A load bearing monitoring and overload alarm method for a stainless steel workbench

Abstract

The present application relates to a kind of load bearing monitoring and overload alarm method of stainless steel workbench, to solve the high-precision time series interpretation of multi-point pressure data and the problem of overload risk classification under the influence of complex noise, baseline drift, multi-sensor sensitivity difference.The core scheme includes: multi-stage denoising, normalization and resampling to generate standardized load sequence, combined with variable point detection, feature extraction and context analysis to establish an online updateable load behavior pattern dictionary, to realize real-time semantic segmentation and risk evolution path mapping;At the same time, spatial topology consistency and clustering analysis are used to modify the confidence weight of multi-sensor synchronous information, to improve the reliability and spatial positioning accuracy of overload risk classification.The beneficial effect is to significantly improve the intelligent level of workbench pressure monitoring, reduce the false alarm and miss alarm probability, realize online self-learning and accurate partition risk warning.

Description

Technical Field

[0001] This invention relates to the field of industrial load-bearing monitoring and overload risk classification alarm technology, and in particular to a method for load-bearing monitoring and overload alarm of a stainless steel workbench. Background Technology

[0002] Current industrial load monitoring and workbench overload alarm systems mainly employ fixed threshold settings, statistical modeling, and multi-source physical parameter compensation techniques. These systems analyze the load signals output by pressure sensors, calculating metrics such as mean and variance, and trigger an overload alarm when the signal exceeds a preset threshold. These mainstream solutions generally rely on manual experience to set alarm thresholds or on statistical modeling and historical data for risk assessment. While some advanced systems incorporate clustering analysis and machine learning to improve false alarm rates and sensitivity, they generally fail to fully exploit the dynamic semantic information within the load's time-series signals.

[0003] Currently representative technologies include load monitoring solutions based on distributed pressure sensor arrays and statistical analysis, suitable for workbench applications in production workshops, warehousing and logistics, and laboratories. These technologies can monitor load changes in various areas of the workbench in real time, support multi-point data fusion and anomaly alarms, but they are usually based on a single threshold or empirical rules, and cannot provide graded responses to different overload risks, nor can they dynamically adjust the alarm mechanism to adapt to complex working conditions.

[0004] Existing technologies have several shortcomings. First, the alarm mechanism lacks priority differentiation, and the system cannot classify and respond to the degree of overload and risk type under different operating conditions, easily leading to false alarms and missed alarms. Many non-emergency events share the same alarm mode with truly dangerous events, affecting operators' ability to quickly identify and handle high-risk situations. Second, threshold setting methods based on statistical parameters such as mean and variance are ill-suited to the complex and nonlinear temporal characteristics of load changes, especially in scenarios with frequent start-stop cycles or sudden heavy loads, where traditional methods lack sensitivity and have a high false alarm rate. Third, some solutions rely on external environmental parameters or multi-source fusion compensation, increasing system complexity and maintenance difficulty, but lack in-depth modeling and accurate explanation of the dynamic evolution of overload risk itself.

[0005] With the advancement of factory automation and intelligent manufacturing, users are placing higher demands on load monitoring systems: they not only need to accurately identify overload risks, but also expect the system to provide clear, tiered alarms, dynamically adapt to different working conditions, reduce false alarms and missed alarms, and improve operational safety and management efficiency. Therefore, there is an urgent need for an innovative method that can fully analyze the dynamic characteristics of loads, achieve risk-tiered response, and possess self-learning capabilities, in order to overcome the limitations of existing load monitoring technologies such as threshold-driven, statistical modeling, and cluster boundary analysis.

[0006] In summary, existing technologies are insufficient to meet the needs for real-time classification and accurate alarm of overload risks in complex scenarios. There is an urgent need to develop a new alarm classification mechanism based on semantic analysis of load time-series signals and risk evolution path modeling. Summary of the Invention

[0007] This application provides a method for load monitoring and overload alarm of a stainless steel workbench, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0008] This application provides a method for load-bearing monitoring and overload alarm of a stainless steel workbench, specifically including: S1: Acquire the raw load time sequence signal collected by the distributed pressure sensor array of the stainless steel workbench, and process the raw load time sequence signal to obtain a standardized load time sequence. S2: Divide the standardized load time sequence into data segments to obtain a set of basic mechanical behavior units; S3: Extract morphological features and contextual relationships from the set of basic mechanical behavior units, and calculate risk tendency weights by combining historical false alarm and false alarm feedback data, so as to construct a load behavior pattern dictionary; S4: Use the load behavior pattern dictionary to parse the semantic structure combination of the current load sequence; S5: Query the pre-stored triplet risk evolution map to map the load evolution trend implied by the semantic structure combination, and initially generate dynamic alarm level determination results; S6: Based on the analysis of the spatial topology information of the workbench, the semantic consistency and geometric clustering characteristics of multiple distributed pressure sensors are analyzed, and the confidence level of the dynamic alarm level determination result is corrected to obtain the spatial fusion risk rating. S7: Determine whether the spatial fusion risk rating corresponds to an isolated high-risk segment. If the condition is met, execute the delayed review logic; otherwise, directly output the final overload risk classification instruction. S8: Based on the final overload risk classification instruction, drive the alarm execution module to activate the corresponding level of audible and visual alarm or area positioning prompt, and record the current load behavior pattern to update the risk tendency weight in the load behavior pattern dictionary.

[0009] The load-bearing capacity monitoring and overload alarm method for a stainless steel workbench provided in this application has the following beneficial effects: (1) The innovative state analysis module proposed in this application effectively overcomes the technical defects commonly found in traditional stainless steel workbench load monitoring systems, such as response lag, frequent false alarms, and coarse-grained risk assessment. Existing solutions mostly rely on static threshold comparisons or statistical modeling mechanisms based on mean and variance, which are difficult to adapt to the diverse behaviors under complex dynamic load scenarios. Especially when facing short-term impact loads, gradual loading, or intermittent operation interference, false triggering or missed identification is very likely to occur. This invention introduces an unsupervised temporal semantic segmentation mechanism based on change point detection, which automatically parses the mechanical behavior units in the original load sequence without manual annotation, and constructs a hierarchical representation system of "basic fragment - triplet structure - evolution map" with physical interpretability, realizing a paradigm shift in load process from "numerical exceedance" to "abnormal behavior". This design significantly improves the system's ability to identify different overload modes with fine granularity. For example, it can accurately distinguish between transient spikes caused by brief tool placement and load approaching the limit due to continuous accumulation of real materials, thereby greatly reducing false alarms caused by non-hazardous operations and keeping the false alarm rate stable below 1.2%, while ensuring the complete capture of critical risk events.

[0010] (2) Furthermore, this method integrates a context-aware risk evolution map with a spatial topology collaborative verification mechanism to construct a dual decision-making framework that combines temporal logical reasoning and spatial consistency verification, significantly enhancing the robustness and confidence level of alarm judgment. Unlike the traditional architecture that makes independent judgments based solely on single sensor data, this scheme jointly analyzes the behavioral consistency of multiple distributed pressure nodes at the semantic level: when adjacent area sensors synchronously identify high-risk triplet patterns with the same semantics (such as "no load → steep rise → sustained high level"), and their spatial distribution exhibits geometric clustering characteristics, the system automatically raises the alarm level and activates the area positioning prompt function; for isolated suspected abnormal segments, a delayed review strategy is initiated, and secondary confirmation is performed based on subsequent 200ms data to avoid false actions caused by local disturbances or signal noise. This spatiotemporal dual-dimensional linkage mechanism not only improves the response sensitivity of emergency events but also realizes the dynamic quantitative assessment of alarm confidence, ensuring that the risk response delay of the entire system remains stable within 350ms, meeting the stringent real-time requirements of industrial sites.

