Digital twinning-based miniature production line real-time monitoring and fault prediction method and system

Abstract

The invention provides a digital twinning-based miniature production line real-time monitoring and fault prediction method and system, and relates to the technical field of fault prediction, and the method comprises the steps: obtaining reference data, real-time data and a continuous waveform signal of a miniature production line, inputting the real-time data and the reference data into a digital twinning model for comparison, and obtaining a digital twinning model; obtaining structure change data; performing sparse transformation on the continuous waveform signal by adopting a sparse representation model to generate a sparse event sequence, and performing plasticity learning according to structure change data to obtain a plurality of suspected fault marks; and finally, through an isolated forest algorithm, screening out an abnormal event queue and carrying out causal verification, and feeding back a verification report to the digital twin model. According to the invention, high-reliability fault source positioning and prediction in a microsecond fault propagation scene can be realized.

Description

Technical Field

[0001] This application relates to the technical field of fault prediction, and in particular to a method and system for real-time monitoring and fault prediction of a micro production line based on digital twins. Background Technology

[0002] In the field of micro high-speed production lines, real-time monitoring and fault prediction of equipment operating status are key to ensuring production continuity and product consistency, and have significant engineering application value.

[0003] Currently, fault diagnosis methods based on digital twins are mainly divided into two categories: one is to collect real-time operating data and compare it with the simulation model, and then use threshold judgment or machine learning model to identify abnormal states; the other focuses on extracting features from waveform signals such as vibration, and then combining them with historical fault databases for pattern matching to achieve fault early warning.

[0004] However, existing methods often struggle to effectively integrate multi-source heterogeneous data when dealing with micro high-speed production lines where faults propagate at extremely fast speeds, and they are highly dependent on fault samples. This makes it impossible to accurately trace the root cause of faults that propagate at the microsecond level and their causal relationships in scenarios where fault data is scarce, resulting in insufficient reliability and interpretability of the prediction results. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for real-time monitoring and fault prediction of micro production lines based on digital twins, so as to solve the problem of low reliability of fault prediction in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for real-time monitoring and fault prediction of a micro production line based on digital twins, comprising: Acquire baseline data of the micro production line at the time of acceptance, real-time data of the micro production line during operation, and continuous waveform signals of the mechanical frame; The real-time data and the baseline data are input into the digital twin model for comparison to obtain structural change data; The continuous waveform signal is sparsely transformed using a sparse representation model to generate a sparse event sequence. Based on the structural change data, the sparse event sequence is subjected to plasticity learning to obtain multiple suspected fault markers. The isolated forest algorithm is used to filter all the suspected fault markers for abnormal patterns to obtain an abnormal event queue. Causal verification is performed on the abnormal event queue to obtain a verification report, and the verification report is fed back into the digital twin model.

[0007] Optionally, the step of performing causal verification on the abnormal event queue to obtain a verification report includes: Using the mechanical positions corresponding to each suspected fault marker in the abnormal event queue as nodes, and the order of pulse arrival times between the mechanical positions as directed edges, a causal structure graph consisting of multiple nodes and multiple directed edges is constructed. Using a counterfactual reasoning method, an intervention operation is performed on each node in the causal structure graph to generate a counterfactual sample corresponding to each node; The counterfactual samples and the original samples in the abnormal event queue are input into the Siamese network. The similarity distance between the corresponding counterfactual samples and the original samples is calculated through the Siamese network. Based on the similarity distance, the contribution factor corresponding to each node is calculated. Based on all the contribution factors and the structural change data corresponding to each mechanical position, calculate the fault confidence level for each mechanical position, and mark the mechanical positions whose fault confidence level exceeds a preset confidence threshold as the root cause of the fault. Based on the location of the fault root cause and the corresponding downstream node in the causal structure diagram, a verification report is generated that includes the location of the fault root cause, the fault confidence level, and the fault propagation path.

[0008] Optionally, the step of inputting the counterfactual sample and the original sample in the anomalous event queue into a Siamese network, and calculating the similarity distance between the corresponding counterfactual sample and the original sample through the Siamese network, includes: The counterfactual sample corresponding to each mechanical position is input into the first feature extraction branch of the Siamese network for full convolution processing to obtain the counterfactual feature vector; The original samples corresponding to the same mechanical position are input into the second feature extraction branch of the twin network for full convolution processing to obtain the original feature vector, wherein the first feature extraction branch and the second feature extraction branch share the network weights. The counterfactual feature vector and the original feature vector are input into the computational branch of the Siamese network, and the computational branch uses a cross-correlation calculation method to calculate the similarity distance between the counterfactual feature vector and the original feature vector.

[0009] Optionally, the step of performing a sparse transformation on the continuous waveform signal using a sparse representation model to generate a sparse event sequence includes: Multiple waveform segments corresponding to the mechanical positions indicated by the structural change data are extracted from the continuous waveform signal, and all the waveform segments are used as a training sample set. The singular value decomposition algorithm is used to iteratively learn the training sample set to construct an overcomplete redundant dictionary. Using the overcomplete redundant dictionary as a sparse representation model, the orthogonal matching pursuit algorithm is used to project and decompose the subsequently acquired continuous waveform signals onto the sparse representation model to obtain a sparse coefficient sequence. Extract the atom index, pulse arrival time, and atom amplitude corresponding to each non-zero projection coefficient from the sparse coefficient sequence, and combine the atom index, the pulse arrival time, and the atom amplitude to form a sparse event sequence.

[0010] Optionally, based on the structural change data, plasticity learning is performed on the sparse event sequence to obtain multiple suspected fault markers, including: Assign a cluster of spiking neurons to the mechanical location indicated by the structural change data; The pulse time-dependent plasticity algorithm is used to calculate the arrival time difference between presynaptic pulses and postsynaptic pulses in each spiking neuron cluster. Based on the arrival time difference, the weight adjustment direction of long-term enhancement or long-term inhibition is determined, and the synaptic weights of the spiking neuron cluster are updated according to the weight adjustment direction. An attention mechanism is used to calculate the attention weights of each mechanical position, and the synaptic weights are modulated according to the attention weights to obtain modulated synaptic weights; The modulated synaptic weights are compared with a preset first activation threshold. When the comparison result indicates that the modulated synaptic weights exceed the first activation threshold, the mechanical position corresponding to the current spiking neuron cluster is recorded as the activation position, and the activation position and the corresponding pulse arrival time are combined as a candidate fault marker. The amplitude of the synaptic weights corresponding to the candidate fault markers is compressed to obtain normalized synaptic weights. A dynamic gating mechanism is used to apply periodic gating signals to the normalized synaptic weights to obtain synaptic weights after the signals are applied. The synaptic weights after the applied signal are compared with a preset second activation threshold. When the result of the second comparison indicates that the synaptic weights after the applied signal exceed the second activation threshold, the mechanical position corresponding to the current spiking neuron cluster is recorded as a suspected fault position, and the suspected fault position and the corresponding pulse arrival time are combined to form a suspected fault marker.

