Real-time monitoring system for manufacturing workshop based on internet of things

Abstract

This invention discloses a real-time monitoring system for manufacturing workshops based on the Internet of Things (IoT), belonging to the field of intelligent manufacturing monitoring technology. It includes a signal processing module, a framework construction module, a situation map module, a health assessment module, and a safety assessment module. The system captures multimodal sensor signals through an IoT sensing layer and decomposes them into signal morphology sets corresponding to different physical processes. These components are projected onto a virtual monitoring framework composed of interconnected topological units, assigned to corresponding units according to source attributes to form state vectors, and iteratively updated topological connection weights to generate a workshop operation situation map reflecting the overall operational status of the workshop. Based on this map, the load-response relationship model of key structural points is reconstructed, dynamic health indicators are derived, and finally, the system-level safety margin is calculated. This invention achieves a leap from multi-source heterogeneous data to deep perception and quantitative assessment of the workshop system-level safety status.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing monitoring technology, specifically a real-time monitoring system for manufacturing workshops based on the Internet of Things. Background Technology

[0002] Current real-time monitoring in manufacturing workshops primarily relies on deploying various IoT sensors to collect and centrally display signals such as vibration, temperature, and current, and setting static thresholds for out-of-limit alarms. Existing technologies typically process or simply fuse these multimodal sensor signals, lacking in-depth analysis of the specific physical mechanisms behind the signals. Monitoring models are mostly digital twins based on fixed rules or preset three-dimensional scenes, with static and predefined internal logical relationships. This approach treats the workshop as a collection of discrete monitoring points, failing to effectively depict the real-time dynamic interactions between multiple elements such as equipment, processes, and material flows.

[0003] This technical approach has limitations. Directly processing raw, mixed signals cannot decompose them and accurately map them to independent physical processes such as cutting, transmission, and heat exchange. This makes it difficult to trace abnormal states back to their specific physical roots, and the assessment remains superficial. Static monitoring models or views cannot dynamically adjust the logical connections and influence weights between their internal monitoring units based on real-time operational data, and therefore cannot characterize how local state changes affect the whole system through system coupling relationships. This makes it difficult for existing systems to achieve accurate fault tracing, and even more difficult to conduct forward-looking quantitative assessments of the overall operational safety status of the workshop.

[0004] The technical challenges that need to be addressed are: how to separate the state components representing independent physical processes from mixed multimodal sensing data, and how to construct a monitoring model that can dynamically evolve its internal correlation structure with real-time data, thereby achieving in-depth perception and quantitative assessment of the system-level operational safety of the manufacturing workshop. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a real-time monitoring system for manufacturing workshops based on the Internet of Things, comprising: The signal processing module is used to continuously capture multimodal sensing signals through the IoT sensing layer deployed in the manufacturing workshop, and to perform signal morphology decomposition on the multimodal sensing signals to obtain a set of component signals corresponding to different physical processes. The framework construction module is used to project the component signal set onto a preset virtual monitoring framework. The virtual monitoring framework is composed of multiple interconnected topological units. During the projection process, the component signals are assigned to the corresponding topological units according to their source attributes to form a topological unit state vector. The topological connection weights of the virtual monitoring framework are iteratively updated using the topological unit state vector. The dynamic health assessment module is used to reconstruct the load-response relationship model of key structural points in the manufacturing workshop based on the workshop operation status map, and to deduce the dynamic health indicators of each key structural point using the load-response relationship model. The safety assessment module aggregates dynamic health indicators for all critical structural points to calculate the system-level safety margin of the manufacturing workshop.

[0006] Preferably, the step of performing signal morphology decomposition on the multimodal sensing signal to obtain a set of component signals corresponding to different physical processes includes: A basis function library for signal decomposition is established, which contains basis function templates characterizing vibration, heat conduction, fluid motion, and electromagnetic interference; For each captured multimodal sensing signal, adaptive matching and tracking are performed in the basis function library to select multiple basis function templates whose matching degree with the modal sensing signal exceeds a preset threshold. The multimodal sensing signal is synchronously expanded using the selected basis function templates to obtain a set of initial decomposition coefficients; The initial decomposition coefficients are optimized by sparsity constraints, and redundant decomposition components are removed to obtain the optimized set of decomposition coefficients. Based on the optimized set of decomposition coefficients and their corresponding basis function templates, multiple independent signal components are reconstructed; The reconstructed signal components are traced back to their physical processes. Based on their energy distribution characteristics and waveform properties, they are classified into mechanical vibration process components, thermodynamic process components, material transport process components, or electromagnetic compatibility process components, ultimately forming the set of component signals.

[0007] Preferably, projecting the component signal set onto a preset virtual monitoring frame includes: The virtual monitoring framework is defined as a multi-dimensional state space, where each dimension of the multi-dimensional state space corresponds to a process element, and each topological unit represents a coordinate point in the multi-dimensional state space. For each signal component in the set of component signals, calculate its characteristic coordinates in the multidimensional state space. The calculation process combines the physical attribute labels and time-frequency characteristics of the signal components. Signal components with similar characteristic coordinates are clustered to form signal component clusters; The center coordinates of each signal component cluster are mapped to the nearest topological unit in the virtual monitoring framework, thus establishing a binding relationship between the signal component cluster and the topological unit; Based on the binding relationship, the energy amplitude, frequency band information and change trend of each signal component cluster are encoded into the topology unit state vector, and each element of the topology unit state vector corresponds to the current state code of a topology unit.