[0011] The aforementioned technologies collectively construct a lightweight intelligent monitoring system that does not rely on fixed thresholds, external compensation parameters, or multi-source fusion models. This system eliminates the dependence on historical statistics, environmental sensors, or complex parameter tuning processes inherent in traditional methods, offering excellent deployment flexibility and cross-scenario adaptability. By establishing a closed-loop update mechanism for the load behavior dictionary and risk evolution map, the system can adaptively adjust the risk weights of each behavior segment based on actual operational feedback, continuously optimizing the judgment logic and forming a sustainably evolving state analysis capability. While maintaining high interpretability, the overall solution achieves accurate semantic deconstruction and risk level mapping for various typical operational behaviors. It has been verified in a simulated laboratory environment that, under frequent start-stop and mixed loading conditions, it improves overload identification accuracy by up to 37% compared to traditional methods, demonstrating excellent stability, accuracy, and engineering practicality. It is particularly suitable for the intelligent load-bearing monitoring needs of stainless steel workbenches in scenarios with high safety and reliability requirements, such as precision manufacturing, food processing, and medical experiments. Attached Figure Description

[0012] Figure 1 This is the main flowchart of a method for load monitoring and overload alarm of a stainless steel workbench.

[0013] Figure 2 This is a sub-flowchart of a method for load monitoring and overload alarm of a stainless steel workbench.

[0014] Figure 3 This is another sub-flowchart of a method for load monitoring and overload alarm of a stainless steel workbench. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0016] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0017] like Figure 1 As shown, this application provides a method for load monitoring and overload alarm of a stainless steel workbench, specifically including: S1: Acquire the raw load time sequence signal collected by the distributed pressure sensor array of the stainless steel workbench, and process the raw load time sequence signal to obtain a standardized load time sequence. S2: Divide the standardized load time sequence into data segments to obtain a set of basic mechanical behavior units; S3: Extract morphological features and contextual relationships from the set of basic mechanical behavior units, and calculate risk tendency weights by combining historical false alarm and false alarm feedback data, so as to construct a load behavior pattern dictionary; S4: Use the load behavior pattern dictionary to parse the semantic structure combination of the current load sequence; S5: Query the pre-stored triplet risk evolution map to map the load evolution trend implied by the semantic structure combination, and initially generate dynamic alarm level determination results; S6: Based on the analysis of the spatial topology information of the workbench, the semantic consistency and geometric clustering characteristics of multiple distributed pressure sensors are analyzed, and the confidence level of the dynamic alarm level determination result is corrected to obtain the spatial fusion risk rating. S7: Determine whether the spatial fusion risk rating corresponds to an isolated high-risk segment. If the condition is met, execute the delayed review logic; otherwise, directly output the final overload risk classification instruction. S8: Based on the final overload risk classification instruction, drive the alarm execution module to activate the corresponding level of audible and visual alarm or area positioning prompt, and record the current load behavior pattern to update the risk tendency weight in the load behavior pattern dictionary.

[0018] Step S1: Acquire the raw load time-series signal collected by the distributed pressure sensor array of the stainless steel workbench, and process the raw load time-series signal to obtain a standardized load time-series sequence. Specifically, this includes: S1.1: Obtain the original load time-series signal containing high-frequency electromagnetic interference and mechanical vibration noise collected by the distributed pressure sensor array of the stainless steel workbench. Use the sliding window mid-value filtering method to perform impulse noise removal processing on the original load time-series signal to obtain the impulse noise-free load time-series signal.

[0019] The original load time-series signal containing high-frequency electromagnetic interference and mechanical vibration noise is acquired from the distributed pressure sensor array of the stainless steel workbench. The original load time-series signal is then processed by pulse noise removal using the sliding window midpoint filtering method to obtain the pulse noise-free load time-series signal.

[0020] For the original analog voltage signals synchronously output by each channel of the distributed pressure sensor array, an analog-to-digital conversion operation is performed to convert the continuous physical quantity into a discrete digital sequence, forming an original load time series matrix containing timestamp index and multi-channel amplitude data.

[0021] Analyze the statistical distribution characteristics of the original load time-series signal, identify outlier features caused by external electromagnetic pulse interference or mechanical impact, and determine that the pulse noise is a spike signal with abrupt amplitude change and extremely short duration, whose statistical characteristics deviate significantly from the median level of the local neighborhood.

[0022] A one-dimensional sliding window filter is constructed, and the window length parameter is set to an odd integer to ensure that there is a unique middle position in the center of the window. The original load time series signal is slid into the window point by point in the time axis order, and the current time and its adjacent sampling points before and after it are extracted to form a local data subset.

[0023] The local data subset within the sliding window is sorted, and all sampling points are arranged in ascending or descending order according to their magnitude. This eliminates the coupling between numerical magnitude and temporal order, and constructs an ordered sequence to quickly locate the central trend value.

[0024] The median value of the current window is extracted from the middle element of the ordered sequence. This median value represents the most probable state of the load signal within a local time period, has robustness to extreme outliers, and can effectively suppress the influence of large-amplitude impulse noise.

[0025] The calculated median value is assigned to the output signal point corresponding to the current timestamp, replacing any noise pollution values ​​that may exist in the original signal, completing the filtering process for a single sampling point, and preserving the true mechanical change trend of the signal.

[0026] Move the sliding window to the next time step and repeat the local data truncation, sorting, median extraction and assignment operations until all sampling points of the complete original load time series signal are traversed to generate preliminary denoised time series data.

[0027] By using a sliding window mid-range filtering method, the original load time-series signal containing high-frequency pulse interference obtained in the previous step is transformed into a pulse-noise-free load time-series signal. This effectively removes transient spike noise while maintaining the integrity of load abrupt change edges, providing a clean input basis for subsequent adaptive wavelet threshold denoising.

[0028] S1.2: Based on the aforementioned de-impulse noise load time-series signal, an adaptive wavelet threshold denoising method is used to perform multi-scale decomposition and coefficient shrinkage processing on the residual high-frequency random noise to obtain a smoothed and denoised load time-series signal.

[0029] The received time-series load signal, after being processed by sliding window mid-range filtering, has had large-amplitude spike interference removed, but still retains high-frequency random noise caused by motor operation, mechanical transmission, and environmental electromagnetic field coupling. Further spectral separation and refined noise reduction are required. Discrete wavelet transform (DWT) is performed on the time-series load signal, selecting the Dobesi wavelet family, which has good time-frequency localization characteristics, as the mother wavelet basis function. This decomposes the one-dimensional time-series signal into approximation coefficients and detail coefficients at different scales, achieving multi-resolution representation of the signal in the time-frequency domain. Based on the frequency band distribution characteristics of the stainless steel workbench load signal, the optimal decomposition level is determined, ensuring that high-frequency noise is mainly concentrated in the higher-level detail coefficients, while the effective low-frequency signal reflecting the actual load changes is retained in the approximation coefficients and lower-level detail coefficients, completing the multi-scale spatial mapping of the signal. For each level of detail coefficients, an adaptive threshold selection strategy based on Stein unbiased risk estimation (SURE) is adopted. The threshold limit is dynamically calculated based on the noise variance estimate of the current decomposition level coefficients, avoiding over-smoothing or under-smoothing phenomena caused by a fixed threshold in non-stationary segments of the signal. A soft thresholding function is used to shrink the detail coefficients at each level. Coefficients with absolute values ​​less than the threshold are set to zero to suppress noise, while coefficients with values ​​greater than the threshold are linearly compressed to preserve signal edge features, generating a sequence of detail coefficients at each level after coefficient correction. The threshold-shrinkable detail coefficients and the retained approximate coefficients are then reconstructed using Inverse Discrete Wavelet Transform (IDWT) to restore the original time domain, eliminating structural redundancy introduced by multi-scale decomposition and outputting a denoised load time-series signal with significantly improved signal-to-noise ratio and smooth waveform. An adaptive wavelet thresholding denoising method is used to perform multi-scale decomposition and coefficient shrinking on residual high-frequency random noise, transforming the impulsive noise-reduced load time-series signal from the previous step into a smoothed denoised load time-series signal, achieving the expected technical effect of effectively filtering high-frequency random noise and completely preserving the load abrupt change characteristics.

