Sensor-based real-time detection system for mold wear in glass bottle machines

Abstract

This invention discloses a sensor-based real-time wear detection system for glass bottle-making machine molds, belonging to the field of glass bottle-making machine mold wear detection technology. The system includes a data acquisition module for synchronously acquiring radial displacement time sequence, contact pressure time sequence, surface temperature spatiotemporal distribution, and ultrasonic echo attenuation data during the mold blowing cycle; a wear feature extraction module for performing multi-scale decomposition of the multi-dimensional response signal using dual-tree complex wavelet transform, obtaining the decomposition coefficient matrix at each scale, and calculating the multimodal wear feature vector; and a wear state determination module for inputting the feature vector into a pre-constructed wear state space, determining the current wear level through local outlier density estimation, and outputting the detection data. This invention integrates multi-source sensor information and multi-scale analysis technology to achieve high-precision online detection and level determination of mold wear, improving the real-time performance, accuracy, and robustness of the detection.

Description

Technical Field

[0001] This invention relates to the field of glass bottle making machine mold wear detection technology, specifically a sensor-based real-time glass bottle making machine mold wear detection system. Background Technology

[0002] Under high temperature, high pressure, and frequent opening and closing conditions, the inner surface of glass bottle-making machine molds wears down due to continuous contact with high-temperature glass material, leading to decreased bottle forming accuracy and increased surface defects. Real-time detection of mold wear is essential to ensure the quality of glass bottle production and the continuity of the production line. Current methods for mold wear detection mostly rely on monitoring a single physical quantity, such as inferring the degree of wear solely from changes in mold temperature or strain signals in a single direction. Such single signal sources cannot comprehensively reflect the evolution of multi-physics responses caused by wear, easily missing early wear characteristics, and limiting detection accuracy and sensitivity. At the signal processing level, conventional discrete wavelet transform or energy spectrum analysis methods are used to extract wear features. However, discrete wavelet transform is translationally sensitive; even small signal shifts can cause drastic fluctuations in decomposition coefficients, resulting in insufficient stability of the extracted features and difficulty in reliably capturing multi-scale wear traces under complex operating conditions. At the state discrimination level, most methods employ threshold comparison or supervised learning-based classification models, requiring pre-defined fixed wear level boundaries or relying on a large number of balanced labeled samples for training. However, the wear process of molds is nonlinear and gradual, with significant differences in wear paths between different batches of molds. Abnormal wear patterns are difficult to enumerate in advance, and traditional classification models have weak discrimination capabilities for unknown wear states and boundary region samples, easily leading to misjudgments and omissions. Therefore, it is necessary to solve how to stably extract multi-scale wear features from multi-source sensor information under complex mold operating conditions, and identify continuous changes and abnormal deviations in wear levels in the state space, so as to achieve real-time and accurate determination of wear state. Summary of the Invention

[0003] This invention provides a sensor-based real-time wear detection system for glass bottle making machine molds. It aims to synchronously collect multi-dimensional response signals of the mold during the blowing cycle, including radial displacement, contact pressure, spatiotemporal distribution of surface temperature, and ultrasonic echo attenuation. The system utilizes dual-tree complex wavelet transform to decompose the multi-dimensional signals at multiple scales and extract multi-modal wear feature vectors. These feature vectors are then projected onto a wear state space constructed based on dimensionality reduction of historical samples. Local outlier density estimation is used to determine the current wear level, enabling stable perception and anomaly detection of mold wear under non-stationary operating conditions, providing real-time detection data for mold maintenance.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a sensor-based real-time wear detection system for glass bottle making machine molds. The system includes a data acquisition module, a wear feature extraction module, and a wear state determination module.

[0005] The data acquisition module is used to synchronously acquire multidimensional response signals of the target mold during the blowing cycle using multiple sensors. These multidimensional response signals include the radial displacement time series of the mold's inner wall, the contact pressure time series, the spatiotemporal distribution of surface temperature, and ultrasonic echo attenuation data. As a technical solution of this invention, the data acquisition module uniformly distributes multiple fiber Bragg grating strain sensors along the axial and circumferential directions on the inner wall of the target mold to measure the radial displacement time series. Simultaneously, a thin-film pressure sensor array is embedded on the surface of the mold cavity to acquire the contact pressure time series. An infrared thermal imager is installed on the outer wall of the mold to obtain the spatiotemporal distribution of surface temperature. An ultrasonic transducer is fixed at the bottom of the mold to excite and receive ultrasonic signals directed towards the inner wall of the mold, recording the ultrasonic echo attenuation data. By synchronously triggering the acquisition of strain, pressure, temperature, and ultrasonic signals within a unified blowing cycle, the multi-source heterogeneous signals are strictly aligned in the time domain, providing consistent original observation information for subsequent multidimensional feature fusion and avoiding phase deviations introduced by asynchronous acquisition.

