A method and system for monitoring the status of industrial IoT devices

Abstract

This invention relates to the field of equipment monitoring technology, specifically to a method and system for monitoring the status of industrial IoT devices. The method includes: acquiring multimodal time-series data of the industrial IoT device during operation; comparing and analyzing the multimodal time-series data with expected multimodal time-series data to obtain the status performance reliability of the industrial IoT device; if status performance degradation is detected in the industrial IoT device based on the status performance reliability, diagnosing the cause of the degradation based on the multimodal time-series data; and generating a control strategy for compensating and adjusting the device operating parameters based on the cause of degradation. This effectively solves the problem that the degradation result cannot be transformed into a compensatory control strategy to maintain or restore device performance, leading to a break in the closed loop from status perception to control execution.

Description

Technical Field

[0001] This invention relates to the field of equipment monitoring technology, and specifically to a method and system for monitoring the status of industrial Internet of Things (IoT) devices. Background Technology

[0002] Industrial Internet of Things (IIoT) equipment status monitoring is a core technology for ensuring production continuity, improving overall equipment efficiency, and enabling predictive maintenance. By sensing and analyzing the equipment's operating status in real time, it can effectively provide early warnings of faults, reduce unplanned downtime, and optimize maintenance decisions.

[0003] Currently, industrial IoT equipment status monitoring methods mainly rely on real-time or periodic analysis of collected single-dimensional or multi-dimensional sensor data such as vibration, temperature, and current. By comparing the monitoring data with preset static thresholds or dynamic warning lines generated by statistical models or machine learning models trained based on historical data, it is determined whether the equipment is in an abnormal or faulty state, and an alarm is triggered when the threshold is exceeded.

[0004] However, this method is essentially a passive response paradigm of monitoring and alarm. Even if it can identify performance degradation or abnormal trends, it can only provide early warning information such as "what happened" or "what might happen," but cannot automatically generate and issue decision instructions on "how to carry out targeted control." In other words, the monitoring system and the control system are disconnected, and the results of state perception cannot be directly and automatically transformed into control actions to maintain or restore equipment performance. This results in a delay in the process from problem discovery to manual intervention and analysis, and then to manual adjustment of control parameters. Moreover, it relies on expert experience and is difficult to achieve autonomous optimization and closed-loop control of equipment operating status. Summary of the Invention

[0005] To address the technical problem that the failure to translate equipment performance degradation into compensatory control strategies to maintain or restore equipment performance leads to a break in the closed loop from state perception to control execution, this application provides an industrial Internet of Things (IoT) equipment state monitoring method and system.

[0006] The industrial IoT device status monitoring method and system provided in this application adopts the following technical solution: A method for monitoring the status of industrial IoT devices, comprising: Acquire multimodal time-series data during the operation of industrial IoT devices; By comparing and analyzing the multimodal time series data with the expected multimodal time series data, the reliability of the state performance of industrial IoT devices can be obtained. If state performance degradation is detected in an industrial IoT device based on state performance reliability monitoring, the cause of the degradation is diagnosed based on multimodal time series data. Based on the causes of degradation, a control strategy is generated to compensate and adjust the operating parameters of industrial IoT devices.

[0007] Furthermore, the steps for acquiring multimodal time-series data during the operation of industrial IoT devices include: Physical sensing signals are acquired through physical sensors, and control status signals are synchronously acquired with the industrial bus through the controller interface on the controller. Interpolation and synchronization processing are performed on physical sensor signals and control status signals to obtain a synchronization signal stream; The frequency domain features of the physical sensing signals are extracted from the synchronization signal stream to form the physical feature vector, and the instruction codes, status codes and numerical sequences of the control state signals are extracted to form the control state feature vector. Based on the equipment process logic relationship between physical feature vectors and control state feature vectors, multimodal feature segments are obtained by fusing them together. Multimodal feature segments are spliced ​​together according to the timeline to obtain multimodal time series data.

[0008] Furthermore, the steps for comparing and analyzing the multimodal time series data with the expected multimodal time series data to obtain the state performance reliability of industrial IoT devices include: Multimodal time series data is input into a preset digital twin, and the expected multimodal time series data is obtained by deducing from the preset digital twin; The time-frequency feature differences between multimodal time-series data and expected multimodal time-series data are calculated to obtain a difference metric. Based on the statistical distribution of the difference metric over the time window, the uncertainty of the current device state of an industrial IoT device compared to the expected device state is assessed, and an uncertainty quantification vector is obtained. The state performance reliability is obtained by weighting the uncertainty quantization vector.

[0009] Furthermore, the steps for calculating the time-frequency feature differences between the multimodal time-series data and the expected multimodal time-series data to obtain the difference metric include: The multimodal time series data and the expected multimodal time series data are aligned in the time domain to obtain the real-time sequence and the expected sequence; Real-time time-domain features and real-time frequency-domain features are extracted from the real-time sequence, and expected time-domain features and expected frequency-domain features are extracted from the expected sequence. The time warping distance between the real-time time-domain features and the expected time-domain features is calculated to obtain the time-domain difference component, and the cosine similarity between the real-time frequency-domain features and the expected frequency-domain features is calculated to obtain the frequency-domain difference component. By using a pre-set adaptive attention weighting network, the weights of the time-domain difference components and the frequency-domain difference components are adjusted respectively to obtain the difference scores of each modality. The difference scores of each modality are aggregated along the time dimension on a preset sliding window, and the mean and standard deviation of the difference scores of each modality on the preset sliding window are calculated to obtain the difference measure.

[0010] Furthermore, if state performance degradation is detected in an industrial IoT device based on state performance reliability monitoring, the steps for diagnosing the cause of state performance degradation based on multimodal time series data include: When the state performance confidence level is lower than the preset confidence level threshold, it is determined that the industrial IoT device has state performance degradation, triggering the diagnostic process. According to the diagnostic process, multimodal time series data are placed into a degradation time window, and abnormal fluctuation patterns of multimodal time series data in the degradation time window are extracted to obtain a candidate abnormal feature set; Based on the preset equipment degradation knowledge graph, graph matching and causal reasoning are performed on the candidate abnormal feature set to obtain an initial diagnosis list; Analyze the temporal causal relationship between multimodal time series data and each initial diagnostic cause in the initial diagnostic list, and output a reliability score for each initial diagnostic cause based on the analysis results; The reliability scores are sorted and filtered, and the initial diagnostic cause corresponding to the reliability score ranked first is determined as the cause of degradation.

[0011] Furthermore, based on the causes of degradation, the steps for generating control strategies to compensate and adjust the operating parameters of industrial IoT devices include: Query the preset cause-compensation mapping knowledge base to obtain several compensation directions associated with the cause of degradation; Based on the real-time status of the equipment reflected by the compensation direction and multimodal time series data, the adjustment parameters under each compensation direction are calculated to obtain several initial compensation strategies. Simulations were performed on each initial compensation strategy, and the impact of each initial compensation strategy on equipment performance indicators was predicted based on the simulation results, thus obtaining the performance score of each initial compensation strategy. The performance scores are sorted and filtered, and the initial compensation strategy corresponding to the performance score ranked first is determined as the control strategy.

[0012] This application also provides an industrial IoT device status monitoring system, including: The data acquisition module is used to acquire multimodal time-series data during the operation of industrial IoT devices; The data analysis module is used to compare and analyze multimodal time series data with expected multimodal time series data to obtain the reliability of the state performance of industrial IoT devices. The data diagnostic module is used to diagnose the cause of state performance degradation based on multimodal time series data when state performance degradation is detected in industrial IoT devices based on state performance reliability monitoring. The strategy output module is used to generate control strategies for compensating and adjusting the operating parameters of industrial IoT devices based on the causes of degradation.