[0011] Optionally, the step of filtering all the suspected fault markers for abnormal patterns using the isolated forest algorithm to obtain an abnormal event queue includes: Based on the rate of change of each mechanical position in the structural change data, a dynamic weight coefficient is assigned to each mechanical position. The dynamic weight coefficient is multiplied by the synaptic weight value after the signal is applied in the suspected fault mark of the corresponding mechanical position to obtain the weighted suspected fault mark. All the weighted suspected fault markers are arranged in chronological order to form a cycle marker sequence for each production cycle. The cycle marker sequence is then subjected to dimensionality reduction transformation using a random projection method to obtain a low-dimensional feature vector. The low-dimensional feature vector is then input into the isolated forest algorithm. In each isolated tree, a feature dimension and a splitting value are randomly selected, and the low-dimensional feature vector is recursively split until each low-dimensional feature vector is split into a leaf node. The path length of each low-dimensional feature vector in each isolated tree is recorded. Calculate the average path length based on the path length of each of the low-dimensional feature vectors in all the isolated trees, and calculate the anomaly index of each of the periodic marker sequences based on the average path length; The suspected fault markers that exceed the preset abnormal threshold in each production cycle are retained, and the retained suspected fault markers are arranged according to the pulse arrival time to form an abnormal event queue.

[0012] Optionally, the step of comparing the real-time data with the reference data in the digital twin model to obtain structural change data includes: The peak frequency corresponding to each mechanical position is extracted from the real-time data. The difference between the peak frequency and the reference peak frequency of the corresponding mechanical position in the reference data is calculated by the comparison engine of the digital twin model to obtain the frequency offset of each mechanical position. The amplitude attenuation rate corresponding to each mechanical position is extracted from the real-time data. The amplitude attenuation rate is then compared with the reference attenuation rate of the corresponding mechanical position in the reference data using the comparison engine to obtain the amplitude change coefficient of each mechanical position. The frequency offset and amplitude change coefficient of each mechanical position are weighted and combined using the calculation engine of the digital twin model to obtain the comprehensive change index of each mechanical position. Using the digital twin model's combination engine, all the aforementioned comprehensive change indices are arranged according to their mechanical positions to form structural change data.

[0013] Secondly, this application provides a real-time monitoring and fault prediction system for a micro production line based on digital twins, comprising: The acquisition module is used to acquire the baseline data of the micro production line at the time of acceptance, the real-time data of the micro production line during the operation period, and the continuous waveform signal of the mechanical frame. The comparison module is used to input the real-time data and the reference data into the digital twin model for comparison to obtain structural change data; The transformation module is used to perform sparse transformation on the continuous waveform signal using a sparse representation model to generate a sparse event sequence. Based on the structural change data, the sparse event sequence is subjected to plasticity learning to obtain multiple suspected fault markers. The verification module is used to filter all the suspected fault markers for abnormal patterns using the isolated forest algorithm to obtain an abnormal event queue, perform causal verification on the abnormal event queue, obtain a verification report, and feed the verification report back to the digital twin model.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor, used to execute the computer program to implement the steps of the digital twin-based micro production line real-time monitoring and fault prediction method as described in the first aspect above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the digital twin-based micro production line real-time monitoring and fault prediction method described in the first aspect above.

[0016] The digital twin-based real-time monitoring and fault prediction method for micro-production lines provided in this application has the following beneficial effects: First, by acquiring benchmark data, real-time data, and continuous waveform signals, this application can provide a multi-dimensional data foundation for subsequent analysis. Then, by comparing the real-time data and benchmark data in the digital twin model, structural change data is obtained, which can achieve a quantitative assessment of the degree of deviation of the mechanical state. Subsequently, a sparse representation model is used to perform sparse transformation on the continuous waveform signals to generate a sparse event sequence. Combined with the structural change data, plasticity learning is performed, which can efficiently extract suspected fault markers related to structural changes from massive signals. Then, the isolated forest algorithm is used to screen all suspected fault markers for abnormal patterns to obtain an abnormal event queue. Causal verification is then performed on the abnormal event queue, and finally, the verification report is fed back to the digital twin model, thereby achieving accurate tracing of microsecond-level fault propagation and model self-optimization.

[0017] Furthermore, this application constructs a causal structure graph with mechanical position as nodes and pulse arrival time as directed edges, and uses counterfactual reasoning to generate counterfactual samples corresponding to each node. Then, it uses a Siamese network to calculate the similarity distance between the counterfactual samples and the original samples to quantify the node contribution. In addition, it combines structural change data to calculate the fault confidence and mark the fault root cause location. Finally, it generates a verification report containing the fault root cause, confidence, and propagation path. This can accurately separate the causal chain from the time-series mixed abnormal events, thereby enhancing the interpretability and credibility of the fault prediction results and providing a reliable decision-making basis for the precise maintenance of micro high-speed production lines. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating the real-time monitoring and fault prediction method for a micro production line based on digital twins provided in this application embodiment; Figure 2 A schematic diagram illustrating a specific implementation of the digital twin-based real-time monitoring and fault prediction method for a micro production line provided in this application embodiment; Figure 3 A schematic diagram of the structure of a real-time monitoring and fault prediction system for a micro production line based on digital twins provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] In fault prediction of micro high-speed production lines, existing methods typically process equipment condition monitoring and waveform signal analysis separately, making it difficult to effectively integrate multi-source heterogeneous data. At the same time, due to the fact that fault propagation speed can reach the microsecond level and fault samples are extremely scarce, traditional machine learning models often fail to accurately capture early signs of faults due to a lack of sufficient training data, resulting in insufficient reliability of prediction results. In addition, existing technologies have difficulty tracing the propagation path of faults after identifying anomalies, making the prediction results lack interpretability and unable to provide effective guidance for precise maintenance.