[0008] Preferably, the step of iteratively updating the topology connection weights of the virtual monitoring framework using the topology unit state vector includes: Initialize the connection weights between all topological units in the virtual monitoring framework; Calculate the difference vector between the current topological unit state vector and the previous topological unit state vector; Based on the direction and magnitude of the difference vector, the set of active topological units whose states have changed significantly is identified; Within the virtual monitoring framework, neighboring topological units that have a physical or logical association with the set of active topological units are identified. Based on the intensity of the state change of the active topology unit and its preset coupling relationship with the neighboring topology units, the connection weight from the active topology unit to the neighboring topology unit is dynamically adjusted to enhance or weaken the connection strength. After each weight adjustment, check whether the overall connection graph of the virtual monitoring framework meets the preset connectivity constraints. If it does not meet the constraints, fine-tune the local weights to ensure global connectivity. Repeat the process from calculating the difference vector to checking and fine-tuning connectivity until the change in the topological unit state vector tends to stabilize or reaches the preset number of iterations, thereby completing the update of the topological connection weights.

[0009] Preferably, based on the updated topology connection weights, activating the critical propagation path in the virtual monitoring framework includes: In the updated virtual monitoring framework, one or more state initiation units are set; Starting from the state initiation unit, a depth-first state propagation simulation is performed along the topological connections; During the state propagation simulation, the attenuation degree and propagation delay of state information as it passes through each connection are recorded. Based on the recorded attenuation level and propagation delay, calculate the cumulative propagation loss from the initial state unit to other topological units; All paths with cumulative propagation loss below a set threshold are selected as candidate propagation paths; Redundancy elimination processing is performed on the candidate propagation paths to remove paths that are spatially overlapping or functionally equivalent. The remaining paths after redundancy elimination are marked as critical propagation paths, and a path identifier is assigned to each critical propagation path.

[0010] Preferably, the topological unit state vectors are subjected to temporal convolution along the key propagation path to generate a workshop operation status map, including: For each marked critical propagation path, extract the sequence of topological units traversed by the critical propagation path according to the path identifier; Obtain the topology unit state vector fragments of the topology unit sequence over multiple consecutive time slices; Design a convolutional kernel suitable for the critical propagation path, the length of which matches the number of topological units in the critical propagation path; Using temporal convolution operations, the convolution kernel is slid along the time dimension to perform convolution calculations on the topological unit state vector fragments; During the convolution calculation, the temporal state correlation of multiple topological units on the fusion path is fused to output a one-dimensional feature sequence that reflects the overall situational change of the key propagation path. Repeat the convolution calculation process from extracting the topological unit sequence to outputting the one-dimensional feature sequence for all critical propagation paths to obtain a set of feature sequences equal to the number of critical propagation paths; All the obtained feature sequences are spatially aligned and concatenated to form a multidimensional feature matrix. The multidimensional feature matrix is ​​reduced in dimensionality and visualized by encoding, and then transformed into a graph structure containing a time axis, a path axis, and an intensity axis, which is the workshop operation status graph.

[0011] Preferably, the step of reconstructing the load-response relationship model of key structural points in the manufacturing workshop based on the workshop operation status map includes: The spatial locations of all key structural points are marked in the three-dimensional digital model of the manufacturing workshop; From the workshop operation status map, extract the local map region features associated with the spatial location of each key structural point; Analyze the patterns of the local map region features at different time points to identify feature change patterns related to load application and feature change patterns related to structural response; Establish a preliminary mapping relationship, quantize the load-related characteristic change patterns into equivalent load vectors, and quantize the response-related characteristic change patterns into structural response vectors; Using a system identification method, a dynamic transfer function is fitted based on the correspondence between the equivalent load vector and the structural response vector in historical data. By combining the dynamic transfer function with the known physical parameters of key structural points, a mathematical relationship describing the transition from load input to structural response output is constructed. This mathematical relationship is the load-response relationship model.

[0012] Preferably, the dynamic health indicators of each key structural point derived using the load-response relationship model include: The equivalent load vector at the current moment is obtained in real time and used as the input to the load-response relationship model; The theoretical structural response vector of the key structural point under the current load is calculated using the load-response relationship model. Meanwhile, the actual observed response vectors of the key structural points are obtained by directly measuring them through sensing devices or by inverting them from the workshop operation status map; Calculate the residual vector between the theoretical structural response vector and the actual observed response vector; Statistical analysis is performed on the residual vector to calculate its norm and covariance, thereby obtaining the response deviation. By combining the material fatigue characteristic curves and historical load spectra of the key structural points, the contribution of the current response deviation to the cumulative structural damage is evaluated. Based on the cumulative damage contribution and the preset safety threshold, a value between zero and one is calculated as the current dynamic health indicator of the key structural point. The lower the dynamic health indicator value, the worse the health status.

[0013] Preferably, the step of calculating the cumulative propagation loss from the initial state unit to other topological units based on the recorded attenuation level and propagation delay includes: Extract the attenuation factor and propagation delay value for each connection from the records of the state propagation simulation; For each state starting unit, all other topological units in the virtual monitoring framework are traversed as target units; For each target unit, enumerate all possible paths from the state starting unit to the target unit, and each path consists of a series of topological connection sequences; For each path, the attenuation factor of all connections on the path is multiplied to obtain the total attenuation factor of the path, and the propagation delay values ​​of all connections on the path are summed to obtain the total propagation delay of the path. Using a preset attenuation-delay joint cost function, the total attenuation factor and total propagation delay are mapped to a scalar cost, which serves as the propagation loss of the path. For each target unit, the value with the minimum propagation loss is selected from all enumerated paths and used as the cumulative propagation loss from the state starting unit to the target unit.