[0006] The wear feature extraction module is used to perform multi-scale decomposition processing on the multidimensional response signal based on dual-tree complex wavelet transform, obtain the decomposition coefficient matrix at each scale, and calculate the multimodal wear feature vector based on the decomposition coefficient matrix. Preferably, independent dual-tree complex wavelet transforms are performed on the radial displacement time series, contact pressure time series, surface temperature spatiotemporal distribution, and ultrasonic echo attenuation data, respectively. After decomposition, the low-frequency approximation coefficients of the real part tree and the imaginary part tree, as well as the high-frequency detail coefficients, are obtained for each signal. The coefficients of the real part tree and the imaginary part tree at corresponding scales are complex-fused to generate the decomposition coefficient matrix at each scale. The dual-tree complex wavelet transform has the advantages of approximate translation invariance and suppression of frequency aliasing, separating multi-scale detail components that are highly correlated with the wear state of the mold from multi-source signals, so that the signal anomalies caused by early weak wear can be clearly expressed at the feature level.

[0007] As a technical solution of this invention, the wear feature extraction module extracts the wavelet coefficient modulus maxima sequence at each scale from the decomposition coefficient matrix corresponding to the radial displacement time series as the first feature vector; extracts the wavelet coefficient energy distribution gradient at each scale from the decomposition coefficient matrix corresponding to the contact pressure time series as the second feature vector; extracts the principal component proportion after singular value decomposition of the wavelet coefficients at each scale from the decomposition coefficient matrix corresponding to the spatiotemporal distribution of surface temperature as the third feature vector; and extracts the wavelet coefficient correlation dimension at each scale from the decomposition coefficient matrix corresponding to the ultrasonic echo attenuation data as the fourth feature vector. The first, second, third, and fourth feature vectors are then concatenated in scale order to obtain a multimodal wear feature vector. Preferably, when extracting the first feature vector, the wavelet coefficient modulus maxima sequence at each scale is normalized to eliminate the influence of differences in sensor sensitivity on the feature amplitude. This process constructs a unified-scale feature representation from four dimensions: mechanical deformation, contact stress, thermal distribution, and acoustic characteristics, enabling different wear modes to exhibit a distinguishable distribution structure in the feature space.

[0008] The wear state determination module inputs multimodal wear feature vectors into a pre-constructed wear state space, determines the current wear level using a density estimation method based on local outlier factor analysis, and outputs real-time wear detection data. The wear state space is constructed by pre-collecting multimodal wear feature vector samples from various wear stages of historical molds, from brand new to failed. Each sample is labeled with a wear level, and an isometric mapping dimensionality reduction algorithm is used to map these multimodal wear feature vector samples into a three-dimensional space. Each wear level corresponds to a dense cluster region in the three-dimensional space.

[0009] The wear state determination module calculates the projection points of the multimodal wear feature vectors in the wear state space. A search radius is set with the projection points as the center. The number of historical sample points with labeled wear levels within this radius is counted, and the local reachability density of each historical sample point is calculated. Then, the local outlier value of the projection point is calculated. When the local outlier value is less than or equal to a first preset threshold, the projection point is assigned to the wear level corresponding to the historical sample point with the largest number of points within the search radius as the current wear level. When the local outlier value is greater than the first preset threshold, the projection point is marked as an anomaly, and the wear level corresponding to the cluster with the smallest Euclidean distance from the projection point to the center of each wear level cluster is selected as the current wear level. Preferably, the first preset threshold is set to 2. When the local outlier value is greater than 2 and the distance between the projection point and the center of all wear level clusters exceeds 1.5 times the radius of the cluster, the current wear state is determined to be an unknown anomaly, and a system alarm signal is triggered. The wear level is determined by local outlier factor analysis, which does not rely on the pre-set assumption of global distribution. It can stably segment the transition samples between adjacent wear stages during the gradual evolution of mold wear, and effectively identify outlier data points caused by short-term sensor failure or process disturbance, thus avoiding the spread of misjudgments.

[0010] When the wear status determination module outputs real-time wear detection data, this data includes the current wear level identifier, projection point coordinates, and local outlier factor values. This data is encapsulated into data frames according to a preset communication protocol and sent to the host computer monitoring system via an industrial Ethernet interface. The host computer not only obtains a qualitative wear level determination but also acquires information on the projection position and outlier degree, facilitating production line operators to trace the wear evolution trajectory and predict the remaining usable life.