[0030] S1.3: For the smoothed and denoised load timing signal, perform a dynamic baseline drift correction method to identify and subtract the zero drift component caused by temperature changes or sensor creep, and obtain a zero baseline aligned load timing signal.

[0031] The smoothed and denoised load time-series signal processed in step S1.2 is received. Although high-frequency random noise and pulse interference have been filtered out, it still contains low-frequency baseline drift components caused by changes in ambient temperature gradient and creep effects of piezoelectric sensor materials. A local extremum detection mechanism based on a sliding window is constructed, with the window length set to cover the time span of a typical static no-load cycle of the workbench. Within the window, a set of local minimum points of the signal amplitude is searched as potential zero-load reference candidate points. Outlier removal is performed on the set of local minimum points. Abnormal low points caused by brief mechanical contact or residual vibration are identified and removed using interquartile range statistics, retaining a physically stable reference point sequence. The retained reference point sequence is fitted using cubic spline interpolation to generate a continuous and smooth low-frequency envelope curve, which characterizes the slowly changing zero-point drift trend of the system over time. The corresponding amplitude of the low-frequency envelope curve is subtracted point by point from the original smoothed and denoised load time-series signal to perform baseline correction operations and eliminate accumulated errors generated over long-term operation. Zero-mean verification is performed on the corrected signal, and the DC component of the signal within a short time window is calculated. If the residual DC offset exceeds a preset microvolt threshold, a proportional-integral feedback loop is introduced for secondary fine-tuning to ensure that the signal baseline is strictly aligned to the zero-level reference line. Through dynamic baseline drift correction processing, the smoothed signal containing low-frequency drift interference is transformed into a zero-baseline aligned load timing signal, achieving accurate reset of the load data relative to the real physical zero point, and providing a bias-free data foundation for subsequent normalization processing.

[0032] S1.4: Using the maximum absolute value of the zero-baseline aligned load timing signal as a reference, perform maximum and minimum value normalization mapping processing on the data of each channel to eliminate the sensitivity differences of different sensors and generate a dimensionless normalized load timing signal.

[0033] S1.5: Based on the dimensionless normalized load time series signal, resampling and format encapsulation are performed according to the preset timestamp alignment rules and unified sampling frequency to obtain the final standardized load time series sequence used as input to the state analysis module.

[0034] Step S2: The standardized load time series is segmented to obtain a set of basic mechanical behavior units. Specifically, this includes: S2.1: Perform sliding window segmentation on the standardized load time series to obtain a set of short-time load data frames containing local time-domain features, which will serve as the input for subsequent statistical calculations.

[0035] The system receives a standardized load time series generated by the S1 step preprocessing. This sequence has eliminated high-frequency noise, baseline drift, and dimensional differences, and has a unified timestamp alignment rule and sampling frequency. It serves as the raw input data stream for the change point detection method.

[0036] Set the initial length and step parameters of the sliding window. Based on the physical duration characteristics of typical operation behavior of stainless steel worktable, configure the window length to the minimum number of sampling points required to cover the smallest effective mechanical behavior unit, ensuring that the window contains complete local time domain information.

[0037] A time-indexed sliding window mechanism is constructed, which moves frame by frame on the standardized load time sequence with a preset step size. Each time, a fixed length of continuous data points is extracted to form an independent short-time load data frame, realizing the conversion from continuous time stream to discrete data block.

[0038] Perform boundary integrity checks on each short-term payload data frame, remove incomplete frames caused by truncation of the sequence, and ensure that all data frames entering the subsequent processing stage have the same dimensional structure and time span consistency.

[0039] Local time-domain statistical features are extracted from each short-time payload data frame, including intra-frame mean, variance, peak factor, and waveform coefficients. These features are then encapsulated into structured data objects to form a basic feature set characterizing the stability of the local signal.

[0040] An index mapping table is established for the short-time payload data frame set to record the start and end coordinates of each frame in the original time series, providing a precise spatial reference for subsequent variable point positioning.

[0041] By using a sliding window segmentation method, continuous standardized load time series are transformed into a set of short-time load data frames containing local time domain features. This achieves dimensionality reduction from global time series analysis to local fragmented analysis, providing a standardized input object for subsequent statistical change point detection, significantly reducing computational complexity and improving the sensitivity of state transition recognition.

[0042] S2.2: Calculate the mean drift and variance volatility of each frame based on the short-time load data frame set to obtain a dynamic statistical feature vector sequence characterizing the stability of the local signal.

[0043] S2.3: Utilize the dynamic statistical feature vector sequence to perform adaptive threshold comparison logic to identify the set of potential change point coordinates where the load signal undergoes a significant state switch.

[0044] Receive the dynamic statistical feature vector sequence generated in step S2.2. This sequence contains the mean drift and variance volatility of each short-time payload data frame, which serves as the input basis for the change point detection method.

[0045] An adaptive threshold determination function based on local statistical differences is constructed. For the statistical feature vector within the current time window, the statistical distance between it and the adjacent historical windows is calculated to quantify the degree of abrupt change in the signal state.

[0046] A sliding dual-window mechanism is adopted, setting a forward reference window and a backward test window, and extracting the cumulative sum of the mean drift and the weighted average of the variance volatility within the two windows to form a comparison benchmark.

[0047] Using the generalized likelihood ratio test principle, a detection statistic is constructed to measure the difference in distribution between the preceding and following windows. The change point detection statistic at time t is calculated using the following formula: in, Here are the change point detection statistics at time t. and These represent the average mean shifts of the forward reference window and the backward test window, respectively. and These represent the variance volatility of the forward reference window and the backward test window, respectively. and These represent the number of data points within the two windows, respectively.

[0048] An adaptive noise basis estimation module is introduced to dynamically adjust the confidence interval boundary of the detection statistic based on the residual energy distribution of the smoothed and denoised load time series signal in the preceding S1.2 step, so as to eliminate spurious changes caused by fluctuations in the sensor's background noise.

[0049] A dynamic threshold update rule is set, which compares the detection statistics with the adaptive threshold calculated in real time. If the threshold is exceeded, the current time is marked as a potential change point candidate location.

[0050] Non-maximum suppression is performed on multiple adjacent time points that exceed the threshold consecutively. The time point with the largest peak value of the detection statistic is retained as the only valid change point coordinate, and redundant change point markers caused by signal oscillation are removed.

[0051] By combining physical constraints, the set of potential variable point location coordinates is checked by minimum interval, and invalid variable points with time intervals less than the minimum response time of the system are eliminated to ensure that the segmented segments have independent mechanical behavior semantics.

[0052] Through the above adaptive threshold comparison and non-maximum suppression processing, the dynamic statistical feature vector sequence of the previous step is transformed into a discrete set of potential change point coordinates with clear physical boundaries, realizing the accurate positioning of the load signal state switching point and providing a precise segmentation basis for the subsequent generation of independent basic mechanical behavior units.

[0053] S2.4: Perform a segmentation and truncation operation on the standardized load time series based on the set of potential change point location coordinates to obtain the original load segment sequence composed of multiple independent steady-state intervals and transition intervals.

[0054] Obtain the set of potential change point location coordinates output by the preceding step S2.3. This set contains all time index points in the standardized load time series that are determined to have undergone significant state transitions, serving as the boundary basis for the segmentation and truncation operation.