[0021] To address this, this application proposes a real-time monitoring and fault prediction method for micro-production lines based on digital twins. The core of this method lies in: firstly, comparing benchmark data with real-time data to quantify equipment structural changes; then, using a sparse representation model to transform waveform signals and generate sparse event sequences, and combining structural change data to simulate the learning mechanism of biological neurons to extract suspected fault markers from massive signals; finally, using the isolated forest algorithm to filter abnormal patterns and performing causal verification on the abnormal event queue, feeding the verification results back into the digital twin model to achieve closed-loop optimization. This method achieves a seamless process from multi-source data fusion and weak symptom capture to causal chain tracing, effectively overcoming the problems of low reliability in fault prediction and difficulty in achieving interpretable traceability in existing technologies, and is particularly suitable for micro-high-speed production line scenarios with extremely fast fault propagation speeds.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The core of this application is to provide a method for real-time monitoring and fault prediction of a micro production line based on digital twins. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Acquire the baseline data of the micro production line at the time of acceptance, the real-time data of the micro production line during the operation period, and the continuous waveform signal of the mechanical frame.

[0024] Among them, a micro production line is a manufacturing system composed of multiple functional units, such as micro motors, grippers, conveyor belts, bases, support columns, guide rails, etc. Its core function is to complete the assembly, processing or testing of micro products. The mechanical skeleton supports and connects these functional units, and includes physical entities such as the basic frame of the production line, equipment mounting base, support columns, and connectors. Baseline data refers to the standard state parameters collected and recorded by various sensors when a micro production line has completed installation, commissioning, and passed acceptance. These parameters include the baseline vibration frequency of each motion axis, the standard temperature value, and the standard current curve of the drive motor. It represents the initial template for the healthy operation of the equipment.

[0025] Real-time data refers to the current status parameters continuously collected by the same monitoring system during the subsequent actual operation of a micro production line, such as the current vibration spectrum, real-time temperature, and real-time power consumption of the motor, which are used to reflect the immediate operating status of the equipment. Continuous waveform signals refer to the original vibration waveforms captured at an extremely high sampling rate by accelerometers or acoustic emission sensors installed on the mechanical frame, which contain high-frequency impact components excited by minor damage or friction to the mechanical structure.

[0026] In step S101, continuous waveform signals of the mechanical skeleton are collected in real time by fiber optic grating sensors or piezoelectric accelerometers deployed at key workstations on the production line; at the same time, the reference data stored at the time of acceptance and the real-time data of the current operating status are acquired synchronously by the programmable logic controller or dedicated data acquisition card of the production line.

[0027] S102. The real-time data and the reference data are input into the digital twin model for comparison to obtain structural change data.

[0028] In one specific implementation, step S102 includes: Step 1021: Extract the peak frequency corresponding to each mechanical position from the real-time data, and use the comparison engine of the digital twin model to perform difference calculation between the peak frequency and the reference peak frequency of the corresponding mechanical position in the reference data to obtain the frequency offset of each mechanical position.

[0029] Among them, peak frequency refers to the frequency value corresponding to the maximum amplitude in the mechanical vibration signal, which represents the main vibration mode of the component; reference peak frequency refers to the standard value of peak frequency recorded at the time of acceptance of the equipment.

[0030] In step 1021, the peak frequency corresponding to each mechanical position is extracted from the real-time data. Then, the peak frequency is compared with the reference peak frequency of the corresponding mechanical position in the reference data through the comparison engine of the digital twin model to obtain the frequency offset of each mechanical position. The frequency offset reflects the stiffness change or loosening degree of the mechanical structure.

[0031] Step 1022: Extract the amplitude attenuation rate corresponding to each mechanical position from the real-time data, and use the comparison engine to perform a ratio calculation between the amplitude attenuation rate and the reference attenuation rate of the corresponding mechanical position in the reference data to obtain the amplitude change coefficient of each mechanical position.

[0032] Among them, the amplitude attenuation rate refers to the rate at which the amplitude of a mechanical vibration signal decays per unit time. It is usually obtained by calculating the amplitude ratio of adjacent vibration cycles and taking the logarithm, reflecting the damping characteristics of the system; the reference attenuation rate refers to the standard value of the amplitude attenuation rate recorded at the time of acceptance of the equipment.

[0033] In step 1022, the amplitude attenuation rate corresponding to each mechanical position is extracted from the real-time data, and the amplitude attenuation rate is compared with the reference attenuation rate of the corresponding mechanical position in the reference data by the comparison engine to obtain the amplitude change coefficient of each mechanical position, which reflects the change of mechanical friction or lubrication state.

[0034] Step 1023: Using the calculation engine of the digital twin model, the frequency offset and amplitude change coefficient of each mechanical position are weighted and combined to obtain the comprehensive change index of each mechanical position.

[0035] In step 1023, the frequency offset and amplitude change coefficient of each mechanical position are weighted and combined using the calculation engine of the digital twin model to obtain the comprehensive change index of each mechanical position.

[0036] Specifically, its calculation engine uses preset weighting coefficients, where the frequency offset reflects changes in structural stiffness and is assigned a weight of α; the amplitude variation coefficient reflects changes in damping characteristics and is assigned a weight of β. The formula for calculating the comprehensive variation index is as follows: ,in, This represents the comprehensive change index of the i-th mechanical position. This represents the frequency offset of the i-th mechanical position. This represents the reference peak frequency of the i-th mechanical position. This represents the amplitude variation coefficient at the i-th mechanical position. and All are preset weighting coefficients and satisfy the following conditions: + =1, and the values ​​of both are usually set according to the type of equipment. For example, for rigid connection components, =0.7、 =0.3, for sliding parts, =0.4、 =0.6.

[0037] Step 1024: Using the combination engine of the digital twin model, arrange all the comprehensive change indices according to their mechanical positions to form structural change data.

[0038] In step 1024, the combined engine sorts and encapsulates the comprehensive change index according to the physical layout of the mechanical positions, such as the feeding end, processing station, and discharging end, thereby generating a structured dataset, in which each element corresponds to the degree of change of a specific mechanical position, providing spatial location dimension input for subsequent plasticity learning.

[0039] This application achieves a quantitative assessment of changes in the structural state of equipment by comparing multidimensional real-time data with benchmark data, thereby providing spatial location guidance for extracting suspected fault markers from continuous waveform signals and thus closely linking signal analysis with structural changes.

