Vacuum press failure early warning method and system

Abstract

The application provides a kind of vacuum press machine fault early warning method and system, obtain the multichannel sensor signal in the running cycle of vacuum press machine, constitute the signal matrix to be analyzed;Build by the first sub-dictionary of representing normal operating state and the second sub-dictionary of representing specific fault mode are spliced into the composite redundant dictionary;Based on the composite redundant dictionary, the sparse decomposition of the signal matrix to be analyzed is carried out by solving the regularization optimization problem, and the sparse coefficient matrix is obtained;The energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix is calculated at each time, and the energy is exponentially weighted moving average, to obtain the time-varying fault cumulative index;When the current value of the fault cumulative index exceeds the preset threshold, and the change rate in the preset time window exceeds the preset change rate threshold, generate fault early warning.

Description

Technical Field

[0001] The present application relates to the field of fault warning, and in particular to a vacuum pressing machine fault warning method and system. Background Art

[0002] Vacuum laminating machines are core equipment in the manufacturing of multilayer printed circuit boards (PCBs) and copper-clad laminates. Their primary function is to laminate the inner core board with prepreg (PP sheet), copper foil, and other materials through heating, pressurization, and vacuuming, creating a multilayer PCB with a specific interlayer structure and electrical properties. Due to the complex operating environment, involving multiple physical processes such as high temperature, high pressure, and high vacuum, and long operating cycles, key components such as the heating, vacuum, and hydraulic systems are prone to performance degradation or sudden failure. These failures, such as vacuum leaks, unstable pressure, and uneven heating, not only directly lead to quality defects such as delamination, blistering, and board warping, resulting in the scrapping of production batches, but also cause production line shutdowns, increasing unplanned downtime and repair costs, and severely impacting production efficiency and profitability. Existing fault diagnosis methods primarily include signal processing-based methods, physical model-based methods, and data-driven methods. Signal processing-based methods such as Fourier transform and wavelet analysis struggle to effectively extract subtle early-stage fault signatures when dealing with the nonstationary and nonlinear signals generated by complex systems like vacuum laminating machines. Physical model-based approaches require building a precise mathematical model of the equipment, but this is extremely difficult for vacuum presses with complex structures and coupled multi-physics fields. Furthermore, the model struggles to cover all unknown failure modes. Data-driven approaches, such as deep learning, while highly effective at feature learning, typically require a large and balanced set of fault samples for training. In real industrial scenarios, fault data, especially early-stage fault data, is often scarce, resulting in insufficient model generalization. Furthermore, the black-box nature of these methods limits the interpretability of diagnostic results. Sparse representation theory posits that a signal can be linearly represented by a small number of atoms in an overcomplete dictionary. By constructing a suitable dictionary and performing sparse decomposition on the signal, the main components, fault signatures, and noise can be effectively separated. Dictionary construction is a key challenge. A single predefined dictionary has poor adaptability, while a dictionary learned solely from unlabeled data cannot distinguish between normal and faulty states. When early fault signatures are weak, the sparse decomposition process is easily affected by strong background noise and fluctuations in operating conditions, resulting in the sparse coefficients associated with the fault being overwhelmed, making accurate identification difficult. Therefore, how to construct a composite dictionary that can simultaneously characterize the normal state and specific fault modes, and construct a sparse decomposition algorithm that preferentially extracts weak fault information is a technical problem that needs to be solved urgently to achieve the accuracy of vacuum pressing machine fault warning. Summary of the Invention

[0003] To address the problem that when the characteristics of a vacuum laminating machine are weak in the early stages of a fault, the sparse decomposition process is easily interfered with by strong background noise and operating condition fluctuations, resulting in the fault-related sparse coefficients being overwhelmed and difficult to accurately identify. The present invention proposes a vacuum laminating machine fault early warning method, comprising the following steps:

[0004] Multi-channel sensor signals are acquired during the operating cycle of a vacuum laminating machine to form a signal matrix to be analyzed. A composite redundant dictionary is constructed by combining a first sub-dictionary representing normal operating conditions and a second sub-dictionary representing specific failure modes. The first sub-dictionary is generated based on a physical model of the vacuum laminating machine, and the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data. The learning method applies structured correlation constraints to atoms belonging to the same fault type but originating from different sensor channels.

[0005] Based on the composite redundant dictionary, the signal matrix to be analyzed is sparsely decomposed by solving a regularized optimization problem to obtain a sparse coefficient matrix; the regularized optimization problem minimizes the reconstruction error while applying a penalty term to preferentially select sparse coefficients associated with fault atoms in the second sub-dictionary and maintain the local neighborhood geometric structure of the signal to be analyzed in the sparse coefficient space;

[0006] The energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment is calculated, and the energy is subjected to an exponentially weighted moving average to obtain a time-series fault accumulation index; when the current value of the fault accumulation index exceeds a preset threshold and the rate of change within a preset time window exceeds a preset change rate threshold, a fault warning is generated.

[0007] Optionally, obtaining multi-channel sensor signals during the operation cycle of the vacuum laminating machine includes:

[0008] Synchronously collecting multiple signals representing key physical states of the vacuum laminating machine during its working cycle, including but not limited to pressure, temperature, and mechanical vibration signals;

[0009] The timing signals of each channel are aligned according to the timestamps and constructed into a signal matrix to be analyzed with sensor channels as rows and sampling time points as columns.