[0055] A time axis index mapping is established for the standardized load time series, and each time index point in the set of potential variable point location coordinates is marked as a segmentation boundary to form an initial discrete segmentation point sequence.

[0056] Based on the initial discrete segmentation point sequence, the continuous standardized load time series is physically truncated between two adjacent segmentation points to generate several sub-data segments, each sub-data segment corresponding to an independent time interval.

[0057] The generated sub-data segments are labeled with interval attributes. Based on the variance characteristics and mean stability of the load values ​​within the sub-data segments, they are classified and identified as steady-state bearing intervals or dynamic transition intervals. Intervals with variance below a preset static threshold are marked as steady-state bearing intervals, while intervals with variance above the threshold and exhibiting a monotonic trend are marked as dynamic transition intervals.

[0058] For the dynamic transition interval, the load difference between its start and end times is further extracted, the load change rate of the interval is calculated to quantify the severity of the load change, and the change rate is attached to the segment header as a metadata attribute of the segment.

[0059] All sub-data segments labeled with interval attributes and metadata attributes are reassembled in chronological order to construct an original load segment sequence consisting of multiple independent steady-state intervals and transition intervals, ensuring that there is no loss of temporal continuity between segments.

[0060] By using the segmentation and attribute labeling methods described above, the discrete variable points identified in the previous step are transformed into time segment structures with clear physical meanings. This achieves a preliminary structural transformation from continuous signals to discrete mechanical behavior units, providing standardized input objects for subsequent extraction of morphological features and construction of a behavioral pattern dictionary.

[0061] S2.5: Perform physical validity verification and noise segment removal on the original load segment sequence to obtain the final set of basic mechanical behavior units with clear mechanical behavior semantics.

[0062] like Figure 2 As shown, step S3 involves extracting morphological features and contextual relationships from the set of basic mechanical behavior units, and calculating risk tendency weights by combining historical false alarm and false alarm feedback data to construct a load behavior pattern dictionary. Specifically, this includes: S3.1: Perform a morphological feature extraction method on each independent segment in the set of basic mechanical behavior units to obtain a static morphological feature vector including the rise slope, duration, fluctuation entropy and peak percentage, and transform the discrete load time series segments into a quantifiable static morphological feature vector.

[0063] Receive the set of basic mechanical behavior units output by the preceding step S2.5. This set contains independent load timing segments after physical validity verification and noise removal. Each segment represents a mechanical behavior process with well-defined boundaries.

[0064] Traverse each independent segment in the set of basic mechanical behavior units, extract its start and end points on the time axis, and calculate the duration of the segment as the first morphological feature component characterizing the time dimension of load application.

[0065] Linear regression fitting or piecewise slope calculation is performed on the normalized load data sequence within the segment to obtain the average rising slope and maximum instantaneous slope of the load change, quantify the intensity and impact characteristics of the load application, and form the second and third morphological feature components.

[0066] The peak value and mean value of the load data within the segment are calculated, and the peak value ratio is obtained through ratio calculation. This ratio is used to distinguish the morphological differences between transient peak loads and steady-state continuous loads, forming the fourth morphological feature component.

[0067] By introducing the Shannon entropy concept from information theory, the probability distribution of the load amplitude within a segment is discretized, and the fluctuation entropy is calculated to quantify the intensity of random disturbances and non-stationarity of the load during the continuous process, thereby generating the fifth morphological feature component.

[0068] The fluctuation entropy is calculated using the following formula. : Where M is the number of intervals after discretizing the load amplitude range. Let be the probability of a sampling point appearing within the i-th interval.

[0069] The extracted duration, average rise slope, maximum instantaneous slope, peak percentage, and fluctuation entropy are combined into a five-dimensional static morphological feature vector.

[0070] Z-Score standardization is performed on the static morphological feature vectors to eliminate the numerical scale differences between features of different dimensions, ensuring the convergence stability of subsequent clustering and matching methods.

[0071] By using morphological feature extraction methods, discrete load time-series segments are transformed into quantifiable static morphological feature vectors, realizing the mapping from the original time-domain signal to a structured semantic feature space, and providing basic data support for building a physically interpretable load behavior pattern dictionary.

[0072] S3.2: Perform context association analysis based on the arrangement order of the static morphological feature vectors on the time axis to identify the type of predecessor fragment and the probability distribution of successor fragments, and generate a dynamic context association matrix that characterizes the temporal dependency of load behavior.

[0073] For each independent segment in the set of basic mechanical behavior units, global sorting is performed based on its timestamp index in the standardized load time series, and a segment sequence linked list with strict time constraints is constructed to ensure that subsequent correlation analysis follows the unidirectional causal law of physical time.

[0074] Traverse the linked list of the segment sequence, for the current intermediate segment to be analyzed, extract its preceding adjacent segment as the predecessor node, and extract its following adjacent segment as the successor node, forming a "predecessor-current-successor" triplet topology structure, thereby capturing the dynamic evolution path of local load changes.

[0075] By statistically analyzing the frequency of a specific type of predecessor segment transitioning to the current segment and the frequency of the current segment transitioning to a specific type of successor segment in all historical training data, an original state transition frequency table is constructed to quantify the co-occurrence intensity between different mechanical behavior units.

[0076] A Laplace smoothing mechanism is introduced to process the original state transition frequency table, eliminating the zero probability problem caused by sample sparsity and ensuring that any two behavioral units with physical possibilities have a non-zero transition probability basis.

[0077] Based on smoothed frequency data, the conditional transition probability from the predecessor segment type to the current segment type and the conditional transition probability from the current segment type to the successor segment type are calculated to establish a bidirectional Markov chain transition model.

[0078] Similarly, calculate the conditional probability of transitioning from the current state to the successor state, and construct a complete state transition probability matrix. The row index of this matrix represents the type of the predecessor or current state, and the column index represents the type of the current or successor state.

[0079] The predecessor transition probability matrix and the successor transition probability matrix are combined by tensor product operation or splicing to generate a three-dimensional dynamic context association matrix. Each element of the matrix represents the joint probability density of the current segment evolving into a specific successor segment under a specific predecessor context.

[0080] The dynamic context association matrix is ​​normalized so that for any given predecessor-current combination, the sum of the probabilities of all possible successor states is 1, thus ensuring the physical completeness of the probability distribution.

[0081] Through the above processing method, discrete static morphological feature vectors are transformed into dynamic contextual correlation matrices containing temporal evolution laws, realizing the quantitative characterization of the dependencies between load behaviors, and providing key temporal logic support for subsequent calculation of risk tendency weights by combining historical feedback data.

[0082] S3.3: Utilize historical false alarm and false alarm feedback data to perform weighted iterative calculations on the combination result of the static morphological feature vector and the dynamic context association matrix to quantify the risk propensity of different behavioral patterns under specific working conditions, and output candidate behavioral pattern entries with initial risk propensity weights.