[0013] Beneficial effects achieved: This application provides a method for monitoring the status of industrial IoT devices, including: acquiring multimodal time-series data of the industrial IoT device during operation; comparing and analyzing the multimodal time-series data with expected multimodal time-series data to obtain the status performance reliability of the industrial IoT device; if status performance degradation of the industrial IoT device is detected based on the status performance reliability, diagnosing the cause of the degradation based on the multimodal time-series data; and generating a control strategy for compensating and adjusting the device operating parameters of the industrial IoT device based on the cause of degradation.

[0014] In this application, acquiring multimodal time-series data of industrial IoT devices during operation provides a comprehensive and integrated real-time status information foundation for the entire analysis, enabling monitoring to go beyond single-signal judgment and more comprehensively capture the actual operating conditions of the equipment. Based on this, by comparing and analyzing the multimodal time-series data with expected multimodal time-series data, the reliability of the industrial IoT device's status performance is obtained. This quantifies the results of multidimensional data fusion analysis into an assessable probability or confidence level of performance degradation, achieving a quantitative judgment on whether performance degradation has occurred, and providing guidance on whether and how subsequent intervention is needed. This provides the triggering basis; then, when performance degradation is determined based on this confidence level, the system executes a diagnostic analysis of the degradation cause based on multimodal time-series data, deeply analyzes the multimodal time-series data to locate the root cause of the degradation, thereby transforming the vague performance degradation alarm into a specific equipment anomaly or component problem; finally, based on the diagnosed degradation cause, a control strategy is generated to correct or compensate for the degradation cause, thus realizing a complete chain from perceiving anomalies, diagnosing the root cause to generating and outputting targeted control strategies, bridging the gap from state perception to control execution, and forming a closed loop of autonomous decision-making and execution. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of an industrial Internet of Things (IoT) device status monitoring method according to this application. Figure 2 This is a schematic diagram of the modules of the industrial IoT device status monitoring system of this application.

[0016] Explanation of reference numerals in the attached figures: 10. Data acquisition module; 20. Data analysis module; 30. Data diagnosis module; 40. Strategy output module. Detailed Implementation

[0017] The following combination Figure 1 and Figure 2 This application will be described in further detail.

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

[0019] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0020] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0021] This application discloses a method for monitoring the status of industrial Internet of Things (IoT) devices.

[0022] Please refer to Figure 1 The industrial IoT device status monitoring method proposed in this embodiment includes steps S10~S40: Step S10: Obtain multimodal time-series data of the industrial IoT device during operation.

[0023] Acquiring multimodal time-series data of industrial IoT devices during operation avoids the limitations of traditional monitoring methods that rely on single-dimensional signals for status judgment. The collected multimodal time-series data includes physical sensor signals and control status signals. By using these two different but interrelated signals, a unified time-stamped data foundation is constructed that comprehensively and accurately reflects the overall operating status of industrial IoT devices from multiple perspectives. Integrating previously isolated signals allows for a more complete and accurate digital characterization of the complex operating behavior of industrial IoT devices at the data level. This provides input for subsequent analysis, ensuring that the system not only observes the physical performance of the equipment, such as excessive vibration, but also perceives its control and execution status, such as slow command response and logical integrity. This lays the data foundation for ultimately achieving accurate status assessment, root cause diagnosis, and the generation of effective control strategies.

[0024] Step S20: Compare and analyze the multimodal time series data with the expected multimodal time series data to obtain the reliability of the state performance of the industrial IoT device.

[0025] By comparing and analyzing multimodal time-series data with expected multimodal time-series data, the state performance reliability of industrial IoT devices can be obtained. This allows for the assessment of the current operational health of industrial IoT devices. By comparing the collected multimodal time-series data with expected multimodal time-series data that characterizes the normal or ideal operating state of industrial IoT devices, the state performance reliability reflecting the degree to which the actual state of industrial IoT devices deviates from the normal state can be calculated from the data comparison.

[0026] This step constructs a bridge for transforming multimodal time-series data into intuitively interpretable state performance reliability. By converting the difference between multimodal time-series data and expected multimodal time-series data into a state performance reliability that characterizes the overall or critical performance reliability of the device, the output state performance reliability can not only reveal whether the device may have anomalies, but also quantitatively indicate the extent to which its performance is trustworthy or has degraded. This provides a quantitative reference for subsequent steps to determine whether diagnostics need to be triggered and for diagnostic analysis.

[0027] Step S30: If state performance degradation is detected in an industrial IoT device based on state performance reliability monitoring, the cause of state performance degradation is diagnosed based on multimodal time series data.

[0028] By deepening our understanding based on the quantified reliability of state performance, we move from "judging whether there is an anomaly" to "locating the specific cause of degradation," aiming to overcome the limitations of traditional monitoring methods that stop at alarms and cannot provide specific fault or degradation indications.

[0029] In this step, a state performance reliability index representing equipment performance degradation is transformed into a degradation cause that can guide specific compensatory control. By utilizing multimodal time-series data as the analysis source, the abnormal correlations and causal relationships between different data modes in time series are explored. This allows the general state performance reliability index to be interpreted as a degradation cause pointing to a specific component, specific link, or specific physical process. This endows the entire monitoring method with attribution and insight capabilities, enabling the system not only to know that the equipment "may fail" (low state performance reliability), but also to clearly infer "where it fails and what might be the cause" (specific degradation cause). This provides input for the subsequent generation of targeted control strategies.

[0030] Step S40: Based on the cause of degradation, generate a control strategy for compensating and adjusting the operating parameters of industrial IoT devices.

[0031] By using the diagnosed causes of degradation as direct input, control strategies designed to specifically offset, compensate for, or adapt to these causes are derived. This transforms the results of the initial analysis and diagnosis into executable commands that can directly drive the action of industrial IoT devices, restoring or maintaining their performance. By translating the degradation causes, which characterize "where the problem lies," into specific control strategies instructing "how to adjust," the final link in the process—from state perception, assessment, and diagnosis to the generation of control actions—is completed. This transforms the entire system from a passive monitoring and alarm system into an intelligent control system capable of proactive intervention and autonomous optimization. It bridges the gap between state perception and low-level control execution in traditional monitoring, enabling industrial IoT devices to automatically, quickly, and accurately implement compensatory adjustments when facing performance degradation, without relying on manual experience analysis and parameter adjustments. This maintains the continuous and stable operation of industrial IoT devices while improving the timeliness and intelligence of maintenance response, achieving a key closed loop from intelligent monitoring to autonomous maintenance of industrial IoT devices.

[0032] In one feasible implementation, step S10 may specifically include steps S11 to S15, wherein the industrial IoT device is equipped with physical sensors and a controller: Step S11: Collect physical sensing signals through physical sensors and synchronously collect control status signals with the industrial bus through the controller interface on the controller.

[0033] By deploying physical sensors such as vibration, temperature, and acoustic emission sensors on the key mechanical structures of industrial IoT devices, physical quantities are converted into analog electrical signals. After signal conditioning circuitry, these signals are acquired in real time by an analog-to-digital converter module and converted into discrete digital signal sequences, forming physical sensing signals. Simultaneously, by accessing the controller interface and parsing real-time communication messages on the industrial bus, the system captures motion control commands issued by the controller, real-time current and position feedback from servo drives, and key state variables within the programmable logic controller (PLC) through event triggering or periodic polling.

[0034] Step S12: Interpolate and synchronize the physical sensing signal and the control state signal to obtain a synchronization signal stream.