[0011] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0012] A dual-tree complex wavelet transform was employed to independently decompose radial displacement time series, contact pressure time series, surface temperature spatiotemporal distribution, and ultrasonic echo attenuation data into multi-scale data. A decomposition coefficient matrix was generated by complex fusion of the scale coefficients of the real and imaginary trees. Modulus maxima sequences, energy distribution gradients, singular value decomposition principal component proportions, and correlation dimensions were extracted from the decomposition coefficients of different signals as feature vectors and concatenated to form a multimodal wear feature vector. The dual-tree complex wavelet transform exhibits near translation invariance and better direction selectivity, suppressing fluctuations in decomposition coefficients caused by small time shifts in the signal. This ensures the stability of extracted wavelet coefficient modulus maxima and other features across multiple scales, overcoming the technical deficiency of conventional discrete wavelet transforms caused by translation sensitivity and feature drift. Furthermore, by fusing wear features from four heterogeneous signals—displacement, pressure, temperature, and ultrasonic echo attenuation—the wear evolution process of the mold can be comprehensively characterized from multiple dimensions, including mechanical response, thermal distribution, and acoustic propagation characteristics. Its ability to characterize weak changes and local anomalies in the early stages of wear is significantly superior to that of single-signal features, facilitating the accurate capture of multi-scale information on wear states under complex process conditions. Multimodal wear feature vectors are input into a 3D wear state space constructed through dimensionality reduction using isometric mapping. Wear levels are then determined by density estimation using local outlier analysis. Historical samples are grouped into dense clusters labeled with wear levels. The local outlier value of the current projection point is calculated. If this value does not exceed a preset threshold, the projection point is assigned to the corresponding wear level based on local density. If the value exceeds the threshold and is far from all cluster centers, it is marked as abnormal or an alarm is triggered. This density-based determination method does not rely on fixed level boundaries and can adaptively handle continuous transitions in wear states and unknown abnormal patterns. It avoids misjudgments caused by fuzzy sample boundaries or rejection of unknown patterns in traditional classification models, thus providing real-time detection results that better reflect actual wear trends. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a schematic diagram of a real-time wear detection system for glass bottle making machine molds;

[0015] Figure 2 This is a schematic diagram of the structure of a multi-physics field sensor deployment and synchronous acquisition system for molds;

[0016] Figure 3 This is a flowchart of wear feature extraction;

[0017] Figure 4This is a flowchart of a wear state determination method based on local outlier detection. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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] See Figure 1 This invention provides a sensor-based real-time wear detection system for glass bottle-making machine molds. The system includes a data acquisition module, a wear feature extraction module, and a wear state determination module. The data acquisition module uses multiple sensors to simultaneously acquire multi-dimensional response signals of the target mold during the blowing cycle. These multi-dimensional response signals include the radial displacement time series of the mold's inner wall, the contact pressure time series, the spatiotemporal distribution of surface temperature, and ultrasonic echo attenuation data. The wear feature extraction module performs multi-scale decomposition processing on the multi-dimensional response signals based on dual-tree complex wavelet transform to obtain decomposition coefficient matrices at each scale, and calculates multi-modal wear feature vectors based on these coefficient matrices. The wear state determination module inputs the multi-modal wear feature vectors into a pre-constructed wear state space, determines the current wear level using a density estimation method based on local outlier analysis, and outputs real-time wear detection data.

[0020] Example 1

[0021] In specific implementation, please refer to Figure 2The data acquisition module uniformly deploys multiple fiber Bragg grating strain sensors along the axial and circumferential directions of the target mold on its inner wall. The fiber Bragg grating strain sensors convert the offset of the reflected wavelength of the internal grating into a radial strain value, and then output the radial displacement time sequence based on a pre-calibrated strain-displacement relationship. A thin-film pressure sensor array is embedded on the cavity surface of the target mold. Each thin-film pressure sensing unit in the array directly contacts the forming glass material, converting the contact surface pressure into a resistance change signal. After amplification and filtering by a signal conditioning circuit, the contact pressure time sequence is acquired in real time. An infrared thermal imager is installed on the outer wall of the target mold. The infrared thermal imager continuously captures image frames of the temperature distribution on the outer surface of the target mold in a non-contact manner. Each frame contains the temperature value of two-dimensional spatial pixels, and the spatiotemporal distribution of the surface temperature is obtained according to a time series. An ultrasonic transducer is fixed at the bottom of the target mold. The ultrasonic transducer emits ultrasonic pulses in a direction perpendicular to the bottom surface of the target mold. The ultrasonic pulses propagate within the target mold material and are reflected after reaching the inner wall of the target mold. The reflected echo is received by the same ultrasonic transducer. The attenuation calculation unit built into the ultrasonic transducer compares the amplitude of the emitted pulse with the amplitude of the received echo, calculates the amplitude attenuation in decibels, and records it as ultrasonic echo attenuation data. Multiple fiber Bragg grating strain sensors, all thin-film pressure sensing units, infrared thermal imagers, and ultrasonic transducers are all connected to a synchronous trigger controller. The synchronous trigger controller receives the blowing cycle start signal output by the bottle-making machine control system. At the start of the blowing cycle, the synchronous trigger controller generates a trigger pulse, driving multiple fiber Bragg grating strain sensors, all thin-film pressure sensing units, infrared thermal imagers, and ultrasonic transducers to synchronously start acquisition, so that the radial displacement time sequence, contact pressure time sequence, surface temperature spatiotemporal distribution, and ultrasonic echo attenuation data remain aligned on the time axis.