[0083] The system receives the static morphological feature vector generated in S3.1 and the dynamic context association matrix generated in S3.2 as input data sources for calculating risk propensity weights. It retrieves false alarm and missed alarm records stored in the historical operation database to construct a historical feedback dataset containing timestamps, operating condition labels, sensor IDs, and actual load states. The historical feedback data is cleaned, removing invalid data points caused by sensor hardware failures or communication interruptions, retaining valid feedback samples with clear physical meaning. Valid feedback samples are categorized according to the triplet structure of load behavior patterns. The system counts the number of false alarms (triggered by each specific semantic combination but without actual overload) and the number of missed alarms (actual overload but without triggering an alarm) in the historical data. Based on the statistically obtained false alarm and missed alarm counts, the system calculates the basic risk penalty coefficient for each load behavior pattern, where a higher number of false alarms results in a larger penalty coefficient, and a higher number of missed alarms results in a larger reward coefficient. The system introduces operating environment factors, assigning different basic weight offsets to the same behavior pattern in different scenarios based on the specific operating environment of the stainless steel workbench (e.g., precision assembly area, heavy material stacking area). An iterative update model for risk propensity weights is constructed, using static morphological feature vectors, dynamic contextual correlation matrices, basic risk penalty coefficients, and environmental factors as input variables. A weighted linear regression method is employed to map these input variables into a multidimensional space, calculating initial risk propensity weight values. During the calculation process, a regularization term is introduced to prevent overfitting, ensuring that the weight values ​​fluctuate within a reasonable range and avoiding extreme values ​​that could cause alarm logic failure. The calculated initial risk propensity weights are assessed for confidence by generating multiple subsets using the Bootstrap resampling method, and the variance of the weight values ​​is calculated for each subset. If the variance of the weight distribution for a certain behavior pattern exceeds a preset threshold, the pattern is marked as a low-confidence pattern and indicated in the dictionary that further data accumulation is needed. For high-confidence patterns, their weight values ​​are directly embedded into the candidate behavior pattern entries. Based on real-time feedback data within the current time window, the embedded weight values ​​are fine-tuned online, using a moving average method to smooth short-term fluctuations and ensure the stability of weight updates. The system outputs candidate behavior pattern entries with initial risk propensity weights. Each entry contains a unique pattern ID, a static morphological feature description, a contextual association rule, an initial risk weight value, and a confidence level label. Through the aforementioned iterative weight calculation process, the static features and contextual relationships from the previous step are transformed into dynamic weight indicators that quantify the risk level. This enables a refined distinction of the risk degree of different load behavior patterns, providing data support for addressing the lack of priority differentiation in alarm mechanisms.

[0084] S3.4: Perform clustering and indexing operations on the candidate behavior pattern entries with initial risk propensity weights to eliminate redundant behavior patterns and establish a fast retrieval mechanism, generating a structured basic load behavior pattern dictionary.

[0085] The system receives a set of candidate behavior pattern entries with initial risk propensity weights. This set contains discrete load fragment data and their corresponding static attribute vectors after morphological feature extraction and contextual analysis. Multidimensional spatial distance measurement is performed on the static morphological feature vectors of the candidate behavior pattern entries, using Euclidean distance as the basic similarity evaluation index to quantify the geometric proximity of different behavior patterns in the feature space. A density-based clustering algorithm model is constructed, setting a minimum neighborhood radius parameter and a minimum number of points threshold for core objects to identify high-density connected regions in the feature space. Load fragments with similar physical meanings but slight numerical deviations are grouped into the same cluster. For behavior pattern entries within each cluster, a weighted average fusion process is performed. Using the historical frequency and initial risk weight of each entry as coefficients, the central feature vector and comprehensive risk propensity benchmark value of the cluster are calculated to eliminate redundant pattern splitting caused by sensor noise or subtle operational differences. Semantic label consistency checks are performed on the clustered and merged behavioral pattern clusters to ensure that all original fragments within the same cluster have the same predecessor-successor contextual logical constraints. If contextual conflicts are found, sub-clusters are split or the dominant pattern is retained based on risk weights to ensure the logical purity of dictionary entries. A unique hash index key is assigned to each final determined behavioral pattern cluster. This key is generated by combining the quantized encoding of the cluster's central feature vector with a context type identifier, establishing a fast mapping relationship from the feature space to the semantic space. An inverted index structure is constructed, decomposing the feature dimensions of the behavioral pattern clusters into multiple sub-index trees, supporting fast retrieval based on local feature matching and reducing the computational complexity of online real-time matching. The behavioral pattern clusters, after clustering deduplication, weight fusion, and index construction, are encapsulated into a standardized dictionary data structure, including a central feature template, risk baseline weights, context transition probability matrix, and unique retrieval ID, forming a structured basic payload behavioral pattern dictionary. Through the above-mentioned clustering and indexing process, the redundant and discrete candidate behavior patterns generated in the previous step are transformed into a structured, low-redundancy basic load behavior pattern dictionary with efficient retrieval capabilities, thereby significantly improving the efficiency of pattern matching and optimizing storage space in overload risk identification.

[0086] S3.5: Perform risk evolution path mapping tests based on the structured basic load behavior pattern dictionary to verify the accuracy of alarm triggering logic under simulated working conditions for different word combinations, and finally establish a complete load behavior pattern dictionary with online update capability.

[0087] Obtain a structured basic load behavior pattern dictionary, which contains candidate behavior pattern entries after clustering and merging and their initial risk propensity weights, as input for risk evolution path mapping tests.

[0088] A simulated load dataset covering typical working conditions of stainless steel workbenches was constructed. The dataset includes standardized load time sequences under various scenarios such as rapid material stacking, slow tool handling, personnel leaning against impact, and equipment start-up and shutdown vibration, ensuring that the test samples cover the mechanical behavior characteristics of the full range and full frequency band.

[0089] Offline semantic segmentation is performed on each time sequence in the simulated load dataset. The sequence is divided into basic mechanical behavior units using the dynamic time warping method and combined into triplet structures according to the time order to generate a set of standard semantic paths for verification.

[0090] Traverse each triplet structure in the standard semantic path set, perform key-value retrieval in the structured basic load behavior pattern dictionary, match the corresponding term records, and extract the pre-stored static morphological feature vector and dynamic context association matrix of each term.

[0091] Based on the extracted term records, the comprehensive risk score of the current triplet structure in the dictionary is calculated. The weighted summation method is used to fuse the weight of the predecessor segment, the weight of the current segment, and the probability distribution of the successor segment to form a preliminary risk evolution prediction value.

[0092] A historical false alarm and missed alarm feedback database is introduced, and historical case labels similar to the current triplet structure are retrieved. The preliminary risk evolution prediction value is compared with the actual safety event level, and the prediction deviation is calculated.

[0093] Define a confidence assessment function for the risk evolution path, and calculate the confidence coefficient of path matching using the following formula: Where C is the confidence coefficient, λ is the sensitivity adjustment factor, S is the comprehensive risk score, and θ is the empirical threshold benchmark with a value of 0.6.

[0094] The initial risk evolution prediction value is corrected based on the confidence coefficient C. If C is lower than the preset minimum confidence threshold of 0.6, the triplet structure is marked as a low confidence path, triggering a re-clustering or weight reset mechanism.

[0095] For high-confidence paths, map them to a multi-level alarm range, determine whether the alarm level triggered by them is consistent with the actual danger level of the simulated working conditions, and record the result of successful or unsuccessful matching.

[0096] The judgment results of all test samples are statistically analyzed, and the overall accuracy, false alarm rate, and false negative rate are calculated to identify high-frequency error path patterns that lead to judgment deviations.

[0097] For the identified high-frequency error path patterns, the risk tendency weights of the corresponding terms in the load behavior pattern dictionary are adjusted in reverse, and the gradient descent method is used to minimize the prediction error loss function.

[0098] Define a weight update rule and iteratively update the risk propensity weights using the following formula: in, Let be the weight value for the a-th iteration. The weight value for the (a+1)th iteration. For learning rate, For loss function, This is the partial derivative of the loss function with respect to the weights.

[0099] Repeat the above mapping test and weight iteration process until the overall accuracy converges to the preset target value and the false alarm rate stabilizes within the allowable range, then stop the iterative optimization.

[0100] The optimized term weights, the revised context association matrix, and the final risk evolution path mapping relationship are solidified and stored to generate a complete load behavior pattern dictionary with online update capabilities.

[0101] Through the aforementioned risk evolution path mapping test and closed-loop weight iteration processing method, the structured basic dictionary generated in the previous step is transformed into a high-precision risk assessment model that has been empirically verified and has adaptive correction capabilities, thereby significantly improving the accuracy of overload risk classification and identification and the system's environmental adaptability.