[0014] Preferably, calculating the residual vector between the theoretical structural response vector and the actual observed response vector includes: Align the dimensions of the theoretical structural response vector with those of the actual observed response vector to ensure that the two vectors contain the same number of response feature components in the same order. Timestamp matching is performed on the two vectors to ensure that the calculation time point of the theoretical structural response vector corresponds precisely to the sampling time point of the actual observed response vector; Subtract the measured value of the corresponding component in the actual observed response vector from the value of each response feature component in the theoretical structural response vector to obtain the original difference of each component. Divide the original difference of each component by the corresponding observation error tolerance or normalization coefficient to obtain the normalized difference; Arrange all normalized differences in their original order to form the residual vector.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Signal morphology decomposition is performed on multimodal sensing signals to obtain a set of component signals that strictly correspond to different physical processes. This process separates the original mixed observation data into pure components that characterize independent physical mechanisms. It achieves a direct correlation between monitoring information and physical essence, enabling state assessment to pinpoint specific physical processes and overcoming the problems of ambiguous root cause identification and diagnostic lag caused by signal coupling in traditional methods.

[0016] Component signals are mapped to a virtual monitoring framework composed of interconnected topological units based on their source attributes, forming unit state vectors and iteratively updating the topological connection weights. This process constructs a monitoring structure whose internal logical relationships can dynamically evolve with the data. The continuous iteration of connection weights enables the framework to automatically quantify the strength and direction of the mutual influence between units, thus depicting the dynamic interaction network within the system in real time. The monitoring perspective thus shifts from a static set of discrete points to a continuous characterization of the overall coupling relationships and state propagation paths of the system.

[0017] The load-response model of key structural points is reconstructed based on dynamic topology relationships, and system-level safety margin is calculated by aggregating health indicators. The load-response model is established based on real-time situational data, reflecting the behavioral characteristics of local nodes under the current coupled state of the system. By aggregating the dynamic health indicators of each node, the system-level safety margin obtained through comprehensive evaluation no longer relies on isolated threshold judgments, but quantitatively expresses the boundary of the workshop's overall ability to resist disturbances and maintain stable operation under the existing operational correlation mode. Attached Figure Description

[0018] Figure 1 This is a timing diagram of the IoT-based real-time monitoring system for manufacturing workshops as described in this invention. Figure 2 A flowchart for signal morphology decomposition; Figure 3A flowchart for projecting component signals; Figure 4 A trend chart showing the dynamic changes in topology connection weights within a virtual monitoring framework for a manufacturing workshop; Figure 5 A bar chart analyzing the contribution of damage to critical structural points in the manufacturing workshop. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This paper describes the implementation of a real-time monitoring system for a manufacturing workshop based on the Internet of Things (IoT). The system continuously captures multimodal sensor signals through an IoT sensing layer deployed within the workshop. A signal processing module decomposes the multimodal sensor signals into signal morphology sets corresponding to different physical processes. A framework construction module projects these signal sets onto a pre-defined virtual monitoring framework, which consists of multiple interconnected topological units. During projection, each signal component is assigned to a corresponding topological unit based on its source attributes, forming a topological unit state vector. The topological connection weights of the virtual monitoring framework are iteratively updated using these topological unit state vectors. Subsequently, based on the updated topological connection weights, critical propagation paths are activated within the virtual monitoring framework. Temporal convolution is performed on the topological unit state vectors along these critical propagation paths to generate a workshop operation status map. A health assessment module reconstructs a load-response relationship model for key structural points within the manufacturing workshop based on the workshop operation status map and uses this model to deduce dynamic health indicators for each key structural point. A safety assessment module aggregates the dynamic health indicators of all key structural points to calculate the system-level safety margin of the manufacturing workshop.

[0021] In one embodiment of the present invention, see [reference] Figure 2The signal processing module performs signal morphology decomposition on the multimodal sensing signals continuously captured by the IoT sensing layer in the manufacturing workshop to obtain component signal sets corresponding to different physical processes. The process begins with establishing a basis function library for signal decomposition. This library contains basis function templates characterizing vibration, heat conduction, fluid motion, and electromagnetic interference. These templates are predefined sets of mathematical functions used to match and characterize the signal morphologies generated by different physical processes. In practice, for each captured multimodal sensing signal, the signal processing module performs adaptive matching pursuit in the basis function library to filter out multiple basis function templates whose matching degree with the modal sensing signal exceeds a preset threshold. Adaptive matching pursuit evaluates the matching degree by calculating the inner product or correlation coefficient between the multimodal sensing signal and each basis function template in the library, and retains the basis function templates exceeding the fixed threshold as candidate templates. In some embodiments, the multimodal sensing signal is synchronously expanded using the selected basis function templates to obtain a set of initial decomposition coefficients. The synchronous expansion operation represents the multimodal sensing signal as a linear combination of all selected basis function templates, and the weight coefficients corresponding to each basis function template in the linear combination are solved by the least squares method. These weight coefficients constitute the initial decomposition coefficients.

[0022] Optionally, sparsity constraint optimization is performed on the initial decomposition coefficients to remove redundant decomposition components and obtain an optimized set of decomposition coefficients. Sparsity constraint optimization is achieved by solving an optimization problem whose objective is to minimize the number of non-zero coefficients in the initial decomposition coefficients while ensuring signal reconstruction accuracy. A usable mathematical expression is minimizing the objective function. :

[0023] Where: vector The matrix represents one captured multimodal sensor signal. Each column represents a selected basis function template, a vector. The scalar represents the set of optimized decomposition coefficients to be determined. It is a regularization parameter that controls the sparsity strength, with the symbol... The L2 norm of a vector is represented by the symbol . Let L1 represent the norm of a vector. By solving the above optimization problem through an iterative algorithm, we finally obtain an optimized set of decomposition coefficients in which most elements are zero or close to zero.