[0022] Example 2

[0023] In specific implementation, please refer to Figure 3 The wear feature extraction module receives radial displacement time series, contact pressure time series, surface temperature spatiotemporal distribution, and ultrasonic echo attenuation data synchronously acquired by the data acquisition module. The wear feature extraction module first performs a dual-tree complex wavelet transform on the radial displacement time series. It selects a biorthogonal wavelet filter bank with approximately symmetry and tight support characteristics as the real part tree filter bank, and a filter bank forming a Hilbert transform pair with the real part tree filter bank as the imaginary part tree filter bank. The number of decomposition layers is set to five layers based on the sampling frequency of the radial displacement time series and the signal frequency band range caused by mold wear. The radial displacement time series is decomposed layer by layer by the real part tree filter bank, outputting low-frequency approximation coefficients and high-frequency detail coefficients of the real part tree at each layer; the radial displacement time series is decomposed layer by layer by the imaginary part tree filter bank, outputting low-frequency approximation coefficients and high-frequency detail coefficients of the imaginary tree at each layer. For the first... The decomposition scale will be the first... The corresponding elements of the low-frequency approximation coefficients of the real part tree and the low-frequency approximation coefficients of the imaginary part tree at each decomposition scale are combined to form a complex number, thus forming the _th_ ... The low-frequency approximate complex coefficients at the decomposition scale; the first... The corresponding elements of the high-frequency detail coefficients of the real part tree and the high-frequency detail coefficients of the imaginary part tree at each decomposition scale are combined to form a complex number, thus forming the th... The high-frequency detail complex coefficients at each decomposition scale are used. The low-frequency approximate complex coefficients and high-frequency detail complex coefficients at all decomposition scales are arranged in scale order to form the decomposition coefficient matrix corresponding to the radial displacement time series. The contact pressure time series is subjected to dual-tree complex wavelet transform in the same way. The surface temperature spatiotemporal distribution is subjected to dual-tree complex wavelet transform according to the time series of each spatial pixel, and the transform results are arranged according to spatial position to form the decomposition coefficient matrix of the surface temperature spatiotemporal distribution. The ultrasonic echo attenuation data is also subjected to dual-tree complex wavelet transform to obtain the decomposition coefficient matrix of the ultrasonic echo attenuation data.

[0024] The wear feature extraction module extracts the wavelet coefficient modulus maxima sequence at each scale from the decomposition coefficient matrix corresponding to the radial displacement time series as the first feature vector. When extracting the first feature vector, for the... The decomposition scale searches for local maxima of the modulus of the complex coefficients of high-frequency details along the time axis, and arranges the found local maxima in chronological order to form the first decomposition scale. The wavelet coefficient modulus maxima sequence is decomposed at each scale, and the wavelet coefficient modulus maxima sequence is normalized using the following formula:

[0025]

[0026] in, Indicates the decomposition scale number. The range of values ​​for is natural numbers. to ; Represents the th wavelet coefficient modulus maxima sequence Each modulo maximum point, The range of values ​​for is positive integers; Indicates the first The sequence of wavelet coefficient modulus maxima at the i-th decomposition scale is the i-th The original amplitude at each modulus maximum point; Indicates the first The arithmetic mean of the original amplitudes of all modulus maxima points in the wavelet coefficient modulus maxima sequence at the nth decomposition scale is obtained by applying the arithmetic mean to the nth decomposition scale. The summation of the original magnitudes of all modulus maxima points in the wavelet coefficient modulus maxima sequence at each decomposition scale, divided by the first decomposition scale, is performed by the summation of the original magnitudes of all modulus maxima points in the sequence. The total number of modulus maxima points of the wavelet coefficient modulus maxima sequence at each decomposition scale is obtained; Indicates the first The standard deviation of the original amplitudes of all modulus maxima points in the modulus maxima sequence at the decomposition scale is calculated first. The original amplitude of each modulus maxima point in the wavelet coefficient modulus maxima sequence at each decomposition scale and Sum the squares of the differences, divide the sum by the total number of points with maximum modulus, and finally take the square root. Indicates the first The sequence of wavelet coefficient modulus maxima at the i-th decomposition scale is the i-th The amplitudes of the normalized modulus maxima are obtained by concatenating the normalized modulus maxima sequences of all decomposition scales. Normalization eliminates the influence of sensitivity differences between different fiber Bragg grating strain sensors on the characteristic amplitudes.

[0027] The wear feature extraction module extracts the wavelet coefficient energy distribution gradient at each scale from the decomposition coefficient matrix corresponding to the contact pressure time series as the second feature vector. For the first... The decomposition scale is used to calculate the sum of the energy values ​​of the high-frequency detail complex coefficients at all time points, resulting in the th ... The wavelet coefficient energy of each decomposition scale is calculated, and the first-order difference of the energy change between adjacent decomposition scales is calculated for the wavelet coefficient energy vector formed by the wavelet coefficient energy of all decomposition scales. The first-order difference sequence is used as the wavelet coefficient energy distribution gradient of the contact pressure time series to form the second feature vector.