[0040] S103. The continuous waveform signal is subjected to sparse transformation using a sparse representation model to generate a sparse event sequence. Based on the structural change data, the sparse event sequence is subjected to plasticity learning to obtain multiple suspected fault markers.

[0041] In one specific implementation, step S103 includes: Step 1031: Extract multiple waveform segments from the continuous waveform signal that correspond to the mechanical positions indicated by the structural change data, and use all the waveform segments as a training sample set. Use the singular value decomposition algorithm to iteratively learn the training sample set and construct an overcomplete redundant dictionary.

[0042] Among them, the overcomplete redundant dictionary is an overcomplete set composed of atoms, where each atom represents a basic waveform template, and the number of columns in the dictionary is much greater than the number of rows. There may be linear correlation between atoms, which can more flexibly approximate waveform signals of various forms.

[0043] In step 1031, firstly, based on the obtained structural change data, the mechanical positions that need to be focused on are determined, and waveform segments corresponding to these positions are extracted from the continuous waveform signal. Then, all waveform segments are arranged into a matrix form as a training sample set. Then, the singular value decomposition algorithm is used to decompose the matrix: the matrix is ​​decomposed into the product of the left singular matrix, the singular value matrix, and the right singular matrix, and the largest K singular values ​​and their corresponding singular vectors are retained. Next, the dictionary is updated through multiple iterations: in each iteration, the dictionary is fixed, and the sparse coefficients of the training sample set under the current dictionary are calculated using the orthogonal matching pursuit algorithm; then, the sparse coefficients are fixed, and the atoms in the dictionary are updated through singular value decomposition, and the above process is repeated until the preset maximum number of iterations is reached, such as 50, or the change in dictionary reconstruction error between two adjacent iterations is less than a preset threshold, such as 10. -6 At this point, the dictionary is considered to have converged, the iteration stops, and an overcomplete redundant dictionary is finally generated.

[0044] Step 1032: Using the overcomplete redundant dictionary as a sparse representation model, the orthogonal matching pursuit algorithm is used to project and decompose the subsequently acquired continuous waveform signals onto the sparse representation model to obtain a sparse coefficient sequence.

[0045] In step 1032, firstly, an empty sparse representation model is created, which reserves two slots: one for storing a dictionary and the other for an association algorithm. Then, the overcomplete redundant dictionary is loaded into the dictionary slot of the sparse representation model, and the code implementation of the orthogonal matching pursuit algorithm is bound to the other slot of the model. At this point, the model is completed, and the model provides a unified interface to the outside world. It stores all the trained atoms in its content, and each atom represents a basic waveform template.

[0046] When subsequent continuous waveform signals arrive, they are input into the sparse representation model through this interface. An orthogonal matching pursuit algorithm is then employed, and the input signal is decomposed using an overcomplete redundant dictionary: the residual vector is initialized with the input continuous waveform signal, and the selected atom set is initialized as an empty set. In each iteration, the inner product of the current residual and each atom in the dictionary is calculated, and the atom with the largest absolute value of the inner product is selected as the best-matching atom. This atom index is then added to the selected atom set. The projection coefficients corresponding to the current selected atom set are then calculated using the least squares method, ensuring that the linear combination of these atoms best approximates the original signal. The residual is then updated to the original signal minus the remaining portion after subtracting the linear combination of the current selected atoms. This iterative process is repeated until a preset sparsity is reached (i.e., at most L atoms are selected), or the residual energy is below a preset threshold. Finally, the projection coefficients corresponding to each atom are obtained, where the atoms corresponding to non-zero coefficients are the atoms used to represent the signal, and all non-zero coefficients constitute a sparse coefficient sequence.

[0047] Step 1033: Extract the atom index, pulse arrival time, and atom amplitude corresponding to each non-zero projection coefficient from the sparse coefficient sequence, and combine the atom index, the pulse arrival time, and the atom amplitude into a sparse event sequence.

[0048] Here, the atom index refers to the unique number of each atom in the overcomplete redundancy dictionary; the pulse arrival time refers to the time point at which a significant impact component appears in the waveform segment; and the atom amplitude refers to the projection coefficient value corresponding to that atom, reflecting the intensity of that waveform component.

[0049] In step 1033, for each non-zero coefficient in the sparse coefficient sequence, the position number of the corresponding atom in the dictionary is recorded as the atom index, and the pulse arrival time is obtained by locating the peak moment of the impact component in the waveform segment. The value of the coefficient is used as the atom amplitude. Then, the above three elements are combined into an event record, and all the event records corresponding to the non-zero coefficients are arranged in chronological order to form a sparse event sequence.

[0050] Step 1034: Assign a cluster of spiking neurons to the mechanical position indicated by the structural change data.

[0051] In step 1034, based on the structural change data, all mechanical positions that need to be monitored are determined, and a cluster consisting of multiple spiking neurons is assigned to each mechanical position. Each cluster is responsible for receiving and processing the spiking events in the sparse event sequence corresponding to that mechanical position. The spiking neuron clusters of different mechanical positions are independent of each other and process the spiking input of their respective positions in parallel.

[0052] Step 1035: Using the pulse time-dependent plasticity algorithm, calculate the arrival time difference between the presynaptic pulse and the postsynaptic pulse in each spiking neuron cluster. Based on the arrival time difference, determine the weight adjustment direction for long-term enhancement or long-term inhibition, and update the synaptic weights of the spiking neuron cluster according to the weight adjustment direction.

[0053] Among them, the presynaptic pulse refers to the pulse signal that arrives at the input end of the neuron; the postsynaptic pulse refers to the pulse signal that is output from the neuron.

[0054] In step 1035, for each spiking neuron cluster, the arrival time tpre of each presynaptic pulse and the emission time tpost of each postsynaptic pulse are recorded, and the time difference Δt = tpost - tpre is calculated. Then, according to the pulse time-dependent plasticity algorithm, the following analysis is performed: when Δt is greater than 0, i.e., the presynaptic pulse arrives before the postsynaptic pulse, long-term enhancement is triggered, and the synaptic weight increases; when Δt is less than 0, i.e., the presynaptic pulse arrives later than the postsynaptic pulse, long-term inhibition is triggered, and the synaptic weight decreases. The change in weight is calculated by the following formula: Where Δw is the change in synaptic weight. and These represent the maximum magnitudes of enhancement and inhibition, respectively. and These are the time constants for enhancement and suppression, respectively. In this embodiment, their values ​​are not limited. Then, the weight changes caused by each pulse pair are accumulated to obtain the final synaptic weight update amount. The current synaptic weight is adjusted based on this update amount.