[0024] In some embodiments, multiple independent signal components are reconstructed based on the optimized set of decomposition coefficients and their corresponding basis function templates. The reconstruction process multiplies each non-zero coefficient in the optimized set of decomposition coefficients with its corresponding basis function template, and sums all the product results to approximate the original multimodal sensing signal. Each term in the summation is considered an independent signal component. It can be understood that physical process tracing is performed on each reconstructed signal component to ultimately form a component signal set. Physical process tracing classifies the signal components according to their energy distribution characteristics and waveform properties, classifying them as mechanical vibration process components, thermodynamic process components, material transport process components, or electromagnetic compatibility process components. For example, a signal component whose energy is mainly concentrated in the high-frequency band and whose waveform exhibits periodic impacts is classified as a mechanical vibration process component, while a signal component whose energy changes slowly and whose trend is consistent with the temperature sensor reading is classified as a thermodynamic process component. In specific implementations, all signal components that have undergone physical process tracing and classification together constitute a component signal set, which serves as the input for subsequent processing.

[0025] In one embodiment of the present invention, see [reference] Figure 3 The framework construction module projects the component signal set onto a preset virtual monitoring framework. The virtual monitoring framework is defined as a multi-dimensional state space, where each dimension corresponds to a process element, including equipment vibration intensity, regional temperature gradient, material flow rate, or electromagnetic field intensity. Each topological unit represents a coordinate point in the multi-dimensional state space, uniquely determined by a set of coordinate values ​​across all dimensions. The module calculates the feature coordinates of each signal component in the multi-dimensional state space. This calculation combines the physical attribute labels and time-frequency features of the signal components. The physical attribute labels are derived from the classification results during signal morphology decomposition, while the time-frequency features are extracted from the signal components using short-time Fourier transform or wavelet transform. In some embodiments, signal components with similar feature coordinates are clustered to form signal component clusters. The clustering operation employs a k-means algorithm based on Euclidean distance, grouping signal components with spatially similar feature coordinates into the same signal component cluster. The center coordinates of each signal component cluster are mapped to the nearest topological unit in the virtual monitoring frame to establish the binding relationship between the signal component cluster and the topological unit. The nearest topological unit is found by calculating the Euclidean distance between the center coordinates of the signal component cluster and the coordinates of all topological units and selecting the topological unit corresponding to the minimum value.

[0026] In practical implementation, the energy amplitude, frequency band information, and trend of each signal component cluster are encoded into a topology unit state vector based on the binding relationship. Each element of the topology unit state vector corresponds to the current state code of a topology unit. The state code is a multi-dimensional vector containing the normalized energy amplitude, the center frequency of the dominant frequency band, and the slope of the signal trend within the recent time window. The topology connection weights of the virtual monitoring framework are iteratively updated using the topology unit state vectors. The implementation process begins by initializing the connection weights between all topology units in the virtual monitoring framework. The initial value of the connection weight can be set to a small constant or based on the physical distance between the topology units. The difference vector between the current topology unit state vector and the previous topology unit state vector is calculated. Each element of the difference vector corresponds to the state change of a topology unit. The set of active topology units with significant state changes is identified based on the direction and magnitude of the difference vector. The identification rule is that the absolute value of the corresponding element in the difference vector exceeds a preset activation threshold.

[0027] Optionally, the virtual monitoring framework identifies neighboring topological units that have physical or logical associations with the active topological unit set. Physical associations are determined based on the actual connection relationships of equipment in the manufacturing workshop layout diagram, while logical associations are determined based on the information flow or material flow relationships in the process flow diagram. The connection weights from active topological units to neighboring topological units are dynamically adjusted based on the intensity of state changes in active topological units and their preset coupling relationships with neighboring topological units, strengthening or weakening the connection strength. The dynamic adjustment process follows these update rules:

[0028] Where: symbol Represents active topology units Pointing to neighbor topology cells The updated connection weights, symbols Represents the connection weights before adjustment, a scalar. It is a preset learning rate coefficient used to control the step size of weight adjustment, function With active topology units state change The input and output is a scalar value representing the intensity of the state change, with the symbol... Represents active topology units With neighboring topology units The preset coupling coefficients between them. It can be understood that after each weight adjustment, it is necessary to check whether the overall connection graph of the virtual monitoring framework meets the preset connectivity constraints. The connectivity constraints require that at least one core topological unit can be reached from any topological unit through a directed connection path. If this is not met, the local weights are fine-tuned to ensure global connectivity. Fine-tuning is achieved by increasing or decreasing the minimum weight value of a specific connection. The process from calculating the difference vector to checking and fine-tuning connectivity is repeated until the change in the topological unit state vector tends to stabilize or the preset number of iterations is reached, thus completing the update of the topological connection weights.

[0029] In some embodiments, the activation threshold used to identify the set of active topological units can be dynamically adjusted according to different stages of workshop operation, for example, a lower activation threshold can be used during the equipment startup phase to capture more subtle state changes. It is understood that a preset coupling coefficient... The value is assigned based on the strength of the causal relationship between the process elements represented by the topological unit. The strength of the causal relationship is obtained by analyzing the mutual information or Granger causality in historical operating data.