[0010] Optionally, the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data, wherein the learning imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensing channels, including:

[0011] Classify historical failure data according to known failure modes;

[0012] For each type of fault data, a joint dictionary learning algorithm is used to impose shared sparsity constraints so that the atoms representing the same fault mode on different sensor channels are structurally associated, thereby obtaining the cross-channel physical characteristics of the specific fault.

[0013] Optionally, the objective function of the regularized optimization problem is composed of the weighted sum of the following three parts:

[0014] (1) Reconstruction error term, which is the Frobenius norm of the difference between the signal matrix to be analyzed and the sparse reconstruction of the signal matrix to be analyzed;

[0015] (2) A weighted L1-norm penalty term, which prioritizes fault-related sparse representations in the solution by imposing a smaller penalty weight on the sparse coefficients corresponding to the fault atoms in the second sub-dictionary than on the corresponding coefficients in the first sub-dictionary;

[0016] (3) Manifold regularization term, which is used to penalize sparse coefficient solutions that are inconsistent with the local geometric structure of the data points in the signal matrix to be analyzed.

[0017] Optionally, the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment is calculated, and an exponentially weighted moving average is performed on the energy to obtain a time-series fault accumulation index, specifically:

[0018] At each sampling time t, the sparse coefficient vector corresponding to the second sub-dictionary is extracted, and the square of the L2 norm of the sparse coefficient vector is calculated as the instantaneous fault energy ;

[0019] Update the fault accumulation index using the exponentially weighted moving average method , the update formula is: ,in is the cumulative fault index of the previous moment, is the smoothing factor.

[0020] Optionally, when the current value of the fault accumulation indicator exceeds a preset threshold and the rate of change within a preset time window exceeds a preset rate of change threshold, a fault warning is generated, specifically:

[0021] a) the current value of the fault accumulation index is greater than a preset threshold value, the preset threshold value being obtained based on a statistical analysis of the growth trend of the fault accumulation index at the early stage of historical fault occurrence;

[0022] b) The rate of change of the fault accumulation indicator within a preset time window is greater than a preset rate of change threshold, and the preset rate of change threshold is obtained by calculating the growth trend of the indicator sequence within the time window.

[0023] The present invention also proposes a vacuum pressing machine fault warning system, comprising the following modules:

[0024] A dictionary construction module is configured to acquire multi-channel sensor signals during the vacuum laminating machine's operating cycle to form a signal matrix to be analyzed. A composite redundant dictionary is constructed by combining a first sub-dictionary representing normal operating conditions and a second sub-dictionary representing specific failure modes. The first sub-dictionary is generated based on a physical model of the vacuum laminating machine, while the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source failure data, imposing structured correlation constraints on atoms belonging to the same failure type but originating from different sensor channels.

[0025] a sparsification module configured to perform a sparse decomposition of the signal matrix to be analyzed by solving a regularized optimization problem based on the composite redundant dictionary to obtain a sparse coefficient matrix; the regularized optimization problem minimizes the reconstruction error while applying a penalty term to preferentially select sparse coefficients associated with fault atoms in the second sub-dictionary and maintain a local neighborhood geometric structure of the signal to be analyzed in the sparse coefficient space;

[0026] An early warning module is used to calculate the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment, and perform an exponentially weighted moving average on the energy to obtain a time-series fault accumulation index; when the current value of the fault accumulation index exceeds a preset threshold and the rate of change within a preset time window exceeds a preset change rate threshold, a fault early warning is generated.

[0027] Optionally, obtaining multi-channel sensor signals during the operation cycle of the vacuum laminating machine includes:

[0028] Synchronously collecting multiple signals representing key physical states of the vacuum laminating machine during its working cycle, including but not limited to pressure, temperature, and mechanical vibration signals;

[0029] The timing signals of each channel are aligned according to the timestamps and constructed into a signal matrix to be analyzed with sensor channels as rows and sampling time points as columns.

[0030] Optionally, the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data, wherein the learning imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensing channels, including:

[0031] Classify historical failure data according to known failure modes;

[0032] For each type of fault data, a joint dictionary learning algorithm is used to impose shared sparsity constraints so that the atoms representing the same fault mode on different sensor channels are structurally associated, thereby obtaining the cross-channel physical characteristics of the specific fault.

[0033] Optionally, the objective function of the regularized optimization problem is composed of the weighted sum of the following three parts:

[0034] (1) Reconstruction error term, which is the Frobenius norm of the difference between the signal matrix to be analyzed and the sparse reconstruction of the signal matrix to be analyzed;

[0035] (2) A weighted L1-norm penalty term, which prioritizes fault-related sparse representations in the solution by imposing a smaller penalty weight on the sparse coefficients corresponding to the fault atoms in the second sub-dictionary than on the corresponding coefficients in the first sub-dictionary;

[0036] (3) Manifold regularization term, which is used to penalize sparse coefficient solutions that are inconsistent with the local geometric structure of the data points in the signal matrix to be analyzed.

[0037] Optionally, the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment is calculated, and an exponentially weighted moving average is performed on the energy to obtain a time-series fault accumulation index, specifically:

[0038] At each sampling time t, the sparse coefficient vector corresponding to the second sub-dictionary is extracted, and the square of the L2 norm of the sparse coefficient vector is calculated as the instantaneous fault energy ;

[0039] Update the fault accumulation index using the exponentially weighted moving average method , the update formula is: ,in is the cumulative fault index of the previous moment, is the smoothing factor.