[0055] Step 1036: Using an attention mechanism, calculate the attention weights of each mechanical position, and modulate the synaptic weights according to the attention weights to obtain modulated synaptic weights.

[0056] In step 1036, an attention mechanism is used to calculate the attention weight of each mechanical position. This weight is determined based on the comprehensive change index in the structural change data. That is, the higher the comprehensive change index, the stronger the correlation between the position and the pulse activity, thus obtaining a higher attention weight. Then, the attention weight is multiplied by the synaptic weight of the corresponding mechanical position to obtain the modulated synaptic weight, so that the position with significant structural changes obtains a greater weight in subsequent processing.

[0057] Step 1037: Compare the modulated synaptic weights with a preset first activation threshold. When the comparison result indicates that the modulated synaptic weights exceed the first activation threshold, record the mechanical position corresponding to the current spiking neuron cluster as the activation position, and combine the activation position with the corresponding pulse arrival time as a candidate fault marker.

[0058] The first trigger threshold is a preset critical value used to initially screen possible fault events in order to capture all possible abnormal events. In this embodiment, the value of this threshold is not limited.

[0059] In step 1037, for each mechanical position, it is determined whether its modulated synaptic weight is greater than the first excitation threshold. If it is greater, it is considered that a noteworthy pulse activity has occurred at the current moment, and the mechanical position and the current pulse arrival time are recorded as a candidate fault marker.

[0060] Step 1038: Perform amplitude compression on the synaptic weights corresponding to the candidate fault markers to obtain normalized synaptic weights. Use a dynamic gating mechanism to apply periodic gating signals to the normalized synaptic weights to obtain the synaptic weights after the signals are applied.

[0061] In step 1038, a nonlinear transformation function, such as a logarithmic transformation or a sigmoid function, is used to map the synaptic weights corresponding to the candidate fault markers to the interval between 0 and 1, resulting in normalized synaptic weights. Then, a dynamic gating mechanism generates a periodic gating signal associated with the production line's operating cycle. This signal is turned on during a specific phase and turned off during other phases. The normalized synaptic weights are multiplied by the gating signal. Only when the gating signal is on and the synaptic weights are high can a synaptic weight with a value greater than the applied signal be obtained.

[0062] Step 1039: Compare the synaptic weights after the applied signal with the preset second activation threshold. When the result of the second comparison indicates that the synaptic weights after the applied signal exceed the second activation threshold, record the mechanical position corresponding to the current spiking neuron cluster as a suspected fault position, and combine the suspected fault position with the corresponding pulse arrival time as a suspected fault marker.

[0063] The second activation threshold is a preset high-confidence threshold for secondary confirmation. Its value is higher than the first activation threshold, which is used to ensure that the suspected fault markers in the final output have high confidence.

[0064] In step 1039, for each candidate fault marker, it is determined whether the synaptic weight after the signal is applied is greater than the second excitation threshold. If it is greater, it is considered that the position still shows significant abnormal activity after plasticity learning and dynamic gating modulation, and the mechanical position and the corresponding pulse arrival time are recorded as a suspected fault marker.

[0065] It is worth noting that the first and second excitation thresholds are related in that the first threshold is smaller than the second threshold. This design has the following advantages: First, the first threshold ensures that all weak signals that may reflect early fault symptoms can be captured, avoiding missed detections. Second, after being modulated by plasticity learning and attention mechanisms, the second threshold is used for secondary screening, which can effectively filter out false triggers caused by noise, ensuring that the final output of suspected fault labels has high confidence. Finally, the two-level threshold combined with the dynamic gating mechanism can achieve high specificity while maintaining high sensitivity, and is suitable for micro high-speed production line scenarios where fault samples are scarce and high-confidence prediction results are required.

[0066] This application transforms continuous waveform signals into sparse event sequences through a sparse representation model. Then, by combining plasticity learning guided by structural change data, it enables the automatic extraction of suspected fault markers related to structural changes from massive waveform data, providing a high-quality set of candidate events for subsequent anomaly screening and causal verification.

[0067] S104. The abnormal patterns of all the suspected fault markers are filtered by the isolated forest algorithm to obtain an abnormal event queue. Causal verification is performed on the abnormal event queue to obtain a verification report, and the verification report is fed back to the digital twin model.

[0068] In one specific implementation, such as Figure 2 As shown, step S104 includes: Step 1041: Based on the rate of change of each mechanical position in the structural change data, assign a dynamic weight coefficient to each mechanical position, and multiply the dynamic weight coefficient by the synaptic weight value after the signal is applied in the suspected fault mark of the corresponding mechanical position to obtain the weighted suspected fault mark.

[0069] In step 1041, the rate of change of the comprehensive change index of each mechanical position over time is first extracted from the structural change data, and the rate of change of each position is calculated. The higher the rate of change, the more rapidly the structural state at that position is deteriorating and the more attention it needs to be paid. Then, according to the magnitude of the rate of change, a dynamic weight coefficient between 0 and 1 is assigned to each mechanical position, and the faster the rate of change, the larger the weight coefficient. After that, the dynamic weight coefficient is multiplied by the synaptic weight value after the signal is applied to the corresponding mechanical position to obtain a weighted suspected fault label, so that the position with drastic structural changes receives greater weight in the subsequent anomaly screening.

[0070] Step 1042: Arrange all the weighted suspected fault markers in chronological order to form a cycle marker sequence for each production cycle. Use a random projection method to perform a dimensionality reduction transformation on the cycle marker sequence to obtain a low-dimensional feature vector. Input the low-dimensional feature vector into the isolated forest algorithm.

[0071] The specific type of random projection method can be sparse random projection. In step 1042, firstly, all weighted suspected fault markers generated in each production cycle are sorted according to the pulse arrival time to form a periodic marker sequence. Since the number of suspected fault markers may be different in each cycle, the sequence needs to be padded or truncated to a fixed length to make it a vector of uniform dimension. Then, sparse random projection is used to perform a dimensionality reduction transformation on this vector: a random projection matrix is ​​generated, and the matrix elements are randomly selected according to a specific distribution. Usually, a sparse random projection matrix is ​​used to reduce computational complexity. Then, the original high-dimensional vector is multiplied by the random projection matrix to obtain a low-dimensional feature vector. This low-dimensional feature vector retains the main structural information of the original data, while the dimensionality is greatly reduced, making it suitable for input into the isolated forest algorithm for processing.