[0035] The physical sensing signals and control status signals are sorted and arranged on the same time axis according to the order of their acquisition time. Since the physical sensing signals are usually sampled at a fixed high frequency, such as 10k points per second, while the updates of the control status signals may be triggered by events or at a lower fixed frequency, such as 100 points per second, the physical sensing signals and control status signals are non-uniform and non-aligned on the time axis, and interpolation algorithms are required for synchronization processing.

[0036] For digital quantities such as control state signals, which typically exhibit step-like changes, forward filling or nearest neighbor hold methods are used to maintain their state values ​​until the next state update time. For analog quantities such as physical sensor signals, which change continuously, linear interpolation or spline interpolation algorithms are used to calculate the corresponding physically continuous sensor signal approximation value at the target time point of the control state signal update.

[0037] Specifically, for control status signals, when the system receives a timestamped control status signal update, such as a state transition from running to stopping, it records the corresponding discrete state value (such as a status code, instruction code, or specific register value) along with its timestamp into the time-series database. If the forward-fill method is used, the discrete state value corresponding to the last state update before the target time point is directly selected, and this discrete state value is regarded as a constant state for all time intervals from the time of the update until the next update. Therefore, for any query of the target time point within this interval, the algorithm returns the same filled discrete state value. If the nearest neighbor hold method is used, the time difference between the target time point and the time point of its last state update is calculated, as well as the time difference between the target time point and the time point of its first subsequent state update. By comparing the magnitude of these two time differences, the discrete state value corresponding to the update point that is closer in time distance is selected as the result.

[0038] For continuously changing analog quantities like physical sensor signals, the system maintains a timestamp-ordered buffer queue of physical sensor signal sampling points in memory. Each physical sensor signal sampling point contains a timestamp and its corresponding signal amplitude. When interpolation calculations are needed, a target time point is received from a control state signal. The system then searches and locates two physical sensor signal sampling points adjacent to the target time point in the buffer queue. One of these sampling points has a timestamp exactly earlier than the target time, and the other has a timestamp exactly later. If a linear interpolation algorithm is used, the amplitudes of these two sampling points are considered as two endpoints on a straight line. The interpolation is then performed based on the relative position of the target time point within the time interval of these two endpoints. The first method is to calculate the amplitude of the target time point by linearly calculating the corresponding amplitude within the amplitude range of the two points according to the proportion of the time difference between the target time point and the previous time point to the total time interval between the two points. This calculation result is the approximate value of the physical sensor signal at the target time point. If a spline interpolation algorithm is used, more physical sensor signal sampling points before and after the target time point will be selected, such as two physical sensor signal sampling points before and after the target time point. Based on these points, a smooth polynomial curve conforming to a specific order, such as the third order, is fitted. This polynomial curve can mathematically guarantee that it passes through these physical sensor signal sampling points and maintains continuous smoothness. Then, the target time point is substituted into the fitted polynomial curve formula, and the result directly calculated is the approximate value of the physical sensor signal.

[0039] After the above processing, each control state signal update point simultaneously has a corresponding physical sensor signal, thus forming a data pair that is completely aligned in time and corresponds one-to-one in value. These data pairs are connected in chronological order to form a synchronous signal stream, ensuring that any cross-modal comparison, correlation and causal analysis in subsequent steps are absolutely accurate and reliable in terms of timing logic.

[0040] Step S13: Extract the frequency domain features of the physical sensing signal from the synchronization signal stream to form a physical feature vector, and extract the instruction code, status code and numerical sequence of the control state signal to form a control state feature vector.

[0041] A segment of physical sensor signal sequence data is extracted from the synchronization signal stream according to a preset time window length (e.g., 1 second) or a specific event window (e.g., a complete operating cycle of the device). For example, a time window containing 1024 physical sensor signal sampling points is windowed to reduce spectral leakage. Then, a short-time Fourier transform or fast Fourier transform algorithm is applied to convert the physical sensor signal sequence data segment from a time-domain signal to a frequency-domain representation, thereby obtaining its power spectrum. Next, a set of statistical features that can characterize the essential properties of the physical sensor signal are calculated from the power spectrum. These statistical features typically include the spectral centroid, root mean square frequency, frequency variance, and energy proportion within a specific frequency band. Arranging these calculated statistical values ​​in a predefined order constitutes a physical feature vector representing the state of the physical sensor signal within the time period.

[0042] Simultaneously, for control status signals aligned with the captured physical sensor signal sequence data segments within the same time window, the protocol fields of the control status signals are directly parsed to extract instruction codes representing the control action type (such as G-code instructions and PLC function codes), status codes reflecting the logical state of the equipment (such as servo enable and alarm codes), and related specific numerical sequences (such as target position, set speed, and actual current value). These discrete symbolic and numerical information are arranged and concatenated according to a predefined field order to form a structured control status feature vector, thereby completing the conversion from signal to feature. This provides input for subsequent cross-modal feature fusion steps, effectively improving the efficiency of data processing and the feasibility and accuracy of subsequent pattern recognition and correlation analysis.

[0043] The predefined field order refers to a fixed data template or format specification that is predefined and used to structure and organize different information in the control status signal.

[0044] Step S14: Based on the equipment process logic relationship between the physical feature vector and the control state feature vector, the multimodal feature segments are obtained by fusing them.

[0045] It should be noted that the industrial IoT equipment status monitoring system has an equipment process logic relationship mapping table. This equipment process logic relationship mapping table defines specific control state feature vectors, such as "instruction code = axial feed, target speed > 1000 rpm". When the industrial IoT equipment is working normally, the range and pattern of changes in the physical characteristics that should be triggered or associated are defined. For example, the "spectral centroid" in the corresponding physical feature vector should be in a certain range, and the "high frequency band energy ratio" should be lower than a certain threshold.

[0046] In actual fusion, physical feature vectors and control state feature vectors generated within the same time window are used. Based on the rules defined in the equipment process logic mapping table, the two vectors are selectively combined and processed. For example, when the control state feature vector indicates that the industrial IoT device is performing a "precision spot welding" process, according to the process logic, special attention is paid to the feature dimensions representing "instantaneous impact energy" and "thermal radiation change" in the physical feature vector. These two physical feature values ​​are then fused with the values ​​of "welding current setpoint" and "pressure holding time" in the control state feature vector through weighted splicing or the construction of joint features, such as calculating the ratio of physical impact energy to welding current, to form a multimodal feature fragment.

[0047] The fusion in this step is not a simple vector connection, but rather generates a multimodal feature fragment based on the causal logic between control intent and physical response. This fragment can simultaneously characterize whether the physical response of an industrial IoT device conforms to the expected process model under a specific control command. This allows subsequent analysis and comparison to directly assess the state of industrial IoT devices based on fused features with clear process implications, thereby greatly improving the accuracy, interpretability, and sensitivity to early minor performance degradation in state assessment.

[0048] Step S15: The multimodal feature segments are spliced ​​together according to the time axis order to obtain multimodal time series data.

[0049] A data buffer is maintained, strictly ordered by timestamps in ascending order. Each multimodal feature segment entering this buffer carries the start timestamp of its corresponding time window. During concatenation, each multimodal feature segment is treated as a complete data unit based on the order of its start timestamps, and they are arranged sequentially end-to-end to form a sequence. To ensure strict temporal continuity and avoid information gaps, the time windows corresponding to adjacent multimodal feature segments are usually consecutive or partially overlapping. This ensures that the state of an industrial IoT device at any given time can be included in the sequence, forming a multimodal time-series data composed of a large number of multimodal feature segments. This data fully preserves the dynamic evolution process and historical context information of the industrial IoT device's operating state, enabling subsequent comparative analysis and other steps to be performed based on a multimodal time-series data with coherent temporal dependencies. This allows for the identification of the gradual change patterns, periodic laws, and causal relationships of the industrial IoT device's state over time.