[0040] Optionally, when the current value of the fault accumulation indicator exceeds a preset threshold and the rate of change within a preset time window exceeds a preset rate of change threshold, a fault warning is generated, specifically:

[0041] a) the current value of the fault accumulation index is greater than a preset threshold value, the preset threshold value being obtained based on a statistical analysis of the growth trend of the fault accumulation index at the early stage of historical fault occurrence;

[0042] b) The rate of change of the fault accumulation indicator within a preset time window is greater than a preset rate of change threshold, and the preset rate of change threshold is obtained by calculating the growth trend of the indicator sequence within the time window.

[0043] The present invention combines the normal state representation based on the physical model with the fault mode representation based on multi-source data learning by constructing a composite redundant dictionary, so that the signal decomposition has a clear physical meaning. During the sparse decomposition process, the regularized optimization problem can preferentially amplify and extract weak fault information while maintaining the intrinsic structure of the signal, solving the problem that early fault characteristics are difficult to detect in a strong noise background and improving the sensitivity of the early warning. In addition, by extracting the coefficients corresponding to specific fault dictionary atoms to calculate the fault index, and combining the dual thresholds of the index amplitude and the rate of change for judgment, it can effectively capture the initiation and development trend of the fault, reducing the false alarms and missed reports that may be caused by the traditional single threshold method, thereby improving the reliability of the vacuum press fault warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the time series of collected data;

[0045] Figure 2 Schematic diagram of the relationship between the signal matrix, composite redundant dictionary and sparse coefficient matrix;

[0046] Figure 3 Schematic diagram of the fault accumulation index. DETAILED DESCRIPTION

[0047] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0048] The terms used in the implementation section of this application are only used to explain the specific embodiments of this application and are not intended to limit this application.

[0049] Specific embodiment 1 proposes a vacuum laminating machine fault warning method, comprising the following steps:

[0050] Step 1: Acquire multi-channel sensor signals during the vacuum laminating machine's operating cycle to form a signal matrix to be analyzed; construct a composite redundant dictionary consisting of a first sub-dictionary representing normal operating conditions and a second sub-dictionary representing specific failure modes; wherein the first sub-dictionary is generated based on a physical model of the vacuum laminating machine; and the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source failure data, which imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensor channels.

[0051] A vacuum press is a device that presses materials together in a vacuum environment. The main operating processes include vacuuming, heating, pressurizing, curing / cooling, etc. Failures may occur in each process. Pressure sensors, temperature sensors, and vibration acceleration sensors are installed in key parts of the vacuum press, such as the hydraulic system, heating system, and frame. Data acquisition cards are used to synchronously collect data on a complete working cycle, such as the entire process from mold closing, vacuuming, pressurizing, heating and maintaining pressure, to cooling and releasing pressure. Figure 1 As shown in the figure, the generated multi-channel time series signal arranges the signal samples of length N from M sensor channels into an M-row and N-column signal matrix to be analyzed. The signal matrix to be analyzed can be approximately represented as the product of the composite redundant dictionary and the sparse coefficient matrix, as shown in the figure. Figure 2 As shown, the sparse coefficient matrix is ​​generally simpler than the signal matrix to be analyzed.

[0052] In one embodiment, the first sub-dictionary is generated from a simplified differential equation model of the vacuum laminating machine's hydraulic or thermal system. The step response or impulse response curves of the simulated model under typical inputs are used as atoms to characterize the normal operating mode. More specifically, a mathematical model of the key physical processes of the vacuum laminating machine under normal operating conditions is established, including: 1) a heat conduction model, which describes how the temperature of the heating plate changes over time based on the set heating power, typically represented by a first- or second-order differential equation; 2) a gas state model, which describes how the pressure within the vacuum chamber changes with the vacuum pump's pumping rate and the set target vacuum level, and can be established based on the ideal gas state equation and flow equation; and 3) a mechanical motion model, which describes the displacement, velocity, and pressure changes during the press's closing and opening processes. Model parameters are then set for a series of typical normal process recipes for the vacuum laminating machine (e.g., different target temperatures, pressures, and dwell times). These differential equations are numerically solved, for example using MATLAB's ODE45 solver, to generate a series of noise-free time series signals, including temperature, pressure, and vibration. The simulated ideal signal segments representing different normal working conditions, after segmentation and normalization, together constitute the atomic set of the first sub-dictionary.

[0053] The second sub-dictionary is learned using an improved group sparse K-SVD algorithm. The pressure and vibration signal segments corresponding to the same confirmed fault in historical data, such as a hydraulic oil leak, are grouped together as a data set. During the iterative dictionary learning process, this data set is forced to share the same sparsity pattern, thereby learning a set of fault atoms with intrinsic physical correlations. One atom represents the impulse response of the fault in the pressure signal, while another represents the high-frequency oscillation in the vibration signal. More specifically, historical fault data is collected and organized, and classified according to known fault labels, such as "vacuum pump performance degradation," "partial heater failure," and "minor seal leakage." For example, for the case of "vacuum pump performance degradation," all multi-channel synchronous signal segments, such as pressure, pump current, and vibration, are extracted when this fault occurs. Joint sparse dictionary learning is then applied to learn a dictionary for this fault type, ensuring that any fault signal segment can be sparsely represented by the dictionary, and that its sparse coefficients across different sensor channels share the same sparsity pattern, meaning that the positions of non-zero elements are the same. Specifically, the multi-channel signal matrices are stacked side by side and then an optimization problem with shared sparsity constraints is solved, for example using functions from the SPAMS library. The learned dictionary atoms inherently capture the common characteristics of a single physical fault, such as "vacuum pump performance degradation," across multiple dimensions such as pressure, current, and vibration. This process is repeated for each fault type, concatenating all the fault-specific sub-dictionaries to generate a second sub-dictionary.

