based on big data according to claim 4, characterized in that, The signed distance between the entropy-valued feature vector and the support vector machine classification hyperplane specifically includes: Read the pre-stored set of support vectors, as well as the Lagrange multipliers and class labels corresponding to each support vector. The set of support vectors is obtained by offline training on historical entropy feature vector samples. Using the radial basis function as the kernel function, the inner product of the current entropy feature vector and each support vector in the support vector set in the high-dimensional feature space is calculated respectively. All inner product values ​​are weighted and summed according to the corresponding Lagrange multipliers and class labels. After adding the bias term, the output value of the decision function is obtained. The sign and absolute value of the output value of the decision function are used as the sign distance of the current entropy feature vector relative to the classification hyperplane, where the sign indicates the side to which it belongs and the absolute value indicates the degree of deviation from the hyperplane.

6. The data management method for a four-line industrial equipment based on big data according to claim 1, characterized in that, S5 specifically includes: Align the attribute tags and their corresponding entropy feature vectors with the current production batch data and operation and maintenance record data using timestamps to generate associated data records containing equipment status and production and operation and maintenance information, and store them in the historical database; Retrieve all relevant data records containing manual verification and confirmation tags within the most recent preset period from the historical database to form a threshold correction sample set. The manual verification and confirmation tags include fault symptom events and operating condition disturbance events that have been corrected after verification by on-site personnel. The statistical threshold correction method analyzes the distribution pattern of the sign distance between the entropy feature vector corresponding to fault symptom events in the sample set and the classification hyperplane. Multiple quantiles of the cumulative distribution function are used as new threshold boundaries, and the high and low thresholds in the softening decision function are updated using the new threshold boundaries.

7. The data management method for a four-line industrial equipment based on big data according to claim 6, characterized in that, The statistical threshold correction corrects the distribution pattern of the sign distance between the entropy feature vectors corresponding to fault symptom events in the sample set and the classification hyperplane, specifically including: Extract all associated data records marked as fault symptom events from the threshold-corrected sample set, obtain the symbolic distance value corresponding to each record, and form a fault symptom symbolic distance set; The kernel density estimation method is used to fit the probability density distribution of the fault symptom symbol distance set, and a smoothing function is used to approximate the true distribution of the symbol distance to generate a continuous probability density curve. The lower and upper quantiles corresponding to the preset confidence levels are located at the tails on both sides of the probability density curve, and the lower and upper quantiles are determined as the low and high threshold update values ​​of the softening decision function, respectively.

8. The data management method for a four-line industrial equipment based on big data according to claim 7, characterized in that, The generation of continuous probability density curves specifically includes: Read all symbol distance values ​​from the fault symptom symbol distance set and determine that the kernel function used for kernel density estimation is a Gaussian kernel function; Based on the sample size and data dispersion of the fault symptom symbol distance set, the optimal smoothing bandwidth corresponding to the current set is calculated by solving the bandwidth selection criterion that minimizes the asymptotic integral mean square error. Centered on the distance value of each fault symptom symbol, probability density components are constructed according to the Gaussian kernel function and the optimal smoothing bandwidth. All components are superimposed and divided by the product of the sample size and the bandwidth to obtain a continuous and smooth probability density curve.

9. A data management system for a four-line industrial equipment machine based on big data, characterized in that: The data management method for a four-line industrial equipment based on big data as described in any one of claims 1-8 includes: The data sensing and acquisition module acquires high-frequency vibration time-domain signals in real time from the vibration sensing unit deployed at the key transmission components of the four-thread machine, as well as operating parameters related to the current speed of the spindle of the four-thread machine. The parameter configuration decomposition module obtains the preset filter bank initialization parameters by looking up the table based on the operating condition parameters, and uses the initialization parameters to perform variational mode decomposition on the high-frequency vibration time domain signal to obtain several finite bandwidth mode components with preset center frequencies. The multi-scale entropy feature extraction module calculates the sample entropy for each finite bandwidth modal component at multiple coarse-grained time scales, obtains the multi-scale entropy value corresponding to each modal component, and constructs an entropy feature vector representing the current running state from all multi-scale entropy values. The support vector classification and discrimination module inputs the entropy feature vector into the classification and discrimination layer based on the support vector structure. The classification and discrimination layer outputs the attribute label of the vibration event according to the preset decision logic. The attribute label is used to indicate whether the current vibration belongs to the working condition disturbance event or the fault symptom event. The closed-loop threshold optimization module associates attribute labels and corresponding entropy feature vectors with the production batch data and operation and maintenance record data of the four-line machine to form a time-series correlated data set, and dynamically corrects the discrimination threshold for subsequent vibration events based on the data set.

Images