[0050] In one feasible implementation, step S20 may specifically include steps S21 to S24: Step S21: Input the multimodal time series data into the preset digital twin, and deduce the expected multimodal time series data through the preset digital twin.

[0051] It should be noted that the preset digital twin is a computational model trained based on the historical health status data of industrial IoT devices. It can simulate the operating behavior of industrial IoT devices under normal and non-degrading conditions. It is usually composed of multiple machine learning sub-models, such as long short-term memory networks or temporal convolutional networks, which can characterize the complex dynamic relationship between the multimodal inputs of industrial IoT devices under different operating conditions, such as the preceding control commands and physical responses, and the multimodal outputs, such as the subsequent expected control states and expected physical responses.

[0052] During the simulation, the multimodal time-series data is input as the initial condition sequence into a preset digital twin. The computational model inside the preset digital twin simulates the idealized operation of the industrial IoT device under fault-free and performance degradation-free conditions based on the input initial condition sequence. It predicts the control state signal sequence and physical sensor signal sequence that the device should generate in the next equal time period. The predicted control state signal sequence and physical sensor signal sequence are constructed into expected multimodal time-series data that are completely consistent with the input multimodal time-series data in terms of time length, data structure, and feature dimensions through the same steps as steps S11 to S15. This provides a benchmark data that matches the current actual operating conditions for subsequent comparative analysis, rather than a general threshold. This allows the detection of performance degradation of industrial IoT devices to exclude the influence of normal operating condition fluctuations and more accurately focus on abnormal deviations caused by the degradation of the industrial IoT devices themselves, greatly improving the accuracy of condition assessment.

[0053] Step S22: Calculate the time-frequency feature differences between the multimodal time series data and the expected multimodal time series data to obtain the difference metric value.

[0054] The time-frequency feature difference is calculated on multimodal time-series data and expected multimodal time-series data to obtain a difference metric value. This allows for the quantitative characterization of the micro-device deviation between the actual operating state of industrial IoT devices and the ideal health state deduced from a preset digital twin from two physical analysis dimensions: time domain and frequency domain. The aim is to go beyond the direct comparison of simple scalar values ​​and thus reveal more comprehensively and profoundly the specific modes, severity, and characteristic space of performance degradation of industrial IoT devices.

[0055] By decomposing and transforming the differences between multimodal time-series data and expected multimodal time-series data into computable difference measures in time-domain features such as waveform morphology and timing alignment, and frequency-domain features such as energy distribution and dominant frequency, these difference measures provide complementary descriptions of the deviation from both the temporal evolution law and the frequency component structure of the signal.

[0056] Furthermore, step S22 may include steps S221 to S224: Step S221: Perform time-domain alignment between the multimodal time-series data and the expected multimodal time-series data to obtain the real-time sequence and the expected sequence.

[0057] Since the actual collected multimodal time series data and the expected multimodal time series data derived from the preset digital twin may have systematic offsets in absolute time starting point and micro time phase, such as offsets caused by communication delays or synchronization errors of the initial state of the model, it is necessary to perform time domain alignment to eliminate the existing offsets and make the real-time sequence and the expected sequence synchronized on the time axis.

[0058] First, find an optimal time offset or warping path that maximizes the overall similarity between the multimodal time series data and the expected multimodal time series data after adjustment. A dynamic time warping algorithm can be used, which involves calculating the local distance matrix between each multimodal feature segment in the multimodal time series data and the expected multimodal time series data, and searching for a path with the minimum cumulative distance from the start point to the end point in the local distance matrix. This path defines how to stretch or compress the multimodal time series data or the expected multimodal time series data along the time axis so that it is morphologically best aligned with the other time series data. In practical applications, multimodal feature segments at various times in multimodal time series data and expected multimodal time series data are used as basic units. The Euclidean distance or cosine distance between units is calculated as a local cost. The optimal alignment path is solved through dynamic programming. Based on this path, the time index of multimodal time series data or expected multimodal time series data is remapped to generate a sequence corresponding to another time series data at a given time point. This method is used to obtain real-time and expected sequences, thereby eliminating spurious differences introduced by degradation factors of non-industrial IoT devices, such as time misalignment, from the time dimension. This ensures that the calculation of feature differences in the time domain and frequency domain is performed between states at the same time.

[0059] Step S222: After extracting real-time time-domain features and real-time frequency-domain features from the real-time sequence, and extracting expected time-domain features and expected frequency-domain features from the expected sequence, calculate the time warping distance between the real-time time-domain features and the expected time-domain features to obtain the time-domain difference component, and calculate the cosine similarity between the real-time frequency-domain features and the expected frequency-domain features to obtain the frequency-domain difference component.

[0060] Extracting time-domain features from real-time and expected sequences involves calculating a set of numerical values ​​that characterize the time-domain waveform statistical properties of each multimodal feature segment at each corresponding time point in the two sequences. For example, the mean, variance, peak value, root mean square value, and zero-crossing rate of a certain multimodal feature are calculated within a short time window. These calculated data are then arranged in a fixed order to form the real-time and expected time-domain features at the corresponding time points.

[0061] Extracting frequency domain features involves treating the same multimodal feature segment as a time-domain signal, obtaining its spectral representation through windowing and fast Fourier transform, and then calculating a set of frequency domain descriptors from this spectral representation, such as spectral centroid, spectral variance, spectral entropy, and energy ratio within a specific frequency band. These descriptors are arranged in a fixed order to form the real-time frequency domain features and expected frequency domain features at the corresponding time points.

[0062] The calculation of the time-domain difference component involves taking the sequences of real-time time-domain features extracted from all time points within the entire analysis period and the sequences of expected time-domain features as two new time series. The dynamic time warping algorithm is then applied again to find an optimal nonlinear alignment path through dynamic programming, such that the cumulative distance between the two time series after considering local scaling is minimized. This minimum cumulative distance value is output as the time-domain difference component, which quantifies the overall dissimilarity of the two time series in terms of the evolution of time-domain features.

[0063] The frequency domain difference component is calculated by calculating the cosine similarity between the real-time frequency domain features and the expected frequency domain features at each time step. The cosine similarity is obtained by dividing the dot product of the vectors of the two frequency domain features by the product of their magnitudes. The closer the value is to 1, the more consistent the directions are, i.e., the more similar the spectral structures are. Then, the cosine similarity results of all time steps in the entire analysis period are averaged and the average value is subtracted from 1. The result is the frequency domain difference component, which quantifies the average deviation between the real-time frequency domain features and the expected frequency domain features in the frequency domain feature distribution.

[0064] Step S223: By using a preset adaptive attention weighting network, the weights of the time-domain difference components and the frequency-domain difference components are adjusted respectively to obtain the difference scores of each modality.

[0065] It should be noted that the preset adaptive attention weighted network is a lightweight feedforward neural network model, which is usually trained offline using a large amount of historical data covering different working conditions and different degradation states. Its core function is to learn how the importance of time domain difference components and frequency domain difference components in reflecting the real performance degradation of industrial IoT devices changes dynamically with specific situations.