[0028] The wear feature extraction module extracts the principal component proportions after singular value decomposition of wavelet coefficients at each scale from the decomposition coefficient matrix corresponding to the spatiotemporal distribution of surface temperature, using these proportions as the third feature vector. For each decomposition scale, the high-frequency detail complex coefficients of that scale in the decomposition coefficient matrix corresponding to the spatiotemporal distribution of surface temperature are rearranged into a two-dimensional complex matrix according to spatial pixels. Singular value decomposition is performed on the two-dimensional complex matrix to obtain a sequence of singular values ​​arranged in descending order. The ratio of the largest singular value to the sum of all singular values ​​is calculated as the principal component proportion. The principal component proportions at all decomposition scales are arranged in order to obtain the third feature vector.

[0029] The wear feature extraction module extracts the correlation dimension of wavelet coefficients at each scale from the decomposition coefficient matrix corresponding to the ultrasonic echo attenuation data as the fourth feature vector. For each decomposition scale, based on the high-frequency detail complex coefficient sequence, the phase space is reconstructed from the high-frequency detail complex coefficient sequence. The number of point pairs with a phase point distance less than a preset neighborhood radius under different embedding dimensions is calculated. The correlation dimension is obtained by fitting the double logarithmic relationship between the number of point pairs and the neighborhood radius. The correlation dimensions at all decomposition scales are arranged in order to obtain the fourth feature vector.

[0030] The wear feature extraction module concatenates the first, second, third, and fourth feature vectors in the natural order of the decomposition scale to obtain a multimodal wear feature vector.

[0031] Example 3

[0032] In practical implementation, the wear state determination module pre-collects multimodal wear feature vector samples from various wear stages of historical molds, from brand new to failed states. These wear stages are divided into initial break-in wear, stable wear, intermediate accelerated wear, and final severe wear stages based on the actual operating time and wear amount of the mold's inner wall. At each wear stage, the data acquisition module and wear feature extraction module acquire the corresponding multimodal wear feature vector according to a preset acquisition cycle, serving as the multimodal wear feature vector sample for that stage. Each multimodal wear feature vector sample is labeled with a corresponding wear level label, using integer encoding: 1 for the initial break-in wear stage, 2 for the stable wear stage, 3 for the intermediate accelerated wear stage, and 4 for the final severe wear stage.

[0033] After acquiring all multimodal wear feature vector samples and their corresponding wear level labels, the wear state determination module uses an isometric mapping dimensionality reduction algorithm to map all multimodal wear feature vector samples to a three-dimensional space. The isometric mapping dimensionality reduction algorithm constructs a nearest neighbor graph based on Euclidean distance between multimodal wear feature vector samples. For any two multimodal wear feature vector samples, the Euclidean distance between the two multimodal wear feature vector samples in the high-dimensional original space is calculated. When the Euclidean distance between the two multimodal wear feature vector samples is less than a preset nearest neighbor threshold, the nodes corresponding to the two multimodal wear feature vector samples are connected in the nearest neighbor graph, and the weight of the connecting edge is assigned to the Euclidean distance between the two multimodal wear feature vector samples. When the Euclidean distance between the two multimodal wear feature vector samples is greater than or equal to the preset nearest neighbor threshold, the nodes corresponding to the two multimodal wear feature vector samples are not directly connected in the nearest neighbor graph. After the nearest neighbor graph is constructed, the isometric mapping dimensionality reduction algorithm calculates the shortest path distance between the nodes corresponding to any two multimodal wear feature vector samples in the nearest neighbor graph, which is used as the geodesic distance between the two multimodal wear feature vector samples. The preset nearest neighbor threshold is set to the eighth percentile of the Euclidean distance between all pairs of multimodal wear feature vector samples. The basis for setting the preset nearest neighbor threshold to a certain percentile is to avoid bridging samples in different wear level clusters while ensuring the connectivity of the nearest neighbor graph. The eighth percentile is determined by connectivity testing when the total number of multimodal wear feature vector samples is three hundred.

[0034] The wear state determination module constructs a geodesic distance matrix from the geodesic distances between all multimodal wear feature vector samples. A double centering operation is then performed on this geodesic distance matrix to obtain a centered inner product matrix. The double centering operation is as follows: The square of each element in the geodesic distance matrix is ​​calculated to obtain a squared geodesic distance matrix. The average values ​​of each row and column of the squared geodesic distance matrix, as well as the overall average value of all elements, are calculated. Then, each element of the squared geodesic distance matrix is ​​subtracted from its corresponding row average, then its corresponding column average, and finally the overall average value is added. This result is multiplied by -1 / 2 to obtain the centered inner product matrix. Eigenvalue decomposition is performed on the centered inner product matrix, and the three largest eigenvalues ​​and their corresponding eigenvectors are selected. Let the three largest eigenvalues ​​be... , and The eigenvectors corresponding to the three largest eigenvalues ​​are respectively , and Then the first The mapping coordinates of the multimodal wear feature vector samples in three-dimensional space are:

[0035]