[0072] Step 1043: Randomly select a feature dimension and a splitting value in each isolated tree, and recursively split the low-dimensional feature vector until each low-dimensional feature vector is split into a leaf node, and record the path length of each low-dimensional feature vector in each isolated tree.

[0073] In step 1043, the isolated forest algorithm consists of multiple isolated trees. For each isolated tree, a subset of samples is randomly selected from the training data as the training data for that tree. Then, a feature dimension is randomly selected from the subset of samples, and a split value is randomly selected between the minimum and maximum values ​​of the feature. The data point is then divided into left and right child nodes according to the split value. Subsequently, the above random splitting process is recursively performed on each child node until each data point is isolated to a leaf node, or the preset maximum tree depth is reached. The path length of each low-dimensional feature vector in the current isolated tree is recorded, that is, the number of edges traversed from the root node to the leaf node where the data point is located.

[0074] Step 1044: Calculate the average path length based on the path length of each of the low-dimensional feature vectors in all the isolated trees, and calculate the anomaly index of each of the periodic marker sequences based on the average path length.

[0075] In step 1044, for each periodic marker sequence, the path lengths of its corresponding low-dimensional feature vector in all isolated trees are summed and divided by the total number of isolated trees to obtain the average path length. Subsequently, since outliers are more easily isolated, their path lengths in isolated trees are shorter than a preset length threshold. Then, through... Calculate the anomaly index, where, An anomaly index representing a periodically labeled sequence, This represents the average path length of the low-dimensional feature vector corresponding to the sequence across all isolated trees. This represents the average path length of a binary search tree containing n data points. It is used to normalize the average path length. The closer the anomaly index is to 1, the more likely the periodic labeled sequence is to be anomalous; the closer the anomaly index is to 0, the more likely it is to be normal.

[0076] Step 1045: Retain the suspected fault markers in each production cycle whose abnormality index exceeds the preset abnormality threshold, and arrange the retained suspected fault markers according to the pulse arrival time to form an abnormal event queue.

[0077] In step 1045, firstly, an anomaly threshold is set according to business requirements, typically between 0.5 and 0.6. Then, for each production cycle, the anomaly index corresponding to its cycle marker sequence is compared with the anomaly threshold. If the anomaly index exceeds the anomaly threshold, an abnormal mode is considered to exist within that cycle, and all weighted suspected fault markers within that cycle are retained. If the anomaly index does not exceed the anomaly threshold, a normal mode is considered to exist within that cycle, and all suspected fault markers within that cycle are discarded. Finally, all retained suspected fault markers are sorted according to pulse arrival time to form an anomaly event queue, which serves as input for subsequent causal verification.

[0078] Step 1046: Using the mechanical positions corresponding to each suspected fault marker in the abnormal event queue as nodes, and the order of pulse arrival times between the mechanical positions as directed edges, construct a causal structure graph consisting of multiple nodes and multiple directed edges.

[0079] In step 1046, all suspected fault markers are extracted from the abnormal event queue, and each marker corresponds to a mechanical position and a pulse arrival time. Then, each mechanical position is regarded as a node in the graph. For any two mechanical positions C and D, if the pulse arrival time of position C is earlier than the pulse arrival time of position D, and the time difference between the two is within the preset propagation time window, then a directed edge from node C to node D is added to the graph, which indicates that the abnormality may have propagated from C to D. Finally, all suspected fault markers are traversed according to this rule to construct a complete causal structure graph.

[0080] Step 1047: Using counterfactual reasoning, perform intervention operations on each node in the causal structure graph to generate counterfactual samples corresponding to each node.

[0081] In step 1047, for each mechanical location node in the causal structure graph, an intervention operation is performed: assuming that no anomaly occurs at this location, while the anomaly states of other locations remain unchanged, the changes in the entire anomaly event queue under this intervention are recalculated through the propagation relationship in the causal structure graph to generate a new counterfactual sample that reflects how the entire anomaly propagation process would change if no anomaly occurred at a specific location.

[0082] Step 1048: Input the counterfactual samples and the original samples in the abnormal event queue into the Siamese network, calculate the similarity distance between the corresponding counterfactual samples and the original samples through the Siamese network, and calculate the contribution factor corresponding to each node based on the similarity distance.

[0083] Step 1048 may specifically include the following steps: Step a1: Input the counterfactual sample corresponding to each mechanical position into the first feature extraction branch of the Siamese network for full convolution processing to obtain the counterfactual feature vector.

[0084] The first feature extraction branch is the feature extractor in the Siamese network responsible for processing counterfactual samples, and it consists of multiple convolutional layers, pooling layers, and activation function layers.

[0085] In step a1, the first feature extraction branch of the Siamese network adopts a fully convolutional neural network structure, which contains 5 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The second and fourth layers also contain max pooling layers. The specific parameters of the convolutional layers are set as follows: the first convolutional layer has a kernel size of 11×11 and 96 output channels; the second convolutional layer has a kernel size of 5×5 and 256 output channels; the third convolutional layer has a kernel size of 3×3 and 384 output channels; the fourth convolutional layer has a kernel size of 3×3 and 384 output channels; and the fifth convolutional layer has a kernel size of 3×3 and 256 output channels. Then, the counterfactual samples are input into this branch, and after layer-by-layer convolution and pooling processing, a fixed-dimensional counterfactual feature vector is finally output.

[0086] Step a2: Input the original samples corresponding to the same mechanical position into the second feature extraction branch of the Siamese network for full convolution processing to obtain the original feature vector, wherein the first feature extraction branch and the second feature extraction branch share the network weights.

[0087] In step a2, the network structure of the second feature extraction branch is exactly the same as that of the first feature extraction branch. The parameters such as the number of convolutional layers, the size of the convolutional kernel, the number of output channels, and the position of the pooling layer are all the same. More importantly, the two branches share the same network weight parameters to ensure that the same feature extraction operation is performed on the two input samples. Then the original sample is input into the second feature extraction branch and undergoes the same layer-by-layer convolution and pooling processing to output the original feature vector.

[0088] Step a3: Input the counterfactual feature vector and the original feature vector into the computation branch of the Siamese network. The computation branch uses a cross-correlation calculation method to calculate the similarity distance between the counterfactual feature vector and the original feature vector.