[0066] In the specific application phase of the network, i.e., during online computation, the input of the pre-defined adaptive attention weighted network includes not only the two scalar values ​​of the time-domain difference component and the frequency-domain difference component, but also contextual information from the multimodal time-series data within the current analysis window. This includes, for example, the signal-to-noise ratio estimation of each modality, the mode encoding of the current industrial IoT device operation, and the statistical characteristics of the difference components themselves. Internally, the pre-defined adaptive attention weighted network calculates and outputs two weight coefficients between 0 and 1, summing to 1, corresponding to the time-domain and frequency-domain difference components, respectively. These weight coefficients are not fixed values ​​but are calculated in real-time by the pre-defined adaptive attention weighted network based on the current input context. For example, when the monitoring system determines that the signal noise is high under the current operating conditions, the pre-defined adaptive attention weighted network may assign higher weights to the more noise-resistant time-domain difference component, while when a suspected periodic fault occurs, it may assign higher weights to the frequency-domain difference component.

[0067] After obtaining the weighting coefficients, the time-domain difference component and the frequency-domain difference component are multiplied by their corresponding weights, and then the two weighted results are added together to obtain a difference score that integrates time-domain and frequency-domain information. This process is performed independently for each mode to obtain its own difference score. Based on the actual data quality and working conditions, the system adaptively decides whether to trust time-domain information or frequency-domain information more in this evaluation. This makes the final difference scores of each mode more accurate and robust in reflecting the true degree of performance degradation of industrial IoT devices under that mode, avoiding the inaccurate evaluation problem caused by fixed weighting strategies in complex industrial scenarios.

[0068] Step S224: Aggregate the difference scores of each modality along the time dimension on a preset sliding window, calculate the mean and standard deviation of the difference scores of each modality on the preset sliding window, and obtain the difference measure value.

[0069] It should be noted that the preset sliding window is a continuous time interval of fixed length defined on the time axis, such as containing the difference scores of the most recent 10 calculation moments.

[0070] During the aggregation process, a preset sliding window moves over time, sliding a fixed time step (e.g., one hour) each time to continuously cover the entire analysis period. For each preset sliding window position, all variance scores of a specific modality falling within that window are extracted, forming a subsequence of variance scores within that preset sliding window. Next, the mean of this subsequence is calculated to obtain the mean reflecting the average level of variance for that modality within that time period. Simultaneously, the standard deviation of this subsequence is calculated to obtain the dispersion index reflecting the degree of variance fluctuation within that time period. Finally, the calculated mean and standard deviation are combined into a tuple [mean, standard deviation]. This tuple represents the variance measure of the corresponding modality within the current preset sliding window, comprehensively characterizing the performance deviation of the corresponding modality within the window from two statistical dimensions: central tendency and dispersion.

[0071] Step S23: Based on the statistical distribution of the difference metric values ​​over the time window, assess the uncertainty of the current device state of the industrial IoT device compared to the expected device state, and obtain the uncertainty quantification vector.

[0072] It should be noted that the expected device state is the state that the industrial IoT device should present under ideal conditions without degradation, which is deduced from the preset digital twin in step S21. Its specific manifestation is the expected multimodal time series data.

[0073] In this step, the uncertainty of the current device state deviating from the expected device state is assessed based on the difference metric. During the assessment, a database of historical difference metrics is maintained, storing the historical statistical distribution of difference metrics for each modality during the health period of the industrial IoT device and under various known degradation modes. For each set of difference metrics most recently calculated within the current time window, it is compared with the historical statistical distribution corresponding to the mode. Its uncertainty is quantified by calculating the probability of anomaly or statistical distance within the historical statistical distribution. For example, assuming the mean and standard deviation of the historical difference metrics follow a multidimensional Gaussian distribution, the currently observed difference metric can be converted into an uncertainty quantification score by calculating the Mahalanobis distance between it and the mean of the historical statistical distribution, or by calculating the probability that it falls outside the confidence interval of the historical statistical distribution. The higher the calculated uncertainty quantification score, the more anomalous and unpredictable the currently observed difference is compared to the historical normal mode.

[0074] The above calculations are performed independently for each mode to obtain the uncertainty quantification score corresponding to each mode. The uncertainty quantification scores of all modes are arranged in a fixed order to form an uncertainty quantification vector. By transforming the difference measure representing the magnitude of the deviation into a probabilistic index representing "the degree of abnormality or unreliability of this deviation", the concept of statistical confidence is introduced into the state assessment. This enables the monitoring system to distinguish between situations where the numerical deviation is large but belongs to normal operating condition fluctuations (low uncertainty) and situations where the numerical deviation seems small but is extremely abnormal and indicates potential failure (high uncertainty). This provides a key probabilistic statistical basis for finally generating a state performance credibility that combines information on the magnitude of the deviation and the degree of abnormality.

[0075] Step S24: Perform weighted calculation on the uncertainty quantization vector to obtain the state performance credibility.

[0076] A weight vector with the same dimension as the uncertainty quantization vector is preset. Each weight coefficient in the weight vector corresponds to a mode. The magnitude of the weight coefficient is based on the criticality of the corresponding mode to the overall performance of the industrial IoT device, the stability of its historical data, and the knowledge of domain experts that is preset or learned from historical data. For example, the mode related to control accuracy may be given a higher weight than the ambient temperature mode.

[0077] During weighted calculation, each uncertainty score in the uncertainty quantification vector is multiplied by its corresponding weight coefficient to obtain a set of weighted uncertainty scores. Then, all the calculated weighted scores are summed to obtain a total uncertainty score. To obtain a confidence score between 0 and 1 (typically 1 represents completely trustworthy and 0 represents completely untrustworthy), this total uncertainty score is processed using a standard Sigmoid function, typically in the form of... Then, the process is reversed to transform the total uncertainty score into a state performance confidence level. Therefore, it is necessary to construct a monotonically decreasing function of the total uncertainty score. ,in, and It is an adjustable parameter; parameter As a steepness factor, it controls the Sigmoid function in... The sensitivity or rate of change in the vicinity is typically set based on the distribution range of the total uncertainty score and historical experience. The larger the value, the higher the confidence level relative to uncertainty at the threshold. The more sensitive the parameters are to small changes in the vicinity. As a baseline offset, it is usually set based on the statistical average or a specific quantile of the total uncertainty score of industrial IoT devices in a healthy state in historical data. Its function is to align the total uncertainty score to the vicinity of the central region of the Sigmoid function.

[0078] During the calculation, the total uncertainty score obtained in the steps will be used. Substituting into the monotonically decreasing function, first calculate Obtain the offset relative to the baseline, and then multiply it by the steepness factor. Get input variables Then input variables Input Sigmoid function The calculation yields an intermediate value between 0 and 1. Finally, subtracting this intermediate value from 1 yields the final state performance reliability, which is also between 0 and 1. This enables the normalization of the total score of uncertainties without clear boundaries and mapping it into a credibility index with clear upper and lower bounds and clear physical meaning. This allows the monitoring system to efficiently determine whether the overall status of the current industrial IoT equipment is healthy and whether it is necessary to immediately trigger the subsequent degradation cause diagnosis process based on this single value.

[0079] In one feasible implementation, step S30 may specifically include steps S31 to S35: Step S31: When the state performance confidence level is lower than the preset confidence level threshold, it is determined that the industrial IoT device has state performance degradation, and the diagnostic process is triggered.

[0080] It should be noted that the preset confidence threshold is a parameter that is dynamically set based on the historical operating stage of the industrial IoT device, the criticality of the current task, and the maintenance strategy. For example, a stricter threshold is used when the industrial IoT device is performing high-precision machining tasks, while a relatively lenient threshold is used during normal operation.