[0054] Step 2: Based on the composite redundant dictionary, the signal matrix to be analyzed is sparsely decomposed by solving a regularized optimization problem to obtain a sparse coefficient matrix; the regularized optimization problem minimizes the reconstruction error while applying a penalty term to preferentially select sparse coefficients related to the fault atoms in the second sub-dictionary and maintain the local neighborhood geometric structure of the signal to be analyzed in the sparse coefficient space;

[0055] The goal of sparse decomposition is to find a sparse coefficient matrix A such that . It is achieved by solving a regularized optimization problem. When solving the problem, an iterative optimization algorithm such as FISTA or ADMM is used. Starting from an initial sparse coefficient matrix A, such as an all-zero matrix, in each iteration, the value of A is updated according to the gradient of the objective function and penalty terms such as the L1 norm and the manifold regularization term, so that it gradually approaches the optimal solution. The iterative process continues until the change in A is less than a preset convergence threshold or the maximum number of iterations is reached, and the obtained A is the desired sparse coefficient matrix.

[0056] In another embodiment, sparse decomposition is achieved by solving an optimization objective function that combines a weighted L1 norm and a graph Laplace regularization term, and an alternating direction multiplier method (ADMM) is used for iterative solution. A larger weight is assigned to the coefficients in the weighted L1 norm term corresponding to the first sub-dictionary, i.e., the normal dictionary atoms, while a smaller weight is assigned to the coefficients corresponding to the second sub-dictionary, i.e., the fault dictionary atoms. This serves as a penalty term to encourage the algorithm to preferentially use the fault atoms to reconstruct the signal. At the same time, a K-nearest neighbor graph based on the Euclidean distance between signal samples is constructed, and the graph Laplace matrix is ​​calculated and introduced into the regularized optimization problem to ensure that adjacent signal points in the original signal space remain adjacent in the sparse coefficient space, thereby enhancing the stability of the decomposition result.

[0057] The two core penalty terms in the regularized optimization problem's objective function consist of two weighted terms added to the overall objective function. Preferential selection of faulty atoms is achieved through a weighted L1-norm penalty term, which assigns smaller penalty weights to coefficients associated with the faulty atom. This ensures that when the optimization solver fits the signal, if a faulty atom can well explain the anomaly in the signal, then it is less costly to select that atom, and thus preferred. Maintaining local neighbor geometry is achieved through a manifold regularization term. This term first analyzes the similarity between signal vectors at each time point in the signal matrix Y to be analyzed—for example, by computing a k-nearest neighbor graph—and then constructs a penalty term that penalizes solutions that are close in the original signal space but far apart in the sparse coefficient space. These two penalty terms, combined with the weighted sum of the reconstruction error term, form the overall objective function of the optimization problem, guiding the solution process to find a sparse solution that accurately reconstructs the signal while being sensitive to faults and exhibiting a stable representation.

[0058] Step three, calculate the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment, and perform exponentially weighted moving average on the energy to obtain a time-series fault accumulation index; when the current value of the fault accumulation index exceeds a preset threshold and the rate of change within a preset time window exceeds a preset change rate threshold, a fault warning is generated.

[0059] From the sparse coefficient matrix obtained by decomposition, all coefficient rows corresponding to the atoms of the second sub-dictionary are separated to form a fault coefficient sub-matrix. For each column of the sub-matrix, that is, each time sampling point, the square sum of all its coefficients is calculated to obtain an instantaneous fault energy sequence. Then, an exponentially weighted moving average filter is applied to the energy sequence. The calculation formula is: the current moment index value is equal to the smoothing factor multiplied by the current moment energy value, plus the difference between the smoothing factor and the index value at that moment, and then multiplied by the index value at that moment. The final smooth and trend-clear fault cumulative index curve is obtained, as shown in the figure below: Figure 3The preset threshold is determined according to the statistical distribution of the fault accumulation index under a large number of normal operating conditions, for example, taking the sum of the mean value and three times the standard deviation; and the preset change rate threshold is obtained by analyzing historical early fault data. First, a fixed length time window is set, such as the last 50 sampling points, and the least square method is used to linearly fit the fault accumulation index sequence in the time window, and the slope of the straight line is the current change rate. When the index value itself is greater than the first threshold, and the change rate slope is also greater than the second threshold, the system determines that the fault is occurring and developing, thereby triggering the early warning.

[0060] In an optional embodiment, the acquisition of the multi-channel sensor signals in the operating cycle of the vacuum laminator comprises:

[0061] Synchronously acquiring a plurality of signals representing the key physical states of the vacuum laminator in the operating cycle of the vacuum laminator, the signals including but not limited to pressure, temperature and mechanical vibration signals;

[0062] Aligning the time series signals of each channel according to the time stamp, and constructing a signal matrix to be analyzed with the sensor channels as rows and the sampling time points as columns.