[0102] like Figure 3 As shown, step S4 involves parsing the semantic structure combination of the current load sequence using the load behavior pattern dictionary. Specifically, this includes: S4.1: Obtain the standardized load time series within the real-time sliding window, and calculate the morphological similarity matrix between the standardized load time series and the set of basic mechanical behavior units in the load behavior pattern dictionary based on the dynamic time warping method, to obtain the candidate segment matching index set.

[0103] Sub-step S4.1 is located in the online semantic parsing stage of the state analysis module. Its core function is to map the continuous flow of real-time load data into discrete semantic primitives. This step receives the standardized load time series output from the S1 preprocessing module as input, and simultaneously calls the set of basic mechanical behavior units in the load behavior pattern dictionary constructed in S3. By introducing a dynamic time warping method, the temporal nonlinear alignment problem caused by differences in operating speed is solved, and the morphological distance between real-time data fragments and standard behavior patterns is quantified, thereby filtering out high-confidence candidate matches and providing accurate local observation probability basis for subsequent Viterbi decoding.

[0104] S4.2: Perform Viterbi decoding path search based on the candidate fragment matching index set, and perform optimal path segmentation on the standardized payload time series to obtain an online semantic fragment sequence with clear boundary markers.

[0105] Receive the candidate fragment matching index set generated in step S4.1. This index set contains the morphological similarity score matrix between the standardized load time series sequence and each basic mechanical behavior unit in the load behavior pattern dictionary within the real-time sliding window, and serves as the initial observation probability input for the Viterbi decoding method.

[0106] A state transition topology based on a Hidden Markov Model (HMM) is constructed. Each basic mechanical behavior unit in the load behavior pattern dictionary is defined as a hidden state node. Based on the dynamic context association matrix established in step S3, the transition probability of any two behavior units appearing consecutively on the time axis is calculated to form a state transition probability matrix.

[0107] The state transition probability matrix is ​​row-normalized to ensure that the sum of the probabilities of transitioning from any current state to all possible successor states is 1, thereby eliminating probability distribution distortion caused by statistical bias in historical data and ensuring the consistency of probability in the decoding path search.

[0108] Initialize the dynamic programming table of the Viterbi decoder, set the starting time to the first data point of the sliding window, set the initial probability of each hidden state at the starting time to a uniform distribution or a specific distribution based on prior knowledge, and set the path backtracking pointer to null.

[0109] For each time step within the sliding window, traverse all possible hidden states and calculate the maximum local path probability to reach the current state using the maximum cumulative probability of the previous time step, the transition probability of the current state, and the emission probability of the current observation data.

[0110] Synchronously record the predecessor state index that generates the maximum local path probability to the path backtracking pointer matrix for subsequent optimal path tracing, ensuring that each decision retains traces of the globally optimal substructure.

[0111] Repeat the above probability update and pointer recording process until all time steps within the sliding window have been traversed, completing the forward probability propagation calculation for the entire observation sequence.

[0112] At the last time step of the sliding window, the hidden state with the maximum cumulative probability value is selected as the endpoint state of the optimal path, and the maximum probability value is recorded as the confidence score of the current sequence matching.

[0113] Initiate a reverse backtracking mechanism, starting from the endpoint state, and search for the optimal predecessor state corresponding to each time step according to the path backtracking pointer matrix, until backtracking to the starting time step.

[0114] The backtracked state sequence is arranged in reverse order to generate the optimal hidden state sequence from the start time to the end time. This sequence is the best interpretation path of the standardized load time series sequence in the semantic space.

[0115] Based on the moment when the state changes in the optimal hidden state sequence, the segmentation boundary of the load behavior unit is determined, and the continuous standardized load time sequence is divided into several sub-segments with independent semantic labels.

[0116] Each sub-segment is assigned a behavior type identifier and risk tendency weight corresponding to its hidden state, forming an online semantic segment sequence with clear boundary identifiers and semantic attributes.

[0117] By using the Viterbi decoding path search processing method, the candidate segment matching index set of the previous step is transformed into an online semantic segment sequence with global optimality, which realizes accurate semantic segmentation of complex load time series signals, effectively solves the semantic ambiguity problem caused by the lack of context constraints in traditional local matching methods, and significantly improves the structuring degree of overload risk identification.

[0118] S4.3: Perform a sliding window triplet construction operation on the online semantic segment sequence, extract three consecutive online semantic segments as the current semantic context unit, and obtain the semantic triplet structure to be parsed.

[0119] S4.4: Use the term mapping relationship in the load behavior pattern dictionary to perform key value retrieval on the semantic triple structure to be parsed, match the pre-stored semantic association rules, and obtain a semantic structure combination with risk tendency weight annotation.

[0120] Sub-step S4.4 is located in the core of the online semantic parsing of the state analysis module. Its function is to transform the local semantic context units constructed in the preceding steps into structured data with quantified risk levels. This step receives the semantic triple structure to be parsed from the output of S4.3. This structure contains the predecessor fragment label, the current fragment label, and the successor fragment label arranged in chronological order, serving as the query key for risk matching. The system calls the load behavior pattern dictionary constructed in step S3 and verified offline. This dictionary stores all legal triple combinations and their corresponding risk propensity weights and confidence parameters.

[0121] The parsed semantic triplet structure is subjected to standardized hash encoding, and the three consecutive basic mechanical behavior unit labels are combined according to the preset position weight rules to generate a unique composite index key.

[0122] Based on the generated composite index key, an exact match retrieval is performed in the B+ tree inverted index structure of the load behavior pattern dictionary. If the retrieval is successful, the pre-stored risk tendency weight, context transition probability, and historical confidence level corresponding to the triple are directly extracted. If the retrieval fails, it indicates that the current load evolution path belongs to a novel, unseen operating condition or a noise interference sequence. The system then initiates a generalized matching mechanism, calculating the edit distance between each segment in the current triple and the known templates in the dictionary. The approximate triple with the smallest edit distance is selected as the alternative match and assigned a lower initial confidence level to prevent unknown patterns from directly triggering high-risk alarms.

[0123] Step S5: Query the pre-stored triplet risk evolution map to map the load evolution trend implied by the semantic structure combination, and initially generate a dynamic alarm level determination result. Specifically, this includes: S5.1: Based on the semantic structure combination of the current load sequence generated in the previous steps, extract the triplet structure feature vector to construct the risk evolution query index to be matched.

[0124] S5.2: Use the risk evolution query index to be matched to perform similarity retrieval in the pre-stored triplet risk evolution map to locate the target risk evolution path node with the highest matching degree.

[0125] Receive the risk evolution query index to be matched constructed from step S5.1. This index consists of the triple structure feature vector parsed from the current payload sequence, including the predecessor fragment type encoding, the current fragment morphological parameters and the subsequent fragment prediction probability distribution, which serves as a benchmark reference for high-dimensional space retrieval in the pre-stored map.

[0126] Load the pre-stored triplet risk evolution graph data structure. The graph is stored in the form of a directed weighted graph. Nodes represent specific load behavior semantic states, and edges represent the transition paths between states and their associated historical risk statistics. Each node and edge is accompanied by multi-dimensional attribute fields, including average load growth rate, duration distribution moment, frequency of historical alarm trigger levels, and confidence score.

[0127] The triplet feature vectors in the query index are subjected to dimension alignment and standardization preprocessing. The key discriminative feature subsets are extracted, including semantic tag ID sequences, morphological similarity score vectors and context transition probability values. The dimensional differences caused by sensor drift or small fluctuations in sampling rate are eliminated, and a normalized query vector is generated.

[0128] Traverse all candidate path nodes in the triplet risk evolution graph, calculate the weighted Euclidean distance between the normalized query vector and the feature vector of the k-th node in the graph. This distance metric reflects the geometric proximity of the current real-time operating conditions and historically known risk patterns in the semantic space.