[0081] When the monitoring system detects that the current state performance confidence value is consistently lower than or momentarily drops to a preset confidence threshold, it determines the time point when this state occurs as the performance degradation judgment moment and generates a performance degradation judgment signal. This judgment signal serves as a trigger condition, automatically activating the diagnostic process scheduler in the monitoring system that is in a standby state. The diagnostic process scheduler then allocates computing resources for subsequent diagnostic processes, initializes the data context required for diagnosis (such as locking the current time point and backtracking to determine the degradation time window used for analysis), and starts the diagnostic job queue. This completes the switch from state assessment mode to root cause diagnosis mode, enabling the monitoring system to initiate relatively resource-intensive in-depth diagnostic analysis only when truly necessary. This achieves intelligent allocation of monitoring system resources, avoids the computational burden caused by continuously running full-scale diagnostics, and ensures that diagnostic conclusions always target the most noteworthy performance degradation events, improving the decision-making efficiency and economy of the entire monitoring system.

[0082] Step S32: According to the diagnostic process, the multimodal time series data is placed into the degradation time window, and the abnormal fluctuation patterns of the multimodal time series data in the degradation time window are extracted to obtain a candidate abnormal feature set.

[0083] It should be noted that the degradation time window is a continuous period of time that extends backward in the historical direction from the time when performance degradation is judged. Its length can be fixed or dynamically determined according to the rate of decline in the reliability of state performance. Its purpose is to define the critical period when the performance of industrial IoT devices begins to deteriorate noticeably, thereby focusing the analysis on the most relevant multimodal feature segments.

[0084] Operationally, the monitoring system extracts all multimodal feature segment sequences corresponding to the degradation time window from the stored multimodal time series data according to a defined time range. These sequences constitute the data placed within the degradation time window. Next, abnormal fluctuation patterns within the data are extracted. Specifically, multi-scale wavelet decomposition is performed on the feature sequence sequences of each modality within the degradation time window to obtain coefficient sequences representing fluctuation components at different time scales. Then, the statistical properties of the coefficient sequences at each time scale, such as mean and variance, are calculated and compared with the baseline statistical distribution of the corresponding modality at the corresponding scale under the historical healthy window. Mahalanobis distance is calculated to identify fluctuation components that significantly deviate at a specific time scale. All fluctuation components marked as significantly deviating (e.g., an abnormal increase in energy of a modality at a time scale of 0.5-1 seconds), along with their corresponding feature dimensions, timestamps, and deviation strengths, are extracted and structured into an abnormal fluctuation pattern.

[0085] Finally, all abnormal fluctuation patterns extracted from all modalities and time scales within the degradation time window are collected to form a candidate abnormal feature set, thereby condensing massive amounts of multimodal time series data into an abnormal evidence set, providing direct and clear information for subsequent steps, and greatly improving the accuracy and efficiency of root cause diagnosis.

[0086] Step S33: Based on the preset device degradation knowledge graph, perform graph matching and causal reasoning on the candidate abnormal feature set to obtain an initial diagnosis list.

[0087] It should be noted that the preset device degradation knowledge graph is a domain knowledge base organized in the form of a graph structure. Its nodes represent specific components of industrial IoT devices, specific states of components (such as "bearing - normal wear", "motor - overheating", etc.), and observable physical characteristics (such as "vibration spectrum energy increases at 1kHz"). Its edges represent the relationships between nodes, including hierarchical relationships (such as "motor - contains - bearing") and causal reasoning relationships (such as "poor lubrication - leads to -> bearing overheating - leads to -> vibration spectrum energy increases at 1kHz").

[0088] During graph matching, each abnormal fluctuation pattern in the candidate abnormal feature set is compared with the nodes representing "observable physical features" in the preset device degradation knowledge graph. For example, cosine similarity is calculated. If the similarity exceeds a certain threshold, the abnormal fluctuation pattern is considered to have been successfully located in the preset device degradation knowledge graph.

[0089] Then, causal reasoning is performed. Starting from each successfully matched node, a reverse traversal is performed along the directed causal edges in the preset device degradation knowledge graph, tracing upstream cause nodes that may lead to this abnormal fluctuation pattern layer by layer. For example, starting from the node "vibration spectrum energy increases at 1kHz", the tracing proceeds along the "cause" edge to the "bearing overheating" node, and further to the "poor lubrication" node, collecting all traced upstream cause nodes. During this process, a preliminary confidence score is calculated for each possible cause based on the edge weights, path lengths, and matching similarities in the preset device degradation knowledge graph. Specifically: ,in, It is a decay factor between 0 and 1.

[0090] The similarity between candidate anomaly features and nodes in a predefined device degradation knowledge graph is used as the initial base value for confidence calculation. During backward reasoning along causal edges, the current confidence value is multiplied by a predefined weight for each directed edge traversed (the weight is typically set based on the statistical probability of the causal relationship in historical data or expert experience, and ranges from 0 to 1). This represents the confidence loss due to the uncertainty of the causal relationship. Simultaneously, the path length increases with each edge traversed, introducing a global path decay factor. The path length raised to the power of (e.g.) =0.9) further decays the current value, which reflects the general logic that the longer the reasoning chain and the more indirect the hypothesis, the lower its credibility. When the reasoning reaches a possible cause node, the final value obtained after the above continuous multiplication operation is assigned the confidence score of that possible cause node.

[0091] Finally, all the inferred, non-repeating possible causal nodes and their corresponding confidence scores were compiled into an initial diagnostic list, which narrowed the scope of manual investigation and provided a clear and efficient diagnostic direction for subsequent steps of data verification based on time-series causality.

[0092] Step S34: Analyze the temporal causal relationship between the multimodal time series data and each initial diagnostic cause in the initial diagnostic list, and output the reliability score of each initial diagnostic cause based on the analysis results.

[0093] For each initial diagnostic cause in the initial diagnostic list, based on the causal chain defined in the preset equipment degradation knowledge graph, the expected temporal relationship that the initial diagnostic cause is expected to lead to is clarified. For example, the initial diagnostic cause "insufficient bearing lubrication" is expected to first cause "a slow rise in temperature characteristics," and then lead to "a significant increase in high-frequency vibration energy characteristics." Then, feature time series directly related to the expected temporal relationship are extracted from multimodal time series data, especially within the degradation time window. Next, in the phase space formed by the two feature time series, the ability to predict the future state of one time series from the past state of the other is evaluated, thereby quantifying the causal strength of the temporal causal relationship between them and obtaining a statistical significance measure. Subsequently, the calculated causal strength and metric are compared with the expected temporal relationships obtained from the preset device degradation knowledge graph to calculate a temporal causal consistency score between 0 and 1. This temporal causal consistency score is then weighted and fused with the confidence score (e.g., weighted average) to generate a reliability score. This process is repeated to obtain the reliability score for each initial diagnostic cause, providing data evidence to support subsequent decisions.

[0094] Step S35: Sort and filter the reliability scores, and determine the initial diagnostic cause corresponding to the reliability score ranked first as the cause of degradation.

[0095] All reliability scores are sorted in descending order, with the initial diagnostic cause having the highest reliability score at the top of the list. The initial diagnostic cause with the highest reliability score is marked and output as the degradation cause determined in this diagnostic process. This provides the entire monitoring system with a clear, unambiguous diagnostic output that can directly guide subsequent maintenance or control actions, completing a complete intelligent diagnostic closed loop from the perception of abnormal status to the accurate location of the root cause.

[0096] In one feasible implementation, step S40 may specifically include steps S41 to S44: Step S41: Query the preset cause-compensation mapping knowledge base to obtain several compensation directions associated with the cause of degradation.