[0063] In a complete lamination operating cycle, for example, a cycle of 300 seconds, the data of a plurality of sensors are synchronously recorded at a sampling frequency of 1 kHz by a unified master clock. These sensors can include a pressure sensor installed in the vacuum chamber to monitor the process of pressure dropping from standard atmospheric pressure to 10 Pa; a plurality of thermocouples installed on the heating plate to monitor the process of temperature linearly rising from room temperature 25 degrees Celsius to 150 degrees Celsius; and an accelerometer installed on the mechanical structure of the device to capture the mechanical vibration generated in the lamination stage. This synchronous acquisition ensures that at any millisecond level time point, a complete physical state snapshot of the device at that instant can be obtained.

[0064] Specifically, the collected data needs to be arranged into a structured matrix for subsequent algorithm analysis. Taking the above 300-second operating cycle and 1-kHz sampling rate as an example, each sensor channel will generate 300,000 data points. Assuming that three sensors of pressure, temperature and vibration are used, the time series data of the three channels are accurately aligned according to the time stamp. Then, these data are organized into a signal matrix, which has 3 rows corresponding to the three sensor channels and 300,000 columns corresponding to the 300,000 sampling time points. Each element in the matrix represents the measurement value of a particular sensor at a particular time. This finally formed signal matrix is the direct input for subsequent sparse representation and fault diagnosis analysis.

[0065] In an optional embodiment, the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data, wherein the learning imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensing channels, including:

[0066] Classify historical failure data according to known failure modes;

[0067] For each type of fault data, a joint dictionary learning algorithm is used to impose shared sparsity constraints so that the atoms representing the same fault mode on different sensor channels are structurally associated, thereby obtaining the cross-channel physical characteristics of the specific fault.

[0068] To build the second sub-dictionary capable of identifying specific faults, historical data must first be collected and organized. For example, historical equipment problems are categorized into several known failure modes, such as vacuum pump performance degradation, localized heating element failure, or uneven pressure in the pneumatic system. For each failure mode, multiple sets of data, including multi-channel sensor signals such as pressure, temperature, and vibration, are collected. For example, a data set from ten vacuum pump performance degradation events on different equipment is labeled as a vacuum pump failure.

[0069] Specifically, a joint dictionary learning algorithm is applied to each type of fault data, such as vacuum pump faults. The key to this algorithm is that, rather than learning signal patterns independently for each sensor channel, it simultaneously learns by imposing a shared sparsity constraint. This constraint enables the algorithm to discover correlated features across channels. For example, vacuum pump degradation may manifest as both a slowing decrease in the pressure signal rate and the appearance of specific high-frequency noise in the vibration signal. The fault atom learned by the algorithm is a structured pattern that simultaneously describes these two phenomena. In this way, a single fault atom in the dictionary inherently encodes the comprehensive manifestations of a single fault across multiple physical dimensions, significantly improving the accuracy and robustness of fault identification.

[0070] In an optional embodiment, the objective function of the regularized optimization problem is composed of the weighted sum of the following three parts:

[0071] (1) Reconstruction error term, which is the Frobenius norm of the difference between the signal matrix to be analyzed and the sparse reconstruction of the signal matrix to be analyzed;

[0072] (2) A weighted L1-norm penalty term, which prioritizes fault-related sparse representations in the solution by imposing a smaller penalty weight on the sparse coefficients corresponding to the fault atoms in the second sub-dictionary than on the corresponding coefficients in the first sub-dictionary;

[0073] (3) Manifold regularization term, which is used to penalize sparse coefficient solutions that are inconsistent with the local geometric structure of the data points in the signal matrix to be analyzed.

[0074] The core goal of the optimization problem is to represent the signal matrix to be analyzed as accurately as possible using the linear combination of atoms in the dictionary, while minimizing the number of atoms used. The first part of the reconstruction error term measures the difference between the original signal matrix and the signal matrix reconstructed by the dictionary and sparse coefficients. The smaller this difference is, the more faithfully the sparse representation can restore the original signal, ensuring the accuracy of the analysis. For example, the square root of the sum of the squares of the differences between the corresponding elements of the two matrices, that is, the Frobenius norm, will be calculated and minimized. At each step of the optimization solution, given the current sparse coefficient matrix A and the composite dictionary D, the sparse reconstructed signal matrix is ​​first calculated. Then, the reconstructed signal is calculated The difference between the original signal Y to be analyzed, that is, the error matrix The reconstruction error term is the square of the Frobenius norm of the error matrix E, denoted as .

[0075] The weighted L1 norm penalty term is created by creating a weight vector w with the same length as the number of atoms in the composite redundant dictionary D. Assume that the first k atoms of D come from the first sub-dictionary , the last m atoms come from the second sub-dictionary Then, the first k elements of the weight vector w are set to a large value, such as 1.0, and the last m elements are set to a significantly smaller value, such as 0.1. The weighted L1 norm penalty term is then ,in are the elements in the sparse coefficient matrix A, is its corresponding weight, and λ is a hyperparameter that balances the penalty term. In the optimization process, when the algorithm tries to reduce the total cost, the weight corresponding to the failure coefficient is very small, even if its absolute value The larger the , the smaller the contribution to the total cost, thus achieving the priority selection of fault-related sparse representation.