[0030] In one embodiment of the invention, critical propagation paths are activated in a virtual monitoring framework based on updated topology connection weights. One or more state-starting units are set in the updated virtual monitoring framework. These state-starting units are designated as topology units reflecting the core equipment or process starting point of a manufacturing workshop, for example, a topology unit representing "CNC machine tool spindle drive" is set as the state-starting unit. A depth-first state propagation simulation is performed along the topology connections starting from the state-starting unit. The state propagation simulation uses the state-starting unit as the root node and recursively visits all reachable neighboring topology units in the direction of the upward connections within the virtual monitoring framework. During the state propagation simulation, the attenuation level and propagation delay of state information as it passes through each connection are recorded. The attenuation level is derived from the connection weights, and the propagation delay is set based on the causal delay between the physical processes represented by the topology units at both ends of the connection. The cumulative propagation loss from the state-starting unit to other topology units is calculated based on the recorded attenuation level and propagation delay. The attenuation factor and propagation delay value for each connection are extracted from the records of the state propagation simulation. The attenuation factor is a value between 0 and 1, and the propagation delay value is in milliseconds. For each state-starting unit, all other topology units in the virtual monitoring framework are traversed as target units. For each target unit, all possible paths from the starting state unit to the target unit are enumerated. Each path consists of a series of topological connections. For each path, the attenuation factors of all connections on the path are multiplied to obtain the total attenuation factor of the path, and the propagation delay values ​​of all connections on the path are summed to obtain the total propagation delay of the path. A preset attenuation-delay joint cost function is used to map the total attenuation factor and the total propagation delay into a scalar cost as the propagation loss of the path. A specific form of the cost function is as follows:

[0031] Where: symbol Represents the calculated path propagation loss, symbol The total attenuation factor of the path, symbol The total propagation delay of the path, coefficient and These are preset weighting parameters used to balance the contributions of the attenuation factor and propagation delay to the cost. For each target unit, the value with the minimum propagation loss is selected from all enumerated paths as the cumulative propagation loss from the starting state unit to the target unit. All paths with cumulative propagation loss below a set threshold are selected as candidate propagation paths. The threshold is, for example, [value missing]. All propagation losses Paths with a propagation loss less than 5.0 are selected. Redundancy elimination is performed on candidate propagation paths to remove spatially overlapping or functionally equivalent paths. For example, two paths that pass through different topological unit sequences but ultimately lead to the same critical process parameter change are considered functionally equivalent, and only the path with the lower propagation loss is retained. The remaining paths after redundancy elimination are marked as critical propagation paths, and each critical propagation path is assigned a path identifier in the format "KP_Starting Unit Number_Sequence Number".

[0032] In some embodiments, the setting of the state starting unit can be dynamically adjusted based on real-time alarm information. When the state code of a certain topology unit exceeds a safety threshold, the unit is automatically added as a new state starting unit to activate a new propagation path analysis. It can be understood that the attenuation factor can be calculated based on the connection weight value through a monotonically decreasing function. In specific implementations, the depth-first state propagation simulation sets a maximum search depth to prevent infinite recursive searches within a large virtual monitoring framework. Optionally, when enumerating all possible paths from the state starting unit to the target unit, a k-shortest path algorithm is used to find the first k paths with the smallest propagation loss for subsequent evaluation. In some embodiments, the weight parameters... and The value can be obtained by optimizing historical operating data so that the selection results of critical propagation paths are consistent with the actual failure propagation chains that have occurred in history.

[0033] In one embodiment of the present invention, temporal convolution is performed on the topology unit state vectors along the critical propagation path to generate a workshop operation status map. For each marked critical propagation path, the sequence of topology units traversed by the critical propagation path is extracted according to the path identifier. For example, for the critical propagation path with the path identifier "KP_S1_03", the sequence of topology units traversed is extracted as [unit A, unit B, unit F, unit J]. Topology unit state vector segments of the topology unit sequence are obtained in multiple consecutive time slices. The time slice can be one segment per second, and the state vector segment of each topology unit is a vector sequence arranged in chronological order. A convolution kernel suitable for the critical propagation path is designed. The length of the convolution kernel matches the number of topology units in the critical propagation path. If the critical propagation path contains 4 topology units, the convolution kernel is designed as a weight vector of length 4, for example, the weight vector is [0.1, 0.4, 0.4, 0.1]. Temporal convolution is used to perform convolution calculations on the state vector segments of the topological unit by sliding the convolution kernel along the time dimension. At each time point, the weight of the convolution kernel is summed with the current and historical state vector values ​​of the corresponding topological unit.

[0034] During convolution computation, the temporal state correlations of multiple topological units along the propagation path are fused, and a one-dimensional feature sequence reflecting the overall situational changes of the key propagation path is output. The specific form of the convolution operation can be represented as follows:

[0035] Where: symbol Indicates a point in time The one-dimensional feature sequence value of the critical propagation path output, symbol Indicates the number of topological units contained in the critical propagation path, symbol [symbol missing]. The vector represents the length of the past time window considered by the convolution kernel in the time dimension. The weight vector representing the convolution kernel, its 6th weight... element Corresponding to the first on the path The weights of each topological unit, the function Return to the path The index of a topological unit, symbol Indicates a point in time Topology The state vector values ​​are obtained. The convolution process, from extracting the topological unit sequence to outputting a one-dimensional feature sequence, is repeated for all critical propagation paths to obtain a set of feature sequences equal in number to the number of critical propagation paths. All obtained feature sequences are spatially aligned and concatenated to form a multi-dimensional feature matrix. Spatial alignment ensures that all feature sequences have the same time length and sampling points, while feature concatenation treats each feature sequence as a row or column of the matrix. The multi-dimensional feature matrix is ​​then dimensionality-reduced and visualized, transforming it into a graph structure containing time, path, and intensity axes—a workshop operation status graph. Dimensionality reduction can be achieved using principal component analysis to extract the main feature dimensions, and visualization encoding maps the feature values ​​to color or brightness for display in a three-dimensional coordinate system.