[0097] It should be noted that the pre-defined cause-compensation mapping knowledge base is a structured rules and relational database. It stores the mapping relationships between various known causes of degradation of industrial IoT devices and potential compensatory control measures, which are summarized and deduced from the experience of domain experts and historical maintenance data. Its storage form can be a production rule of "IF (cause) THEN (compensation direction)" or a "compensable as" relation edge between cause nodes and compensation action nodes in the knowledge graph.

[0098] During the query, the identified degradation cause is used as the input keyword or query node. Semantic similarity matching is performed in the preset cause-compensation mapping knowledge base. After finding the corresponding cause entry, all preset compensation directions directly associated with that entry are retrieved, such as IF “insufficient bearing lubrication” THEN “increase lubrication frequency”, “increase lubricant supply pressure”, and “reduce the load condition of the bearing position”. The retrieved preset compensation directions are output in list form, ensuring that the compensation measures proposed by the monitoring system are targeted and physically feasible, and avoiding blind or ineffective control attempts.

[0099] Step S42: Based on the real-time status of the equipment reflected by the compensation direction and multimodal time series data, calculate the adjustment parameters under each compensation direction to obtain several initial compensation strategies.

[0100] Each compensation direction is analyzed and decomposed into control variables and desired adjustment targets. For example, the compensation direction "increase lubrication frequency" is analyzed into the control variable "lubrication interval time" and the desired adjustment target "shorten this time". Then, the real-time status of the equipment is queried, and key status parameters directly related to the current compensation direction are extracted from the latest multimodal time-series data. For example, for the "increase lubrication frequency" direction, the current lubrication interval time, bearing temperature, and vibration amplitude are extracted as calculation benchmarks. Next, the specific adjustment amount to satisfy the compensation direction is calculated. For example, if the current bearing temperature exceeds the baseline by 10%, the lubrication interval time is shortened by 20%, or a suggested interval time value is directly output through preset adjustment rules. After the above process, each compensation direction generates an initial compensation strategy containing specific control variables, target setpoints, or adjustment ranges.

[0101] Step S43: Simulate each initial compensation strategy and predict the impact of each initial compensation strategy on the equipment performance indicators based on the simulation results, and obtain the performance score of each initial compensation strategy.

[0102] During simulation, a preset digital twin is invoked. The specific control variables included in each initial compensation strategy are used as input conditions for the preset digital twin, while the real-time state of the device is set to the initial state of the simulation. Subsequently, the preset digital twin is driven to run a future operating cycle in the virtual environment at a speed several times faster than in reality, simulating the dynamic response process of the industrial IoT device after the implementation of the corresponding initial compensation strategy.

[0103] During the simulation, the monitoring system continuously extracts equipment performance indicators, such as predicted values ​​of machining accuracy, total vibration, and energy efficiency, from the output of the preset digital twin. After the simulation, by comparing the changes in performance indicators before and after implementing the initial compensation strategy, such as the percentage reduction in total vibration or the improvement in the standard deviation of machining accuracy, and taking into account the execution cost (such as increased energy consumption) and potential risks (such as whether it causes other parameters to exceed limits), the system calculates the effectiveness score of each initial compensation strategy through a preset multi-objective evaluation function.

[0104] The preset multi-objective evaluation function defines a set of target dimensions that need to be optimized, such as performance improvement (e.g., percentage reduction in vibration), quality improvement (e.g., improvement in accuracy standard deviation), efficiency impact (e.g., energy consumption change rate), and risk coefficient (e.g., whether it is close to other safety thresholds). For each initial compensation strategy, the specific predicted value of the initial compensation strategy on each target dimension is extracted from the simulation results.

[0105] For positive benefit-oriented target dimensions, such as vibration reduction percentage, a baseline value (usually the simulation prediction value or historical health average without compensation strategy) and an ideal target value (possibly the theoretical optimum or the allowable best limit) are set for this type of target dimension. Then, the calculation is performed using the formula (specific predicted value - baseline value) / (ideal target value - baseline value). If the result is greater than 1, it is truncated to 1; if it is less than 0, it is truncated to 0, thus mapping the specific predicted value to the interval [0, 1]. 1 represents achieving the ideal target, and 0 represents no improvement. For negative cost-oriented target dimensions, such as energy consumption increase rate, the processing is similar but in the opposite direction. An acceptable upper limit value (such as the maximum allowable energy consumption increase) and an ideal value (usually 0 growth) are set. The calculation is performed using the formula (acceptable upper limit value - specific predicted value) / (acceptable upper limit value - ideal value). The result is also truncated to the interval [0, 1]. However, in this case, 1 represents the best cost control (i.e., no growth), and 0 represents reaching the cost upper limit, thus normalizing the specific predicted value.

[0106] Next, each normalized predicted value is multiplied by a weighting coefficient reflecting the importance of that objective dimension in the global evaluation to obtain a weighted score. This process is repeated for each predicted value. Finally, all weighted scores are summed sequentially, and the resulting sum is the performance score. The score directly represents the overall expected performance of the initial compensation strategy after balancing various indicators.

[0107] Step S44: Sort and filter the performance scores, and determine the initial compensation strategy corresponding to the performance score ranked first as the control strategy.

[0108] All performance scores are sorted in descending order, with the initial compensation strategy having the highest performance score at the top of the list. This strategy is then marked and output as the control strategy. This process selects the optimal solution from multiple feasible initial compensation schemes, providing an optimal decision output for the entire industrial IoT device. This completes a full intelligent closed loop from state perception and degradation diagnosis to strategy generation and optimization, ensuring that the final control strategy is the best solution after thorough verification and comparison. This achieves a fundamental shift in equipment maintenance from passive response to proactive optimization and control.

[0109] This application also provides an industrial IoT device status monitoring system, referring to... Figure 2 As shown, the industrial IoT device status monitoring system includes: Data acquisition module 10 is used to acquire multimodal time-series data during the operation of industrial IoT devices; The data analysis module 20 is used to compare and analyze the multimodal time series data with the expected multimodal time series data to obtain the reliability of the state performance of industrial IoT devices. The data diagnostic module 30 is used to diagnose the cause of state performance degradation based on multimodal time series data if state performance degradation of industrial IoT devices is detected based on state performance reliability monitoring. The strategy output module 40 is used to generate a control strategy for compensating and adjusting the operating parameters of industrial IoT devices based on the cause of degradation.

[0110] Optionally, the data acquisition module 10 is also used for: Physical sensing signals are acquired through physical sensors, and control status signals are synchronously acquired with the industrial bus through the controller interface on the controller. Interpolation and synchronization processing are performed on physical sensor signals and control status signals to obtain a synchronization signal stream; The frequency domain features of the physical sensing signals are extracted from the synchronization signal stream to form the physical feature vector, and the instruction codes, status codes and numerical sequences of the control state signals are extracted to form the control state feature vector. Based on the equipment process logic relationship between physical feature vectors and control state feature vectors, multimodal feature segments are obtained by fusing them together. Multimodal feature segments are spliced ​​together according to the timeline to obtain multimodal time series data.

[0111] Optionally, the data analysis module 20 is also used for: Multimodal time series data is input into a preset digital twin, and the expected multimodal time series data is obtained by deducing from the preset digital twin; The time-frequency feature differences between multimodal time-series data and expected multimodal time-series data are calculated to obtain a difference metric. Based on the statistical distribution of the difference metric over the time window, the uncertainty of the current device state of an industrial IoT device compared to the expected device state is assessed, and an uncertainty quantification vector is obtained. The state performance reliability is obtained by weighting the uncertainty quantization vector.