[0076] The manifold regularization term is to preprocess the input signal matrix Y to be analyzed, treating each column, i.e., the multi-channel signal vector at each time point, as a data point. The pairwise distances between these data points, such as the Euclidean distance, are calculated, and a k-nearest neighbor graph is constructed. That is, for each data point, the k nearest neighbors are found. Based on this graph, an adjacency weight matrix W is constructed. If point i and point j are neighbors, then is a positive value, such as the Gaussian kernel function , otherwise 0. Calculate the graph Laplacian matrix L=DW, where D is a diagonal matrix, The manifold regularization term is , where A is the sparse coefficient matrix, Tr is the trace of the matrix, and β is the hyperparameter that balances this term. If the original signal and Very similar, then their sparse coefficients will be given and The difference between By minimizing the manifold regularization term, we can ensure that the local manifold structure in the original data is preserved in the sparse coefficient space, making the sparse representation more stable.

[0077] Specifically, in order to achieve effective fault identification, two key regularization terms are introduced into the objective function. The weighted L1 norm penalty term is the key to distinguishing normal from faulty states. A larger penalty weight, such as 1.5, is applied to the sparse coefficients corresponding to the second sub-dictionary representing the fault mode, while a smaller weight, such as 1.0, is applied to the coefficients corresponding to the first sub-dictionary representing the normal mode. This means that the algorithm will pay a greater price when using fault atoms to reconstruct the signal, so it will only do so when there are very obvious fault features in the signal, so that it has to be explained by fault atoms. The manifold regularization term ensures the smoothness and continuity of the analysis results. It requires that data points that are adjacent in time and have similar signal characteristics should also have similar sparse representations, which helps to suppress noise interference and avoid isolated misjudgments.

[0078] In an optional embodiment, the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment is calculated, and an exponentially weighted moving average is performed on the energy to obtain a time-series fault accumulation index, specifically:

[0079] At each sampling time t, the sparse coefficient vector corresponding to the second sub-dictionary is extracted, and the square of the L2 norm of the sparse coefficient vector is calculated as the instantaneous fault energy ;

[0080] Update the fault accumulation index using the exponentially weighted moving average method , the update formula is: ,in is the cumulative fault index of the previous moment, is the smoothing factor.

[0081] After solving the regularized optimization problem, the sparse coefficients at each moment are obtained. At any sampling time t, special attention is paid to the coefficients associated with the second sub-dictionary. These coefficients are formed into a vector, and the square of the L2 norm of the vector is calculated, which is the sum of the squares of all elements in the vector. This value is the instantaneous fault energy at that moment. If the device is operating normally, the coefficients associated with the second sub-dictionary will be very small or even zero, and the instantaneous fault energy will also be close to zero. Conversely, if fault characteristics appear in the signal, the corresponding fault atoms will be activated, resulting in non-zero coefficients and a significant increase in the instantaneous fault energy.

[0082] Since the instantaneous fault energy may fluctuate violently due to factors such as noise, using it directly will make the fault judgment unstable. Therefore, the exponentially weighted moving average method is used to generate a smoother fault accumulation index that can better reflect the long-term trend. For example, the smoothing factor α can be set to 0.1. Then, the fault accumulation index at the current moment is 10% of the current instantaneous fault energy and 90% of the previous moment's fault accumulation index By filtering out occasional spikes, the indicator can stably reflect the continuous accumulation of fault energy. When a fault begins to develop, the indicator will show a continuous upward trend rather than violent fluctuations.

[0083] In an optional embodiment, when the current value of the fault accumulation indicator exceeds a preset threshold and the rate of change within a preset time window exceeds a preset rate of change threshold, a fault warning is generated, specifically:

[0084] a) the current value of the fault accumulation index is greater than a preset threshold value, the preset threshold value being obtained based on a statistical analysis of the growth trend of the fault accumulation index at the early stage of historical fault occurrence;

[0085] b) The rate of change of the fault accumulation indicator within a preset time window is greater than a preset rate of change threshold, and the preset rate of change threshold is obtained by calculating the growth trend of the indicator sequence within the time window.

[0086] The first trigger condition is a static judgment based on amplitude. This involves analyzing a large amount of cumulative fault indicator data, such as thousands of complete normal operating cycles, and calculating the mean and standard deviation of this health data. Assuming a mean of 0.8 and a standard deviation of 0.2, the first preset threshold is set to the mean plus three times the standard deviation, or 3 times 0.2, based on the statistical three-sigma principle. Only when the real-time calculated cumulative fault indicator exceeds 1.4 is the device health status considered statistically significantly abnormal. This is a necessary condition for triggering an early warning, preventing false alarms caused by minor fluctuations within the normal operating range.

[0087] The second trigger condition is a dynamic judgment based on trends. It observes the fault accumulation indicator series over a recent period of time, such as the past 200 sampling points, and calculates its growth slope, or rate of change, through methods such as linear regression. The rate of change reflects the speed at which the fault develops. In order to set the preset change rate threshold, the confirmed early fault cases in the historical data are analyzed, and their indicator change rates at the initial stage of the fault are calculated. A specific quantile of the distribution of these change rates, such as the 80th percentile, is taken as the threshold. The final issuance of an early warning signal must meet two conditions at the same time: the value of the fault accumulation indicator itself exceeds the static amplitude threshold of 1.4, and it has recently shown a rapid growth momentum that is greater than the preset change rate threshold. The double verification mechanism ensures the reliability of the early warning because it not only identifies the existence of the problem, but also confirms that the problem is worsening.