[0036] in, This represents the index number of the multimodal wear feature vector sample. The value range is a positive integer from 1 to the total number of samples in the multimodal wear feature vector; Indicates the first The mapping coordinates of a multimodal wear feature vector sample in three-dimensional space, the mapping coordinates being a three-dimensional vector containing three components; This represents the largest eigenvalue of the centered inner product matrix. It is obtained by performing eigenvalue decomposition on the centered inner product matrix and then taking the maximum eigenvalue; This represents the second largest eigenvalue of the centered inner product matrix. It is obtained by performing eigenvalue decomposition on the centered inner product matrix and then taking the second largest eigenvalue. This represents the third largest eigenvalue of the centered inner product matrix. It is obtained by performing eigenvalue decomposition on the centered inner product matrix and taking the third largest eigenvalue. Representing the eigenvector The Middle The value of each element, the feature vector for The corresponding feature vector; Representing the eigenvector The Middle The value of each element, the feature vector for The corresponding feature vector; Representing the eigenvector The Middle The value of each element, the feature vector for The corresponding feature vector.

[0037] After all multimodal wear feature vector samples are mapped, the three-dimensional space formed by all mapped coordinates constitutes the wear state space. In the wear state space, mapping points of multimodal wear feature vector samples with the same wear level label cluster to form dense cluster regions. The set of mapping points corresponding to wear level label 1 forms the first dense cluster region, the set of mapping points corresponding to wear level label 2 forms the second dense cluster region, the set of mapping points corresponding to wear level label 3 forms the third dense cluster region, and the set of mapping points corresponding to wear level label 4 forms the fourth dense cluster region. The wear state determination module stores the wear state space and the wear level label corresponding to each mapping point in the wear state space in a memory. When the wear feature extraction module generates a new multimodal wear feature vector, the wear state determination module inputs the new multimodal wear feature vector into the wear state space. Following the nearest neighbor graph structure and geodesic distance calculation method trained in the metric mapping dimensionality reduction algorithm, the module calculates the coordinates of the projection points of the new multimodal wear feature vector in the wear state space for subsequent wear level determination.

[0038] Example 4

[0039] In specific implementation, please refer to Figure 4 After the wear state determination module obtains the current multimodal wear feature vector output by the wear feature extraction module, it calls the pre-stored wear state space and the coordinates of historical sample points labeled with wear levels in the wear state space to calculate the projection point of the current multimodal wear feature vector in the wear state space. The calculation method of the projection point is as follows: determine the Euclidean distance between the current multimodal wear feature vector and all historical multimodal wear feature vector samples in the high-dimensional original space, select the three historical multimodal wear feature vector samples with the smallest Euclidean distance as the nearest neighbor reference points, and calculate the coordinates of the projection point of the current multimodal wear feature vector in the wear state space by interpolation using the centroid coordinate method based on the known coordinates of the three nearest neighbor reference points in the wear state space.

[0040] The wear state determination module sets the search radius with the projection point as the center. The search radius is set as follows: The set of distances between all historical sample points in the wear state space and their respective k-th nearest neighbor sample points is calculated. The median of this distance set is then multiplied by a preset radius coefficient to obtain the search radius. Here, k is set to 10, and the preset radius coefficient is set to 1.5. The reason for setting k to 10 is that, with 300 historical sample points stored in the wear state space, the 10 nearest neighbors can cover a sufficient local area without crossing the boundaries of different wear level clusters. The reason for setting the preset radius coefficient to 1.5 is that, in the wear state space, the average intra-cluster point spacing of the four dense cluster regions is approximately 1.2 times the median distance; a value of 1.5 ensures that cluster edge sample points are included while avoiding crosstalk between clusters of different wear levels.

[0041] The wear state determination module counts the number of historical sample points within a search radius centered on the projection point, and calculates the local reachability density of each historical sample point within the search radius. For the β-th historical sample point within the search radius, the local reachability density calculation process is as follows: all sample points within the search radius centered on the β-th historical sample point constitute the neighborhood point set of the β-th historical sample point; for each neighbor point in the neighborhood point set of the β-th historical sample point, calculate the reachability distance from the β-th historical sample point to that neighbor point, where the reachability distance is the larger of the Euclidean distance from the β-th historical sample point to that neighbor point and the k-th nearest neighbor distance of the β-th historical sample point; take the reciprocal of the arithmetic mean of all reachability distances as the local reachability density of the β-th historical sample point.

[0042] The wear state determination module calculates the local outlier factor value of the projection point based on the local reachability density of historical sample points in the neighborhood point set. The formula for calculating the local outlier factor value of the projection point is:

[0043]

[0044] in, This represents the local outlier factor value of the projection point. The range of values ​​for is real numbers greater than 0; This represents the neighborhood of the projected point, which is composed of all historical sample points within the search radius centered at the projected point. This represents the total number of historical sample points in the neighborhood of the projected point. It is a positive integer; Represents the first point in the neighborhood set of the projection point. One historical sample point, exist Iterate through the data to retrieve values; Indicates the first Local reachability density of historical sample points; The local reachability density of the projection point is obtained by imagining the projection point as a neighborhood point and concentrating the geometric centroid of the three nearest historical sample points of the projection point, and then calculating the reciprocal of the reachability distance of the imaginary location.