[0036] In some embodiments, a sliding time window mechanism is used when acquiring topological unit state vector fragments. The window length is 60 seconds, and the sliding step is 1 second, thereby obtaining a continuous sequence of fragments. It can be understood that the values ​​of the convolutional kernel weight vectors can be assigned based on the importance of the topological units in the path, and the importance can be determined by the frequency of occurrence of the topological units in historical faults. In specific implementations, the time window length... The time scale can be set according to the characteristics of the monitored physical process, for slowly changing thermal processes. It can be set to 300 for rapidly changing vibration processes. It can be set to 30. Optionally, the color mapping used for visual encoding adopts a gradient of "blue-green-yellow-red", where blue represents low-intensity feature values ​​and red represents high-intensity feature values. In some embodiments, the topological unit state vector fragment needs to be normalized before input convolution calculation to eliminate the influence of different physical dimensions. The normalization adopts the z-score method. Refer to Table 1, which shows a simplified topological unit state vector fragment.

[0037] Table 1: Time Series Table of State Vector Fragments of Topological Units

[0038] It can be understood that the dimension of the multidimensional feature matrix formed by feature concatenation is... ,in It is the number of critical transmission paths. This refers to the number of time points. Optionally, dimensionality reduction can also employ the t-SNE method to better preserve the situational differences between paths in a lower-dimensional space. In practical implementation, the workshop operation status map can be dynamically updated and displayed on the monitoring interface in the form of a heatmap, with the time axis as the horizontal axis, the path axis as the vertical axis, and the intensity represented by pixel color.

[0039] See Figure 4 This is a trend chart showing the dynamic changes in the topological connection weights within a virtual monitoring framework for a manufacturing workshop. All three curves exhibit a slow upward trend, indicating that the weights of these three topological connections in the virtual monitoring framework have been continuously increasing over 20 iterations, reflecting the accumulating correlation of state changes in the corresponding topological units. The stable vertical spacing between the curves indicates that the relative importance of the three connections has not reversed during the iteration process, which perfectly matches the algorithm's design logic of "dynamically adjusting weights based on the intensity of state changes in active units." The smooth upward trend of the curves directly proves the stability of the "topological connection weight iterative update" algorithm, demonstrating that the virtual monitoring framework has not experienced oscillations or divergence during state changes, and conforms to the preset connectivity constraints.

[0040] In one embodiment of the present invention, the health assessment module reconstructs a load-response relationship model of key structural points in the manufacturing workshop based on the workshop operation status map. The spatial locations of all key structural points are marked in the three-dimensional digital model of the manufacturing workshop. Key structural points include the front bearing housing of the machine tool spindle, stress concentration areas of large welded structures, and support connection points of the overhead conveyor rail. Each key structural point has a unique coordinate identifier in the three-dimensional digital model. Local map region features associated with the spatial location of each key structural point are extracted from the workshop operation status map. These local map region features refer to data blocks composed of path feature sequences adjacent to the key structural points in spatial coordinates within the workshop operation status map. The patterns of local map region features at different time points are analyzed to identify feature change patterns related to load application and feature change patterns related to structural response. Feature change patterns related to load application are characterized by a sharp increase in the amplitude of a specific path feature sequence, while feature change patterns related to structural response are characterized by oscillations at a specific frequency in another set of path feature sequences. A preliminary mapping relationship is established to quantify load-related characteristic change patterns into equivalent load vectors and response-related characteristic change patterns into structural response vectors. The equivalent load vector is a multidimensional vector containing components such as force, torque, or heat flux density, while the structural response vector is a multidimensional vector containing components such as displacement, strain, acceleration, or temperature rise. A dynamic transfer function is fitted based on the correspondence between the equivalent load vector and the structural response vector using a system identification method, employing either recursive least squares or a state-space model. The dynamic transfer function is then combined with the known physical parameters of key structural points to construct a mathematical relationship describing the transition from load input to structural response output. This mathematical relationship is the load-response relationship model, which can be in the form of a differential equation or a state-space equation.

[0041] In practical implementation, the dynamic health indicators of each key structural point are derived using a load-response relationship model. The equivalent load vector at the current moment is acquired in real time and used as input to the load-response relationship model. The theoretical structural response vector of the key structural point under the current load is calculated through the load-response relationship model. The calculation process substitutes the equivalent load vector into the load-response relationship model and solves to obtain the theoretical structural response output. At the same time, the actual observed response vector of the key structural point is obtained by directly measuring it through sensing devices or by inverting it from the workshop operation status map. For the main shaft bearing housing, the actual observed response vector is directly measured by the installed vibration acceleration sensor. The residual vector between the theoretical structural response vector and the actual observed response vector is calculated. The dimensions of the two vectors are aligned to ensure they contain the same number and consistent order of response feature components. Timestamp matching is performed to ensure a precise correspondence between the calculation time of the theoretical structural response vector and the sampling time of the actual observed response vector. The original difference between each component is obtained by subtracting the measured value of the corresponding component in the actual observed response vector from the original value of each response feature component in the theoretical structural response vector. The original difference is then divided by the corresponding observation error tolerance or normalization coefficient to obtain the normalized difference. All normalized differences are arranged in their original order to form the residual vector. Statistical analysis is performed on the residual vector to calculate its norm and covariance to obtain the response deviation, which is a comprehensive scalar indicator. The contribution of the current response deviation to the cumulative structural damage is assessed by combining the material fatigue characteristic curves of key structural points with historical load spectra. The material fatigue characteristic curves are pre-stored in the form of SN curves. Based on the cumulative damage contribution and a preset safety threshold, a value between zero and one is calculated as the current dynamic health index of the critical structural point. The lower the dynamic health index value, the worse the health status. The calculation of the dynamic health index can be expressed as:

[0042] Where: symbol Represents the calculated dynamic health index, symbol This represents the current cycle damage contribution calculated from the material fatigue characteristic curve based on the current response deviation, with the sign... The symbol represents the accumulated damage calculated from the historical load spectrum according to Miner's linear cumulative damage rule. This represents the preset damage threshold that would cause structural failure.