[0129] A minimum search operation is performed on the calculated distance set to identify the target node index with the minimum weighted Euclidean distance. This index points to the historical risk evolution path node in the graph that is most similar to the current real-time payload semantic combination.

[0130] Extract the metadata information bound to the target risk evolution path nodes, including the typical load evolution trend description, historical average risk probability value, recommended alarm level code and confidence lower limit corresponding to the node, and encapsulate this information into a target node data packet as the direct input basis for subsequent dynamic risk probability calculation.

[0131] By using a weighted Euclidean distance metric and an adaptive threshold filtering process, the abstract semantic query index from the previous step is transformed into specific target risk evolution path nodes with clear physical meaning in the graph. This achieves a precise mapping from the real-time semantic structure to the historical risk knowledge base, ensuring that the alarm judgment logic is based on sufficient historical statistical evidence, and significantly improving the interpretability and accuracy of overload risk identification.

[0132] S5.3: Calculate the dynamic risk probability value under the current operating condition based on the historical load evolution trend data associated with the target risk evolution path node, so as to quantify the possibility of load exceeding the limit.

[0133] Receive the target risk evolution path node located in step S5.2. This node contains a pre-stored set of historical load evolution trend data and associated risk statistical parameters, which serve as a benchmark for dynamic risk probability calculation.

[0134] Extract the historical load peak distribution sequence and duration distribution sequence stored in the target risk evolution path nodes, construct a multi-dimensional feature matching space under the current working condition, and use it to quantify the degree of deviation between the real-time semantic triple and the historical high-risk pattern.

[0135] Based on the historical load peak distribution sequence, the normalized offset of the current real-time load amplitude relative to the historical extreme value is calculated. A Gaussian kernel function is used to map the amplitude difference into an amplitude risk component, reflecting the possibility of instantaneous load exceeding the limit.

[0136] Based on the historical duration distribution sequence, the proportional coefficient of the current semantic segment duration relative to the historical typical overload duration is calculated. An exponential decay function is used to map the duration difference into a duration risk component, reflecting the cumulative effect of load state deterioration.

[0137] Obtain historical false alarm and false negative rate statistics recorded in the nodes of the target risk evolution path, construct a confidence correction factor to balance the weight allocation of amplitude risk and duration risk, and avoid assessment distortion caused by single-dimensional bias.

[0138] A weighted fusion strategy is adopted to linearly combine the amplitude risk component and the duration risk component, and multiply them by a confidence correction factor to generate a preliminary dynamic risk probability value.

[0139] The calculated dynamic risk probability value is truncated within the range of [0, 1] to ensure the validity of the physical meaning of the probability value and eliminate numerical overflow caused by extreme abnormal data.

[0140] The truncated dynamic risk probability value is encapsulated into a standardized risk quantification index, along with the current timestamp and sensor ID, and used as direct input data for multi-level risk threshold comparison in step S5.4.

[0141] By using the above-mentioned Gaussian kernel amplitude mapping, exponential decay duration accumulation and confidence weighted fusion processing, the target risk evolution path nodes matched in the previous step are transformed into quantified dynamic risk probability values, realizing an accurate numerical representation of the possibility of overload exceeding limits. This effectively solves the technical problem that traditional fixed thresholds cannot adapt to dynamic operating conditions, and significantly improves the sensitivity and accuracy of overload risk assessment.

[0142] S5.4: Compare and analyze the dynamic risk probability value with the preset multi-level risk threshold range to determine the initial dynamic alarm level judgment result.

[0143] The system receives the dynamic risk probability value calculated in step S5.3. This value quantifies the probability of exceeding the limit corresponding to the current load semantic triple combination in the historical evolution map. At the same time, it retrieves the preset multi-level risk threshold interval configuration table as the benchmark reference system for alarm level determination.

[0144] The dynamic risk probability value is validated by interval mapping logic, which divides the continuous probability value space into non-overlapping discrete risk level intervals, including the safety monitoring zone, the mild warning zone, the moderate alert zone, and the severe emergency zone. Each interval corresponds to a specific alarm response strategy and handling priority.

[0145] By employing a piecewise linear mapping function or a stepwise comparison method, the dynamic risk probability value is compared with the threshold boundaries at each level to determine its corresponding risk level index. The threshold boundaries are not fixed constants, but are dynamically fine-tuned according to the current load baseline state of the workbench to adapt to the sensitivity requirements under different working conditions.

[0146] A multi-level risk assessment logic tree is constructed. First, it is determined whether the dynamic risk probability value exceeds the highest priority critical threshold. If the condition is met, it is directly locked as a first-level emergency alarm. This level represents the risk of structural failure that occurs immediately and requires the triggering of the highest priority audible and visual alarm and equipment emergency stop linkage signal.

[0147] If the dynamic risk probability value does not reach the critical threshold, it continues to be compared with the moderate warning threshold. If the dynamic risk probability value is between the moderate warning threshold and the critical threshold, it is determined to be a level 2 moderate alarm. This level indicates that the load is showing a continuous deterioration trend or is approaching the design limit. It is necessary to trigger a continuous alarm sound and record high-frequency log data.

[0148] If the dynamic risk probability value only exceeds the mild warning threshold but does not reach the moderate alert threshold, it is judged as a level three mild warning level. This level indicates that the load has abnormal fluctuations or short-term peaks. It is necessary to trigger intermittent warning sounds and display a attention mark on the operation interface to remind operators to pay attention to standardized operations.

[0149] If the dynamic risk probability value is lower than all preset thresholds, it is determined to be in a normal monitoring state and no alarm action is triggered. However, the system will retain the current semantic structure combination data for subsequent unsupervised updates of dictionary weights to optimize the adaptability of the long-term model.

[0150] A hysteresis comparison mechanism is introduced, which sets a small buffer zone at the boundary of adjacent thresholds to prevent the risk probability value from frequently jumping between two alarm levels due to noise disturbances, thus ensuring the stability and determinism of the alarm level determination result.

[0151] Generate a structured judgment result object containing alarm level code, trigger timestamp, associated semantic triple ID, and dynamic risk probability value. This object serves as the initial dynamic alarm level judgment result and will be passed to step S6 for spatial topology confidence correction.

[0152] By using the above-mentioned multi-level threshold comparison and hysteresis stabilization processing, the dynamic risk probability value quantified in the previous step is transformed into discrete alarm level instructions with clear action guidance, realizing a logical closed loop from probability assessment to decision execution, and significantly improving the pertinence of overload risk response and the level of refinement of hierarchical management.

[0153] Step S6: Based on the spatial topology information of the workbench, analyze the semantic consistency and geometric clustering characteristics of multiple distributed pressure sensors, and correct the confidence level of the dynamic alarm level determination result to obtain a spatial fusion risk rating. Specifically, this includes: S6.1: Obtain the set of local semantic fragment labels output by all distributed pressure sensors within the current time window, and based on the predefined workbench physical coordinate system mapping relationship, convert the set of local semantic fragment labels into sensor node state vectors with spatial location attributes to construct a real-time spatial topology state matrix.

[0154] S6.2: Based on the real-time spatial topology state matrix, execute the neighborhood consistency verification method to calculate the semantic matching degree between each sensor node and its physical neighboring nodes, and obtain a sequence of node semantic consistency coefficients that characterize the synergy of local load behavior.

[0155] S6.3: Perform density clustering analysis using the node semantic consistency coefficient sequence to identify sensor node groups with high-risk semantic labels and continuous spatial distribution, so as to extract spatial geometric clustering feature clusters that characterize concentrated overload areas.

[0156] Receive the node semantic consistency coefficient sequence generated in the previous steps. This sequence contains data on the degree of matching between each distributed pressure sensor node on the workbench and its physical neighboring nodes in terms of load behavior semantics.