[0045] After obtaining the local outlier factor value of the projection point, the wear state determination module compares the local outlier factor value with a first preset threshold. The first preset threshold is set to 2. The basis for setting the first preset threshold to 2 is that when the local outlier factor value is equal to 2, it means that the local density of the projection point is about half of the average local density of the neighboring points. In the gradual wear process of the mold, the half density difference corresponds to a significant offset boundary of the wear state.

[0046] When the local outlier factor value is less than or equal to a first preset threshold, the wear state determination module counts the number of historical sample points corresponding to each wear level in the neighborhood of the projected point, and selects the wear level with the most historical sample points in the neighborhood as the current wear level. When the local outlier factor value is greater than the first preset threshold, the wear state determination module marks the projected point as an anomaly and calculates the Euclidean distance between the projected point and the center of each wear level cluster in the wear state space. The center of the wear level cluster is a three-dimensional coordinate point formed by the arithmetic mean of the coordinates of all historical sample points corresponding to the wear level. The wear state determination module compares the Euclidean distances between the projected point and the center of each wear level cluster, and selects the wear level corresponding to the wear level cluster with the smallest Euclidean distance as the current wear level.

[0047] During the anomaly detection process, the wear status determination module further checks whether the Euclidean distance between the projected point and the center of all wear level clusters exceeds 1.5 times the radius of the corresponding wear level cluster. The wear level cluster radius is defined as the maximum Euclidean distance from all historical sample points of the corresponding wear level to the center of that wear level cluster. When the local outlier value is greater than 2 and the Euclidean distance between the projected point and the center of all wear level clusters exceeds 1.5 times the radius of the corresponding wear level cluster, the wear status determination module determines the current wear status as an unknown anomaly and triggers a system alarm signal. The system alarm signal is output to the host computer monitoring system via the industrial Ethernet interface, and includes the projected point coordinates, local outlier value, and anomaly identifier code in the alarm signal.

[0048] Example 5

[0049] In practical implementation, after the wear state determination module determines the current wear level, it generates real-time wear detection data. This real-time wear detection data includes three data fields: the first field is the current wear level identifier, stored as an unsigned integer. Identifier value 1 represents the initial break-in wear stage, identifier value 2 represents the stable wear stage, identifier value 3 represents the mid-term accelerated wear stage, identifier value 4 represents the final severe wear stage, and identifier value 5 represents an unknown abnormal state; the second field is the projection point coordinate value, which is composed of the three coordinate components of the projection point in the wear state space arranged sequentially, with each coordinate component stored as a single-precision floating-point number; the third field is the local outlier factor value, also stored as a single-precision floating-point number.

[0050] The wear status determination module encapsulates real-time wear detection data into data frames according to a preset communication protocol. The preset communication protocol defines a data frame as consisting of a frame header field, a length field, a command code field, a data payload field, and a frame check field, in that order. The frame header field is fixed as two consecutive hexadecimal values, 0xAA and 0x55. The length field occupies two bytes, and its value equals the total number of bytes in the data payload field plus one byte in the command code field. The command code field occupies one byte, and its value is set to 0x10, indicating that the data frame type is a mold wear detection data reporting frame. The data payload field stores the encoded byte sequence of the real-time wear detection data. The encoding rule for this byte sequence is as follows: first, the current wear level identifier is stored, occupying one byte; then, the first coordinate component of the projected point coordinate value is stored, occupying four bytes; then, the second coordinate component of the projected point coordinate value is stored, occupying four bytes; then, the third coordinate component of the projected point coordinate value is stored, occupying four bytes; finally, the local outlier factor value is stored, occupying four bytes. The total length of the data payload field is seventeen bytes. The frame check field occupies two bytes. The value of the frame check field is the cyclic redundancy check value of all bytes from the first byte of the frame header field to the last byte of the data payload field. The cyclic redundancy check uses the CRC-16 standard polynomial. Perform calculations, where Represents a polynomial variable.

[0051] The wear condition determination module sends the encapsulated data frame to the host computer monitoring system via an industrial Ethernet interface. The industrial Ethernet interface uses an RJ45 physical interface, with the physical layer conforming to the 100BASE-TX standard. The network layer uses Internet Protocol version 4 (IP4), the transport layer uses User Datagram Protocol (UDP), and the application layer uses the frame encapsulation format defined in the preset communication protocol. The wear condition determination module treats the data frame as the data area of ​​a UDP datagram, adding the UDP header, Internet Protocol header, and Ethernet frame encapsulation layer by layer. Then, it drives the RJ45 physical interface through the Media Access Control (MAC) layer controller to send electrical signals to the designated Internet Protocol address and port of the host computer monitoring system. The wear condition determination module performs a data frame encapsulation and transmission operation once after each wear condition determination cycle, and the wear condition determination cycle is synchronized with the glass bottle blowing cycle.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A sensor-based real-time wear detection system for glass bottle making machine molds, characterized in that, The system includes: The data acquisition module is used to synchronously acquire multidimensional response signals of the target mold during the blowing cycle using multiple sensors. The multidimensional response signals include the radial displacement time sequence of the inner wall of the mold, the contact pressure time sequence, the spatiotemporal distribution of surface temperature, and ultrasonic echo attenuation data. The wear feature extraction module is used to perform multi-scale decomposition processing on the multidimensional response signal based on dual-tree complex wavelet transform to obtain the decomposition coefficient matrix at each scale, and calculate the multimodal wear feature vector based on the decomposition coefficient matrix. The wear state determination module is used to input the multimodal wear feature vector into a pre-constructed wear state space, determine the current wear level through a density estimation method based on local outlier factor analysis, and output real-time wear detection data.

2. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 1, characterized in that, The data acquisition module includes: Multiple fiber Bragg grating strain sensors are uniformly arranged along the axial and circumferential directions on the inner wall of the target mold for measuring the radial displacement timing. An array of thin-film pressure sensors embedded in the surface of the mold cavity is used to collect contact pressure timing data. An infrared thermal imager is installed on the outer wall of the mold to obtain the spatiotemporal distribution of surface temperature; An ultrasonic transducer fixed at the bottom of the mold is used to excite and receive ultrasonic signals directed at the inner wall of the mold, and to record the ultrasonic echo attenuation data.

3. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 1, characterized in that, The wear feature extraction module performs multi-scale decomposition processing on the multidimensional response signal based on dual-tree complex wavelet transform, including: Independent dual-tree complex wavelet transforms are performed on the radial displacement time series, contact pressure time series, surface temperature spatiotemporal distribution, and ultrasonic echo attenuation data. After decomposition, the low-frequency approximation coefficients and high-frequency detail coefficients of the real and imaginary trees are obtained. The coefficients of the real and imaginary trees at corresponding scales are complexly fused to generate the decomposition coefficient matrix at each scale.

4. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 3, characterized in that, The wear feature extraction module calculates a multimodal wear feature vector based on the decomposition coefficient matrix, including: Extract the wavelet coefficient modulus maxima sequence at each scale from the decomposition coefficient matrix corresponding to the radial displacement time series as the first feature vector; The wavelet coefficient energy distribution gradient at each scale is extracted from the decomposition coefficient matrix corresponding to the contact pressure time series as the second feature vector; The principal component proportions of the wavelet coefficient singular value decomposition at each scale are extracted from the decomposition coefficient matrix corresponding to the spatiotemporal distribution of the surface temperature and used as the third feature vector. The wavelet coefficient correlation dimension at each scale is extracted from the decomposition coefficient matrix corresponding to the ultrasonic echo attenuation data as the fourth feature vector. The first feature vector, the second feature vector, the third feature vector, and the fourth feature vector are concatenated in scale order to obtain the multimodal wear feature vector.

5. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 1, characterized in that, The wear state determination module inputs the multimodal wear feature vector into a pre-constructed wear state space, including: Multimodal wear feature vector samples from historical molds at multiple wear stages from brand new to failure are collected in advance. Each sample is labeled with a wear level label, and the multimodal wear feature vector samples are mapped to a three-dimensional space using an equal metric mapping dimensionality reduction algorithm. The three-dimensional space is the wear state space, where each wear level corresponds to a dense cluster region.

6. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 5, characterized in that, The wear condition determination module determines the current wear level using a density estimation method based on local outlier factor analysis, including: Calculate the projection point of the multimodal wear feature vector in the wear state space, set the search radius with the projection point as the center, count the number of historical sample points with labeled wear levels within the radius, calculate the local reachability density of each historical sample point, and then calculate the local outlier value of the projection point.

7. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 6, characterized in that, The wear status determination module determines the current wear level based on the local outlier factor value, including: When the local outlier factor value is less than or equal to the first preset threshold, the wear level corresponding to the most numerous historical sample point within the search radius is assigned to the projection point as the current wear level. When the local outlier factor value is greater than the first preset threshold, the projection point is marked as an anomaly point, and the wear level corresponding to the cluster with the smallest distance from the projection point to the center of each wear level cluster is selected as the current wear level.

8. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 1, characterized in that, The wear condition determination module outputs real-time wear detection data, including: The real-time wear detection data includes the current wear level identifier, projection point coordinates, and local outlier factor values, and is encapsulated into data frames according to a preset communication protocol and sent to the host computer monitoring system through an industrial Ethernet interface.

9. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 4, characterized in that, When extracting the first feature sub-vector, the wavelet coefficient modulus maxima sequence at each scale is normalized to eliminate the influence of differences in sensor sensitivity on the feature amplitude.

10. The sensor-based real-time wear detection system for glass bottle making machine molds according to claim 7, characterized in that, The first preset threshold is set to 2. When the local outlier factor value is greater than 2 and the distance between the projection point and the center of all wear level clusters exceeds 1.5 times the radius of the cluster, the current wear state is determined to be an unknown abnormality, and a system alarm signal is triggered.

Images