[0043] In some embodiments, the actual observed response vector is retrieved from the workshop operation status map through a pre-trained inverse mapping network. This network takes local map region features as input and outputs estimated structural response values. It is understood that the observation error tolerance is pre-calibrated based on sensor accuracy and the noise level of the measurement environment; for example, the error tolerance for a displacement sensor is ±0.01 mm. In a specific implementation, timestamp matching uses a linear interpolation method to align the sampling time point of the actual observed response vector to the calculation time point of the theoretical structural response vector. Optionally, the L2 norm is used when calculating the residual vector norm to comprehensively measure the overall deviation magnitude. In some embodiments, material fatigue characteristic curves are retrieved from a material database based on the material grade and heat treatment process of key structural points.

[0044] See Figure 5 This is a bar chart analyzing the damage contribution of key structural points in a manufacturing workshop. It visually displays the current damage contribution, historical cumulative damage, and damage threshold for different structural points. The historical cumulative damage of all three structural points has not exceeded the damage threshold, indicating that they are currently operating safely. The historical cumulative damage (0.47) and current cyclic damage contribution (0.12) are the highest of the three, and the remaining safety margin is the smallest, making them key areas of focus for workshop maintenance. The current cyclic damage contribution value shows that the machine tool spindle has the fastest damage growth rate, requiring shorter inspection cycles and close monitoring of its vibration, temperature, and other parameters. Combining the current damage contribution and historical cumulative data, the remaining lifespan of structural points can be predicted to avoid sudden failures. If the current damage contribution of a structural point remains consistently high, the corresponding workstation's machining parameters (such as cutting force and operating speed) can be optimized to reduce the damage accumulation rate.

[0045] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A real-time monitoring system for a manufacturing workshop based on the Internet of Things, characterized in that, include: The signal processing module is used to continuously capture multimodal sensing signals through the IoT sensing layer deployed in the manufacturing workshop, and to perform signal morphology decomposition on the multimodal sensing signals to obtain a set of component signals corresponding to different physical processes. The framework construction module is used to project the component signal set onto a preset virtual monitoring framework. The virtual monitoring framework is composed of multiple interconnected topological units. During the projection process, the component signals are assigned to the corresponding topological units according to their source attributes to form a topological unit state vector. The topological connection weights of the virtual monitoring framework are iteratively updated using the topological unit state vector. The dynamic health assessment module is used to reconstruct the load-response relationship model of key structural points in the manufacturing workshop based on the workshop operation status map, and to deduce the dynamic health indicators of each key structural point using the load-response relationship model. The safety assessment module aggregates dynamic health indicators for all critical structural points to calculate the system-level safety margin of the manufacturing workshop.

2. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 1, characterized in that, The step of performing signal morphology decomposition on the multimodal sensing signal to obtain a set of component signals corresponding to different physical processes includes: A basis function library for signal decomposition is established, which contains basis function templates characterizing vibration, heat conduction, fluid motion, and electromagnetic interference; For each captured multimodal sensing signal, adaptive matching and tracking are performed in the basis function library to select multiple basis function templates whose matching degree with the modal sensing signal exceeds a preset threshold. The multimodal sensing signal is synchronously expanded using the selected basis function templates to obtain a set of initial decomposition coefficients; The initial decomposition coefficients are optimized by sparsity constraints, and redundant decomposition components are removed to obtain the optimized set of decomposition coefficients. Based on the optimized set of decomposition coefficients and their corresponding basis function templates, multiple independent signal components are reconstructed; The reconstructed signal components are traced back to their physical processes. Based on their energy distribution characteristics and waveform properties, they are classified into mechanical vibration process components, thermodynamic process components, material transport process components, or electromagnetic compatibility process components, ultimately forming the set of component signals.

3. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 2, characterized in that, The step of projecting the set of component signals onto a preset virtual monitoring frame includes: The virtual monitoring framework is defined as a multi-dimensional state space, where each dimension of the multi-dimensional state space corresponds to a process element, and each topological unit represents a coordinate point in the multi-dimensional state space. For each signal component in the set of component signals, calculate its characteristic coordinates in the multidimensional state space. The calculation process combines the physical attribute labels and time-frequency characteristics of the signal components. Signal components with similar characteristic coordinates are clustered to form signal component clusters; The center coordinates of each signal component cluster are mapped to the nearest topological unit in the virtual monitoring framework, thus establishing a binding relationship between the signal component cluster and the topological unit; Based on the binding relationship, the energy amplitude, frequency band information and change trend of each signal component cluster are encoded into the topology unit state vector, and each element of the topology unit state vector corresponds to the current state code of a topology unit.

4. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 3, characterized in that, The step of iteratively updating the topology connection weights of the virtual monitoring framework using the state vector of the topology unit includes: Initialize the connection weights between all topological units in the virtual monitoring framework; Calculate the difference vector between the current topological unit state vector and the previous topological unit state vector; Based on the direction and magnitude of the difference vector, the set of active topological units whose states have changed significantly is identified; Within the virtual monitoring framework, neighboring topological units that have a physical or logical association with the set of active topological units are identified. Based on the intensity of the state change of the active topology unit and its preset coupling relationship with the neighboring topology units, the connection weight from the active topology unit to the neighboring topology unit is dynamically adjusted to enhance or weaken the connection strength. After each weight adjustment, check whether the overall connection graph of the virtual monitoring framework meets the preset connectivity constraints. If it does not meet the constraints, fine-tune the local weights to ensure global connectivity. Repeat the process from calculating the difference vector to checking and fine-tuning connectivity until the change in the topological unit state vector tends to stabilize or reaches the preset number of iterations, thereby completing the update of the topological connection weights.

5. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 4, characterized in that, Based on the updated topology connection weights, key propagation paths are activated in the virtual monitoring framework, including: In the updated virtual monitoring framework, one or more state initiation units are set; Starting from the state initiation unit, a depth-first state propagation simulation is performed along the topological connections; During the state propagation simulation, the attenuation degree and propagation delay of state information as it passes through each connection are recorded. Based on the recorded attenuation level and propagation delay, calculate the cumulative propagation loss from the initial state unit to other topological units; All paths with cumulative propagation loss below a set threshold are selected as candidate propagation paths; Redundancy elimination processing is performed on the candidate propagation paths to remove paths that are spatially overlapping or functionally equivalent. The remaining paths after redundancy elimination are marked as critical propagation paths, and a path identifier is assigned to each critical propagation path.

6. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 5, characterized in that, Perform temporal convolution on the topological unit state vectors along the key propagation path to generate a workshop operation status map, including: For each marked critical propagation path, extract the sequence of topological units traversed by the critical propagation path according to the path identifier; Obtain the topology unit state vector fragments of the topology unit sequence over multiple consecutive time slices; Design a convolutional kernel suitable for the critical propagation path, the length of which matches the number of topological units in the critical propagation path; Using temporal convolution operations, the convolution kernel is slid along the time dimension to perform convolution calculations on the topological unit state vector fragments; During the convolution calculation, the temporal state correlation of multiple topological units on the fusion path is fused to output a one-dimensional feature sequence that reflects the overall situational change of the key propagation path. Repeat the convolution calculation process from extracting the topological unit sequence to outputting the one-dimensional feature sequence for all critical propagation paths to obtain a set of feature sequences equal to the number of critical propagation paths; All the obtained feature sequences are spatially aligned and concatenated to form a multidimensional feature matrix. The multidimensional feature matrix is ​​reduced in dimensionality and visualized by encoding, and then transformed into a graph structure containing a time axis, a path axis, and an intensity axis, which is the workshop operation status graph.

7. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 6, characterized in that, Based on the workshop operation status map, a load-response relationship model for key structural points within the manufacturing workshop is reconstructed, including: The spatial locations of all key structural points are marked in the three-dimensional digital model of the manufacturing workshop; From the workshop operation status map, extract the local map region features associated with the spatial location of each key structural point; Analyze the patterns of the local map region features at different time points to identify feature change patterns related to load application and feature change patterns related to structural response; Establish a preliminary mapping relationship, quantize the load-related characteristic change patterns into equivalent load vectors, and quantize the response-related characteristic change patterns into structural response vectors; Using a system identification method, a dynamic transfer function is fitted based on the correspondence between the equivalent load vector and the structural response vector in historical data. By combining the dynamic transfer function with the known physical parameters of key structural points, a mathematical relationship describing the transition from load input to structural response output is constructed. This mathematical relationship is the load-response relationship model.

8. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 7, characterized in that, The dynamic health indicators of each key structural point are derived using the load-response relationship model, including: The equivalent load vector at the current moment is obtained in real time and used as the input to the load-response relationship model; The theoretical structural response vector of the key structural point under the current load is calculated using the load-response relationship model. Meanwhile, the actual observed response vectors of the key structural points are obtained by directly measuring them through sensing devices or by inverting them from the workshop operation status map; Calculate the residual vector between the theoretical structural response vector and the actual observed response vector; Statistical analysis is performed on the residual vector to calculate its norm and covariance, thereby obtaining the response deviation. By combining the material fatigue characteristic curves and historical load spectra of the key structural points, the contribution of the current response deviation to the cumulative structural damage is evaluated. Based on the cumulative damage contribution and the preset safety threshold, a value between zero and one is calculated as the current dynamic health indicator of the key structural point. The lower the dynamic health indicator value, the worse the health status.

9. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 5, characterized in that, The calculation of the cumulative propagation loss from the initial state unit to other topological units, based on the recorded attenuation level and propagation delay, includes: Extract the attenuation factor and propagation delay value for each connection from the records of the state propagation simulation; For each state starting unit, all other topological units in the virtual monitoring framework are traversed as target units; For each target unit, enumerate all possible paths from the state starting unit to the target unit, and each path consists of a series of topological connection sequences; For each path, the attenuation factor of all connections on the path is multiplied to obtain the total attenuation factor of the path, and the propagation delay values ​​of all connections on the path are summed to obtain the total propagation delay of the path. Using a preset attenuation-delay joint cost function, the total attenuation factor and total propagation delay are mapped to a scalar cost, which serves as the propagation loss of the path. For each target unit, the value with the minimum propagation loss is selected from all enumerated paths and used as the cumulative propagation loss from the state starting unit to the target unit.

10. The IoT-based real-time monitoring system for manufacturing workshops as described in claim 8, characterized in that, The calculation of the residual vector between the theoretical structural response vector and the actual observed response vector includes: Align the dimensions of the theoretical structural response vector with those of the actual observed response vector to ensure that the two vectors contain the same number of response feature components in the same order. Timestamp matching is performed on the two vectors to ensure that the calculation time point of the theoretical structural response vector corresponds precisely to the sampling time point of the actual observed response vector; Subtract the measured value of the corresponding component in the actual observed response vector from the value of each response feature component in the theoretical structural response vector to obtain the original difference of each component. Divide the original difference of each component by the corresponding observation error tolerance or normalization coefficient to obtain the normalized difference; Arrange all normalized differences in their original order to form the residual vector.

Images