[0112] Optionally, the data analysis module 20 is also used for: The multimodal time series data and the expected multimodal time series data are aligned in the time domain to obtain the real-time sequence and the expected sequence; Real-time time-domain features and real-time frequency-domain features are extracted from the real-time sequence, and expected time-domain features and expected frequency-domain features are extracted from the expected sequence. The time warping distance between the real-time time-domain features and the expected time-domain features is calculated to obtain the time-domain difference component, and the cosine similarity between the real-time frequency-domain features and the expected frequency-domain features is calculated to obtain the frequency-domain difference component. By using a pre-set adaptive attention weighting network, the weights of the time-domain difference components and the frequency-domain difference components are adjusted respectively to obtain the difference scores of each modality. The difference scores of each modality are aggregated along the time dimension on a preset sliding window, and the mean and standard deviation of the difference scores of each modality on the preset sliding window are calculated to obtain the difference measure.

[0113] Optionally, the data diagnostic module 30 is also used for: When the state performance confidence level is lower than the preset confidence level threshold, it is determined that the industrial IoT device has state performance degradation, triggering the diagnostic process. According to the diagnostic process, multimodal time series data are placed into a degradation time window, and abnormal fluctuation patterns of multimodal time series data in the degradation time window are extracted to obtain a candidate abnormal feature set; Based on the preset equipment degradation knowledge graph, graph matching and causal reasoning are performed on the candidate abnormal feature set to obtain an initial diagnosis list; Analyze the temporal causal relationship between multimodal time series data and each initial diagnostic cause in the initial diagnostic list, and output a reliability score for each initial diagnostic cause based on the analysis results; The reliability scores are sorted and filtered, and the initial diagnostic cause corresponding to the reliability score ranked first is determined as the cause of degradation.

[0114] Optionally, the policy output module 40 is also used for: Query the preset cause-compensation mapping knowledge base to obtain several compensation directions associated with the cause of degradation; Based on the real-time status of the equipment reflected by the compensation direction and multimodal time series data, the adjustment parameters under each compensation direction are calculated to obtain several initial compensation strategies. Simulations were performed on each initial compensation strategy, and the impact of each initial compensation strategy on equipment performance indicators was predicted based on the simulation results, thus obtaining the performance score of each initial compensation strategy. The performance scores are sorted and filtered, and the initial compensation strategy corresponding to the performance score ranked first is determined as the control strategy.

[0115] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An industrial internet of things device state monitoring method, characterized by, The method comprises the following steps: acquiring multi-modal time series data of an industrial Internet of Things device in operation; comparing the multi-modal time series data with expected multi-modal time series data to obtain a state performance credibility of the industrial Internet of Things device; if state performance degradation of the industrial Internet of Things device is monitored based on the state performance credibility, diagnosing a degradation cause of the state performance degradation according to the multi-modal time series data; generating a control strategy for compensating and adjusting a device operation parameter of the industrial Internet of Things device according to the degradation cause.

2. The industrial internet of things device state monitoring method of claim 1, wherein, The industrial Internet of Things device is provided with a physical sensor and a controller, and the step of acquiring multi-modal time series data of an industrial Internet of Things device in operation comprises the following steps: collecting a physical sensing signal through the physical sensor and collecting a control state signal through a controller interface of the controller and synchronizing the control state signal with an industrial bus; performing interpolation synchronization processing on the physical sensing signal and the control state signal to obtain a synchronized signal stream; extracting a frequency domain feature of the physical sensing signal to form a physical feature vector and extracting an instruction code, a state code and a numerical sequence of the control state signal to form a control state feature vector from the synchronized signal stream; fusing the physical feature vector and the control state feature vector according to a device process logic relationship to obtain a multi-modal feature segment; splicing the multi-modal feature segment according to a time axis sequence to obtain the multi-modal time series data.

3. The industrial IoT device state monitoring method of claim 1, wherein, The step of comparing the multi-modal time series data with expected multi-modal time series data to obtain a state performance credibility of the industrial Internet of Things device comprises the following steps: inputting the multi-modal time series data into a preset digital twin to obtain the expected multi-modal time series data through deduction of the preset digital twin; performing time-frequency feature difference calculation on the multi-modal time series data and the expected multi-modal time series data to obtain a difference measure value; evaluating an uncertainty of a current device state of the industrial Internet of Things device compared with an expected device state according to a statistical distribution of the difference measure value on a time window to obtain an uncertainty quantization vector; performing weighted calculation on the uncertainty quantization vector to obtain the state performance credibility.

4. The industrial internet of things device state monitoring method of claim 3, wherein, The step of performing time-frequency feature difference calculation on the multi-modal time series data and the expected multi-modal time series data to obtain a difference measure value comprises the following steps: performing time domain alignment on the multi-modal time series data and the expected multi-modal time series data to obtain a real-time sequence and an expected sequence; extracting a real-time time domain feature and a real-time frequency domain feature from the real-time sequence and an expected time domain feature and an expected frequency domain feature from the expected sequence, calculating a time regularity distance between the real-time time domain feature and the expected time domain feature to obtain a time domain difference component, and calculating a cosine similarity between the real-time frequency domain feature and the expected frequency domain feature to obtain a frequency domain difference component; adjusting a weight of the time domain difference component and a weight of the frequency domain difference component through a preset adaptive attention weighting network to obtain a difference score of each modality. The difference scores of each modality are aggregated along the time dimension on a preset sliding window, and the mean and standard deviation of the difference scores of each modality on the preset sliding window are calculated to obtain the difference metric.

5. The industrial internet of things device state monitoring method of claim 1, wherein, If state performance degradation is detected in the industrial IoT device based on the state performance reliability, the step of diagnosing the cause of state performance degradation based on the multimodal time series data includes: When the reliability of the state performance is lower than a preset reliability threshold, it is determined that the industrial IoT device has state performance degradation, triggering a diagnostic process; According to the diagnostic process, the multimodal time series data is placed into a degradation time window, and the abnormal fluctuation patterns of the multimodal time series data in the degradation time window are extracted to obtain a candidate abnormal feature set. Based on a preset device degradation knowledge graph, graph matching and causal reasoning are performed on the candidate abnormal feature set to obtain an initial diagnosis list; Analyze the temporal causal relationship between the multimodal time series data and each initial diagnosis cause in the initial diagnosis list, and output a reliability score for each initial diagnosis cause based on the analysis results; The reliability scores are sorted and filtered, and the initial diagnostic cause corresponding to the reliability score ranked first is determined as the cause of degradation.

6. The industrial internet of things device state monitoring method of claim 1, wherein, The step of generating a control strategy for compensating and adjusting the operating parameters of the industrial IoT device based on the cause of degradation includes: Query the preset cause-compensation mapping knowledge base to obtain several compensation directions associated with the degradation cause; Based on the compensation direction and the real-time state of the device reflected by the multimodal time series data, the adjustment parameters under each compensation direction are calculated to obtain several initial compensation strategies. Simulations are performed on each of the initial compensation strategies, and the impact of each initial compensation strategy on the equipment performance indicators is predicted based on the simulation results, so as to obtain the performance score of each initial compensation strategy. The performance scores are sorted and filtered, and the initial compensation strategy corresponding to the performance score ranked first is determined as the control strategy.

7. An industrial internet of things device state monitoring system, characterized by, include: The data acquisition module is used to acquire multimodal time-series data during the operation of industrial IoT devices; The data analysis module is used to compare and analyze the multimodal time series data with the expected multimodal time series data to obtain the state performance reliability of the industrial IoT device. The data diagnostic module is used to diagnose the cause of state performance degradation based on the multimodal time series data if state performance degradation of the industrial IoT device is detected based on the state performance reliability. The strategy output module is used to generate a control strategy for compensating and adjusting the operating parameters of the industrial IoT device based on the degradation cause.

Images