[0088] Specific embodiment 2 proposes a vacuum laminating machine fault warning system, including the following modules:

[0089] A dictionary construction module is configured to acquire multi-channel sensor signals during the vacuum laminating machine's operating cycle to form a signal matrix to be analyzed. A composite redundant dictionary is constructed by combining a first sub-dictionary representing normal operating conditions and a second sub-dictionary representing specific failure modes. The first sub-dictionary is generated based on a physical model of the vacuum laminating machine, while the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source failure data, imposing structured correlation constraints on atoms belonging to the same failure type but originating from different sensor channels.

[0090] a sparsification module configured to perform a sparse decomposition of the signal matrix to be analyzed by solving a regularized optimization problem based on the composite redundant dictionary to obtain a sparse coefficient matrix; the regularized optimization problem minimizes the reconstruction error while applying a penalty term to preferentially select sparse coefficients associated with fault atoms in the second sub-dictionary and maintain a local neighborhood geometric structure of the signal to be analyzed in the sparse coefficient space;

[0091] An early warning module is used to calculate the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment, and perform an exponentially weighted moving average on the energy to obtain a time-series fault accumulation index; when the current value of the fault accumulation index exceeds a preset threshold and the rate of change within a preset time window exceeds a preset change rate threshold, a fault early warning is generated.

[0092] In a specific embodiment, the step of obtaining multi-channel sensor signals during the operation cycle of the vacuum laminating machine includes:

[0093] Synchronously collecting multiple signals representing key physical states of the vacuum laminating machine during its working cycle, including but not limited to pressure, temperature, and mechanical vibration signals;

[0094] The timing signals of each channel are aligned according to the timestamps and constructed into a signal matrix to be analyzed with sensor channels as rows and sampling time points as columns.

[0095] In a specific embodiment, the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data, wherein the learning imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensing channels, including:

[0096] Classify historical failure data according to known failure modes;

[0097] For each type of fault data, a joint dictionary learning algorithm is used to impose shared sparsity constraints so that the atoms representing the same fault mode on different sensor channels are structurally associated, thereby obtaining the cross-channel physical characteristics of the specific fault.

[0098] Optionally, the objective function of the regularized optimization problem is composed of the weighted sum of the following three parts:

[0099] (1) Reconstruction error term, which is the Frobenius norm of the difference between the signal matrix to be analyzed and the sparse reconstruction of the signal matrix to be analyzed;

[0100] (2) A weighted L1-norm penalty term, which prioritizes fault-related sparse representations in the solution by imposing a smaller penalty weight on the sparse coefficients corresponding to the fault atoms in the second sub-dictionary than on the corresponding coefficients in the first sub-dictionary;

[0101] (3) Manifold regularization term, which is used to penalize sparse coefficient solutions that are inconsistent with the local geometric structure of the data points in the signal matrix to be analyzed.

[0102] In a specific embodiment, the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment is calculated, and the energy is subjected to an exponentially weighted moving average to obtain a time-series fault accumulation index, specifically:

[0103] At each sampling time t, the sparse coefficient vector corresponding to the second sub-dictionary is extracted, and the square of the L2 norm of the sparse coefficient vector is calculated as the instantaneous fault energy ;

[0104] Update the fault accumulation index using the exponentially weighted moving average method , the update formula is: ,in is the cumulative fault index of the previous moment, is the smoothing factor.

[0105] In a specific embodiment, when the current value of the fault accumulation indicator exceeds a preset threshold and the rate of change within a preset time window exceeds a preset rate of change threshold, a fault warning is generated, specifically:

[0106] a) the current value of the fault accumulation index is greater than a preset threshold value, the preset threshold value being obtained based on a statistical analysis of the growth trend of the fault accumulation index at the early stage of historical fault occurrence;

[0107] b) The rate of change of the fault accumulation indicator within a preset time window is greater than a preset rate of change threshold, and the preset rate of change threshold is obtained by calculating the growth trend of the indicator sequence within the time window.

[0108] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present application or certain parts of the embodiments.

[0109] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0110] The above describes in detail the method and electronic device for providing commodity object information provided by this application. Specific examples are used herein to illustrate the principles and implementation methods of this application. The description of the above embodiments is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the contents of this specification should not be construed as limiting this application.

Claims

1. A vacuum pressing machine fault warning method, characterized in that: The following steps are involved: Multi-channel sensor signals are acquired during the operating cycle of a vacuum laminating machine to form a signal matrix to be analyzed. A composite redundant dictionary is constructed by combining a first sub-dictionary representing normal operating conditions and a second sub-dictionary representing specific failure modes. The first sub-dictionary is generated based on a physical model of the vacuum laminating machine, and the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data. The learning method applies structured correlation constraints to atoms belonging to the same fault type but originating from different sensor channels. Based on the composite redundant dictionary, the signal matrix to be analyzed is sparsely decomposed by solving a regularized optimization problem to obtain a sparse coefficient matrix; the regularized optimization problem minimizes the reconstruction error while applying a penalty term to preferentially select sparse coefficients associated with fault atoms in the second sub-dictionary and maintain the local neighborhood geometric structure of the signal to be analyzed in the sparse coefficient space; The energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment is calculated, and the energy is subjected to an exponentially weighted moving average to obtain a time-series fault accumulation index; when the current value of the fault accumulation index exceeds a preset threshold and the rate of change within a preset time window exceeds a preset change rate threshold, a fault warning is generated.