[0089] In step a3, the computation branch receives counterfactual feature vectors and original feature vectors from two feature extraction branches, and calculates the similarity between the two vectors using a cross-correlation calculation method. Specifically, the counterfactual feature vector is used as a convolution kernel to perform a convolution operation on the original feature vector to obtain a cross-correlation response map. The higher the peak value in the response map, the more similar the two vectors are. Then, the mean of the cross-correlation response map is calculated as the similarity distance. Based on the magnitude of the similarity distance, the contribution factor corresponding to each node is obtained through normalization. The smaller the distance, the more similar the counterfactual sample is to the original sample, that is, the smaller the contribution of the node to the overall anomaly. The larger the distance, the greater the difference between the two, that is, the greater the contribution of the node to the overall anomaly. Furthermore, the larger the contribution factor, the more critical the role of the position in the anomaly propagation.

[0090] Step 1049: Based on all the contribution factors and the structural change data corresponding to each mechanical position, calculate the fault confidence level for each mechanical position, and mark the mechanical positions whose fault confidence level exceeds a preset confidence threshold as the root cause of the fault.

[0091] In step 1049, for each mechanical location, its contribution factor is weighted and fused with the structural change data of that location obtained in step S102 to obtain the fault confidence. The weighted fusion formula is as follows: ,in, Let represent the fault confidence level at the i-th mechanical position. This represents the contribution factor of the mechanical position. This represents the comprehensive change index at that location. This represents the maximum value of the comprehensive change index of all mechanical positions. c and d are preset weighting coefficients that satisfy c+d=1. Usually, c=0.6 and d=0.4 are chosen to highlight the importance of causal contribution. Then, the fault confidence of all mechanical positions is calculated and compared with the preset confidence threshold. Mechanical positions with fault confidence exceeding the confidence threshold are marked as the root cause of the fault.

[0092] Step 10410: Based on the fault root cause location and the corresponding downstream node in the causal structure diagram, generate a verification report containing the fault root cause location, fault confidence, and fault propagation path.

[0093] In step 10410, in the causal structure graph, starting from the root cause location of the fault, all downstream nodes are traversed along the directed edges to form a complete propagation path from the root cause to each affected location. Then, for each propagation path, the information of each mechanical location on the path and their temporal relationship are recorded, and the root cause location of the fault, its corresponding fault confidence, all propagation paths, and key node information on the path are integrated into a structured verification report. This report is output to the digital twin model to update the model's state assessment and optimize subsequent fault prediction.

[0094] This application uses the isolated forest algorithm to screen suspected fault markers for abnormal patterns, effectively filtering out normal fluctuations and noise interference, and obtaining a high-confidence abnormal event queue. Then, through the construction of causal structure graphs and counterfactual reasoning, it accurately isolates causal chains from the temporally mixed abnormal events to identify the true root cause of the fault and its propagation path, thereby enhancing the interpretability and credibility of the fault prediction results and providing a reliable decision-making basis for the precise maintenance of micro high-speed production lines.

[0095] Figure 3 This application provides a schematic diagram of a specific implementation of a digital twin-based real-time monitoring and fault prediction system for a micro production line. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire the baseline data of the micro production line at the time of acceptance, the real-time data of the micro production line during the operation period, and the continuous waveform signal of the mechanical skeleton.

[0096] The comparison module 32 is used to input the real-time data and the reference data into the digital twin model for comparison to obtain structural change data.

[0097] The transformation module 33 is used to perform sparse transformation on the continuous waveform signal using a sparse representation model to generate a sparse event sequence, and to perform plasticity learning on the sparse event sequence based on the structural change data to obtain multiple suspected fault markers.

[0098] The verification module 34 is used to filter all the suspected fault markers for abnormal patterns using the isolated forest algorithm to obtain an abnormal event queue, perform causal verification on the abnormal event queue, obtain a verification report, and feed the verification report back to the digital twin model.

[0099] The digital twin-based micro production line real-time monitoring and fault prediction system of this application embodiment is used to implement the aforementioned digital twin-based micro production line real-time monitoring and fault prediction method. Therefore, the specific implementation of the digital twin-based micro production line real-time monitoring and fault prediction system can be found in the embodiment section of the digital twin-based micro production line real-time monitoring and fault prediction method above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0100] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described digital twin-based micro production line real-time monitoring and fault prediction methods.

[0101] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for real-time monitoring and fault prediction of micro production lines based on digital twins.

[0102] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0103] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the digital twin-based micro-production line real-time monitoring and fault prediction method.

[0104] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0105] The foregoing has provided a detailed description of a method and system for real-time monitoring and fault prediction of a micro production line based on digital twins, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for real-time monitoring and fault prediction of a micro production line based on digital twins, characterized in that, include: Acquire baseline data of the micro production line at the time of acceptance, real-time data of the micro production line during operation, and continuous waveform signals of the mechanical frame; The real-time data and the baseline data are input into the digital twin model for comparison to obtain structural change data; The continuous waveform signal is sparsely transformed using a sparse representation model to generate a sparse event sequence. Based on the structural change data, the sparse event sequence is subjected to plasticity learning to obtain multiple suspected fault markers. The isolated forest algorithm is used to filter all the suspected fault markers for abnormal patterns to obtain an abnormal event queue. Causal verification is performed on the abnormal event queue to obtain a verification report, and the verification report is fed back into the digital twin model.

2. The method according to claim 1, characterized in that, The causal verification of the abnormal event queue, resulting in a verification report, includes: Using the mechanical positions corresponding to each suspected fault marker in the abnormal event queue as nodes, and the order of pulse arrival times between the mechanical positions as directed edges, a causal structure graph consisting of multiple nodes and multiple directed edges is constructed. Using a counterfactual reasoning method, intervention operations are performed on each node in the causal structure graph to generate a counterfactual sample corresponding to each node; The counterfactual samples and the original samples in the abnormal event queue are input into the Siamese network. The similarity distance between the corresponding counterfactual samples and the original samples is calculated through the Siamese network. Based on the similarity distance, the contribution factor corresponding to each node is calculated. Based on all the contribution factors and the structural change data corresponding to each mechanical position, calculate the fault confidence level for each mechanical position, and mark the mechanical positions whose fault confidence level exceeds a preset confidence threshold as the root cause of the fault. Based on the location of the fault root cause and the corresponding downstream node in the cause-effect diagram, a verification report is generated that includes the location of the fault root cause, the fault confidence level, and the fault propagation path.