2. The method according to claim 1, characterized in that The method of obtaining multi-channel sensor signals during the operation cycle of the vacuum laminating machine includes: Synchronously collecting multiple signals representing key physical states of the vacuum laminating machine during its working cycle, including but not limited to pressure, temperature, and mechanical vibration signals; The timing signals of each channel are aligned according to the timestamps and constructed into a signal matrix to be analyzed with sensor channels as rows and sampling time points as columns.

3. The method according to claim 1, characterized in that The second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data. The learning imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensing channels, including: Classify historical failure data according to known failure modes; For each type of fault data, a joint dictionary learning algorithm is used to impose shared sparsity constraints so that the atoms representing the same fault mode on different sensor channels are structurally associated, thereby obtaining the cross-channel physical characteristics of the specific fault.

4. The method according to claim 1, wherein The objective function of the regularized optimization problem is composed of the weighted sum of the following three parts: The reconstruction error term is the Frobenius norm of the difference between the signal matrix to be analyzed and the sparse reconstruction of the signal matrix to be analyzed; A weighted L1-norm penalty term prioritizes fault-related sparse representations in the solution by imposing a smaller penalty weight on the sparse coefficients corresponding to fault atoms in the second sub-dictionary than on the corresponding coefficients in the first sub-dictionary. The manifold regularization term is used to penalize sparse coefficient solutions that are inconsistent with the local geometric structure of the data points in the signal matrix to be analyzed.

5. The method according to claim 1, wherein Calculate the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment, and perform an exponentially weighted moving average on the energy to obtain a time-series fault accumulation index, specifically: At each sampling time t, the sparse coefficient vector corresponding to the second sub-dictionary is extracted, and the square of the L2 norm of the sparse coefficient vector is calculated as the instantaneous fault energy ; Update the fault accumulation index using the exponentially weighted moving average method , the update formula is: ,in is the cumulative fault index of the previous moment, is the smoothing factor.

6. The method according to claim 1, characterized in that The triggering conditions for generating a fault warning are: a) the current value of the fault accumulation index is greater than a preset threshold value, the preset threshold value being obtained based on a statistical analysis of the growth trend of the fault accumulation index at the early stage of historical fault occurrence; b) The rate of change of the fault accumulation indicator within a preset time window is greater than a preset rate of change threshold, and the preset rate of change threshold is obtained by calculating the growth trend of the indicator sequence within the time window.

7. A vacuum pressing machine fault warning system, characterized in that: Includes the following modules: A dictionary construction module is configured to acquire multi-channel sensor signals during the vacuum laminating machine's operating cycle to form a signal matrix to be analyzed. A composite redundant dictionary is constructed by combining a first sub-dictionary representing normal operating conditions and a second sub-dictionary representing specific failure modes. The first sub-dictionary is generated based on a physical model of the vacuum laminating machine, while the second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source failure data, imposing structured correlation constraints on atoms belonging to the same failure type but originating from different sensor channels. a sparsification module configured to perform a sparse decomposition of the signal matrix to be analyzed by solving a regularized optimization problem based on the composite redundant dictionary to obtain a sparse coefficient matrix; the regularized optimization problem minimizes the reconstruction error while applying a penalty term to preferentially select sparse coefficients associated with fault atoms in the second sub-dictionary and maintain a local neighborhood geometric structure of the signal to be analyzed in the sparse coefficient space; An early warning module is used to calculate the energy of the fault coefficient component corresponding to the second sub-dictionary in the sparse coefficient matrix at each moment, and perform an exponentially weighted moving average on the energy to obtain a time-series fault accumulation index; when the current value of the fault accumulation index exceeds a preset threshold and the rate of change within a preset time window exceeds a preset change rate threshold, a fault early warning is generated.

8. The system according to claim 7, characterized in that The method of obtaining multi-channel sensor signals during the operation cycle of the vacuum laminating machine includes: Synchronously collecting multiple signals representing key physical states of the vacuum laminating machine during its working cycle, including but not limited to pressure, temperature, and mechanical vibration signals; The timing signals of each channel are aligned according to the timestamps and constructed into a signal matrix to be analyzed with sensor channels as rows and sampling time points as columns.

9. The system according to claim 7, wherein: The second sub-dictionary is generated by performing constrained dictionary learning on historical multi-source fault data. The learning imposes structured correlation constraints on atoms belonging to the same fault type but originating from different sensing channels, including: Classify historical failure data according to known failure modes; For each type of fault data, a joint dictionary learning algorithm is used to impose shared sparsity constraints so that the atoms representing the same fault mode on different sensor channels are structurally associated, thereby obtaining the cross-channel physical characteristics of the specific fault.

10. The system according to claim 7, wherein: The objective function of the regularized optimization problem is composed of the weighted sum of the following three parts: The reconstruction error term is the Frobenius norm of the difference between the signal matrix to be analyzed and the sparse reconstruction of the signal matrix to be analyzed; A weighted L1-norm penalty term prioritizes fault-related sparse representations in the solution by imposing a smaller penalty weight on the sparse coefficients corresponding to fault atoms in the second sub-dictionary than on the corresponding coefficients in the first sub-dictionary. The manifold regularization term is used to penalize sparse coefficient solutions that are inconsistent with the local geometric structure of the data points in the signal matrix to be analyzed.

Images