3. The method according to claim 2, characterized in that, The step of inputting the counterfactual samples and the original samples in the anomalous event queue into a Siamese network, and calculating the similarity distance between the corresponding counterfactual samples and the original samples through the Siamese network, includes: The counterfactual sample corresponding to each mechanical position is input into the first feature extraction branch of the Siamese network for full convolution processing to obtain the counterfactual feature vector; The original samples corresponding to the same mechanical position are input into the second feature extraction branch of the twin network for full convolution processing to obtain the original feature vector, wherein the first feature extraction branch and the second feature extraction branch share the network weights. The counterfactual feature vector and the original feature vector are input into the computational branch of the Siamese network, and the computational branch uses a cross-correlation calculation method to calculate the similarity distance between the counterfactual feature vector and the original feature vector.

4. The method according to claim 1, characterized in that, The step of performing a sparse transformation on the continuous waveform signal using a sparse representation model to generate a sparse event sequence includes: Multiple waveform segments corresponding to the mechanical positions indicated by the structural change data are extracted from the continuous waveform signal, and all the waveform segments are used as a training sample set. The singular value decomposition algorithm is used to iteratively learn the training sample set to construct an overcomplete redundant dictionary. Using the overcomplete redundant dictionary as a sparse representation model, the orthogonal matching pursuit algorithm is used to project and decompose the subsequently acquired continuous waveform signals onto the sparse representation model to obtain a sparse coefficient sequence. Extract the atom index, pulse arrival time, and atom amplitude corresponding to each non-zero projection coefficient from the sparse coefficient sequence, and combine the atom index, the pulse arrival time, and the atom amplitude to form a sparse event sequence.

5. The method according to claim 1, characterized in that, The process involves performing plasticity learning on the sparse event sequence based on the structural change data to obtain multiple suspected fault markers, including: Assign a cluster of spiking neurons to the mechanical location indicated by the structural change data; The pulse time-dependent plasticity algorithm is used to calculate the arrival time difference between presynaptic pulses and postsynaptic pulses in each spiking neuron cluster. Based on the arrival time difference, the weight adjustment direction of long-term enhancement or long-term inhibition is determined, and the synaptic weights of the spiking neuron cluster are updated according to the weight adjustment direction. An attention mechanism is used to calculate the attention weights of each mechanical position, and the synaptic weights are modulated according to the attention weights to obtain modulated synaptic weights; The modulated synaptic weights are compared with a preset first activation threshold. When the comparison result indicates that the modulated synaptic weights exceed the first activation threshold, the mechanical position corresponding to the current spiking neuron cluster is recorded as the activation position, and the activation position and the corresponding pulse arrival time are combined as a candidate fault marker. The amplitude of the synaptic weights corresponding to the candidate fault markers is compressed to obtain normalized synaptic weights. A dynamic gating mechanism is used to apply periodic gating signals to the normalized synaptic weights to obtain synaptic weights after the signals are applied. The synaptic weights after the applied signal are compared with a preset second activation threshold. When the result of the second comparison indicates that the synaptic weights after the applied signal exceed the second activation threshold, the mechanical position corresponding to the current spiking neuron cluster is recorded as a suspected fault position, and the suspected fault position and the corresponding pulse arrival time are combined to form a suspected fault marker.

6. The method according to claim 1, characterized in that, The step of filtering all suspected fault markers for abnormal patterns using the isolated forest algorithm to obtain an abnormal event queue includes: Based on the rate of change of each mechanical position in the structural change data, a dynamic weight coefficient is assigned to each mechanical position. The dynamic weight coefficient is multiplied by the synaptic weight value after the signal is applied in the suspected fault mark of the corresponding mechanical position to obtain the weighted suspected fault mark. All the weighted suspected fault markers are arranged in chronological order to form a cycle marker sequence for each production cycle. The cycle marker sequence is then subjected to dimensionality reduction transformation using a random projection method to obtain a low-dimensional feature vector. The low-dimensional feature vector is then input into the isolated forest algorithm. In each isolated tree, a feature dimension and a splitting value are randomly selected, and the low-dimensional feature vector is recursively split until each low-dimensional feature vector is split into a leaf node. The path length of each low-dimensional feature vector in each isolated tree is recorded. Calculate the average path length based on the path length of each of the low-dimensional feature vectors in all the isolated trees, and calculate the anomaly index of each of the periodic marker sequences based on the average path length; The suspected fault markers that exceed the preset abnormal threshold in each production cycle are retained, and the retained suspected fault markers are arranged according to the pulse arrival time to form an abnormal event queue.

7. The method according to claim 1, characterized in that, The step of comparing the real-time data with the baseline data in the digital twin model to obtain structural change data includes: The peak frequency corresponding to each mechanical position is extracted from the real-time data. The difference between the peak frequency and the reference peak frequency of the corresponding mechanical position in the reference data is calculated by the comparison engine of the digital twin model to obtain the frequency offset of each mechanical position. The amplitude attenuation rate corresponding to each mechanical position is extracted from the real-time data. The amplitude attenuation rate is then compared with the reference attenuation rate of the corresponding mechanical position in the reference data using the comparison engine to obtain the amplitude change coefficient of each mechanical position. The frequency offset and amplitude change coefficient of each mechanical position are weighted and combined using the calculation engine of the digital twin model to obtain the comprehensive change index of each mechanical position. Using the digital twin model's combination engine, all the aforementioned comprehensive change indices are arranged according to their mechanical positions to form structural change data.

8. A real-time monitoring and fault prediction system for a micro production line based on digital twins, characterized in that, include: The acquisition module is used to acquire the baseline data of the micro production line at the time of acceptance, the real-time data of the micro production line during the operation period, and the continuous waveform signal of the mechanical frame. The comparison module is used to input the real-time data and the reference data into the digital twin model for comparison to obtain structural change data; The transformation module is used to perform sparse transformation on the continuous waveform signal using a sparse representation model to generate a sparse event sequence. Based on the structural change data, the sparse event sequence is subjected to plasticity learning to obtain multiple suspected fault markers. The verification module is used to filter all the suspected fault markers for abnormal patterns using the isolated forest algorithm to obtain an abnormal event queue, perform causal verification on the abnormal event queue, obtain a verification report, and feed the verification report back to the digital twin model.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the method for real-time monitoring and fault prediction of a micro production line based on digital twins as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the method for real-time monitoring and fault prediction of a micro production line based on digital twins as described in any one of claims 1 to 7.

Images