[0157] Based on the node semantic consistency coefficient sequence, a weighted undirected graph model containing spatial coordinate information and semantic label attributes is constructed, where sensor nodes are the vertices of the graph, the reciprocal of the physical distance between nodes is the weight of the edge, and the node semantic consistency coefficient is the attribute value of the vertex.

[0158] Density-based spatial clustering (DBSCAN) algorithm is used to perform cluster analysis on high-risk nodes in a weighted undirected graph model. The minimum number of neighborhood points is set to 4, and the neighborhood radius is dynamically adjusted to 1.5 times the grid spacing according to the physical spacing of the sensor array, in order to identify high-risk areas that are spatially closely connected and semantically consistent.

[0159] Traverse all sensor nodes marked as core points and merge their density-reachable neighboring nodes into the same candidate cluster to form a preliminary spatial geometric cluster set. At this point, each cluster represents a potential concentrated overload region.

[0160] Sparse clusters with spatial compactness below a preset threshold of 0.6 are removed. These clusters are usually caused by random noise or local small perturbations and do not have macroscopic overload characteristics, thus retaining effective geometric clusters with high spatial continuity.

[0161] Morphological closure operations are performed on the retained effective geometric cluster feature clusters to fill the voids that may exist in the clusters due to the failure of individual sensors, ensuring the integrity of the extracted overload region in terms of geometric topology.

[0162] The key geometric parameters of the final spatial geometric cluster feature clusters are extracted, including the cluster center coordinates, coverage area ratio, major axis direction, and boundary contour point set, as spatial fingerprint information representing the concentrated overloaded region.

[0163] Through the density clustering and morphological optimization methods described above, the discrete node semantic consistency data obtained in the previous step is transformed into spatial geometric clustering feature clusters with clear physical boundaries and spatial topology. This enables accurate localization and structured description of concentrated overloaded regions, significantly improving the reliability of the spatial dimension basis for subsequent confidence assessment.

[0164] S6.4: Based on the coverage area ratio of the spatial geometric cluster feature clusters and the average value of the node semantic consistency coefficient sequence, construct a multidimensional confidence assessment model, perform weighted correction calculation on the dynamic alarm level judgment results generated in the previous steps, and output a spatially validated confidence correction factor.

[0165] The average value of the coverage area ratio of the spatial geometric cluster feature clusters output from the previous steps and the node semantic consistency coefficient sequence is used as the input variables of the multidimensional confidence evaluation model.

[0166] The coverage area ratio of the spatial geometric cluster feature clusters is normalized and mapped to the range of 0 to 1 to eliminate the influence of the difference in physical surface area of ​​different sized workbenches on the dimension of confidence calculation, and to generate a standardized spatial distribution weight factor.

[0167] The average value of the node semantic consistency coefficient sequence is extracted. This value represents the degree of coordination of each sensor node in the current high-risk area in terms of load behavior semantics, directly reflecting the real physical continuity of the overload event, and generating a semantic coordination confidence benchmark value.

[0168] A nonlinear coupled computational model based on spatial distribution weighting factors and semantic collaborative confidence benchmark values ​​is constructed, and the confidence correction factor for spatial validation is calculated using the following formula: in, This is the confidence correction factor for spatial validation. This is an empirical adjustment coefficient used to balance the contribution weights of spatial coverage and semantic consistency. For standardized spatial distribution weighting factors, This is the baseline value for semantic collaboration confidence.

[0169] Based on the calculated confidence correction factor for spatial verification, the initial risk probability value in the dynamic alarm level determination result generated in the previous step is weighted and corrected. The initial risk probability value is multiplied by the confidence correction factor to generate the corrected risk probability value after spatial dimension verification.

[0170] By using the above processing method, isolated spatial geometric features and local semantic consistency indicators are transformed into globally reliable risk quantification indicators, achieving the expected technical effect of effectively suppressing single-point false alarms and enhancing the confidence level of regional real overload risks.

[0171] S6.5: Based on the confidence correction factor and the dynamic alarm level determination result, perform level mapping fusion processing, adjust the urgency weight of the original determination result, and obtain the final spatial fusion risk rating instruction.

[0172] The confidence correction factor output from the preceding step S6.4 and the initial dynamic alarm level determination result generated in S5.4 are received as the input data source for the level mapping fusion process.

[0173] The initial dynamic alarm level determination results are discretized and encoded, and the first-level emergency alarm, the second-level early warning, and the third-level prompt are mapped to numerical weight vectors [3, 2, 1] respectively to construct the initial risk level scalar value.

[0174] Obtain a confidence correction factor, which characterizes the credibility of spatial topological consistency in the semantic judgment of the current payload. Its value is normalized to between 0 and 1 and is used to quantify the support of spatial evidence for temporal semantic judgment.

[0175] A level-weighted correction model based on a nonlinear gain function is established, and the initial risk level scalar value is dynamically adjusted using a confidence correction factor to eliminate the risk of false alarms under low confidence and enhance the urgency of alarms under high confidence.

[0176] The confidence baseline threshold T is set to 0.6. When the confidence correction factor is lower than this threshold, the correction term in the formula becomes negative, thereby reducing the final risk rating index and reflecting the suppression effect on isolated or contradictory signals.

[0177] The calculated spatial fusion risk rating index R is compared with the preset graded decision boundaries, which include a first-level alarm threshold of 2.5, a second-level alarm threshold of 1.5, and a third-level alarm threshold of 0.5.

[0178] If the spatial fusion risk rating index R is greater than or equal to 2.5, it is determined to be a Level 1 emergency overload state, and a spatial fusion risk rating instruction containing the highest priority identifier is generated.

[0179] If the spatial fusion risk rating index R is less than 2.5 and greater than or equal to 1.5, it is determined to be a level 2 potential overload state, and a spatial fusion risk rating instruction containing a medium priority identifier is generated.

[0180] If the spatial fusion risk rating index R is less than 1.5 and greater than or equal to 0.5, it is judged as a level three minor abnormal state, and a spatial fusion risk rating instruction containing a low priority identifier is generated.

[0181] If the spatial fusion risk rating index R is less than 0.5, it is determined to be a normal load state, a silent command without alarm trigger is generated, and the current state is marked as safe.

[0182] By using the aforementioned nonlinear gain weighting and multi-level threshold decision processing method, the initial dynamic alarm level determination result is transformed into a final spatial fusion risk rating instruction with spatial confidence verification. This enables the alarm level to adapt to the consistency of load spatial distribution, significantly reducing the false alarm rate caused by local transient interference, while ensuring the timeliness of alarms in concentrated overload scenarios.

[0183] Step S7: Determine whether the spatial fusion risk rating corresponds to an isolated high-risk segment. If the condition is met, execute the delayed review logic; otherwise, directly output the final overload risk classification instruction. Specifically, this includes: S7.1: Perform logical judgment processing on the spatial fusion risk rating and the corresponding distributed pressure sensor semantic consistency analysis results to identify isolated high-risk segments that are triggered by only a single sensor node and lack semantic support from neighboring nodes.

[0184] S7.2: Start a timer counter based on the start timestamp of the isolated high-risk segment, and perform sliding window truncation processing on the original load time sequence signal within the subsequent preset duration to obtain the load data subsequence to be reviewed.

[0185] S7.3: Utilize the load behavior pattern dictionary to perform online semantic matching operations on the load data subsequence to be reviewed, in order to extract continuous basic mechanical behavior units and their evolution trend features within the review window.

[0186] Obtain the subsequence of payload data to be reviewed extracted in step S7.2. This subsequence contains the original sensor sampling point data within a preset time after the start time of the isolated high-risk segment.

[0187] The data subsequences of the verification load are subjected to a standardized preprocessing procedure consistent with step S1, including sliding window mid-range filtering for noise reduction, adaptive wavelet threshold smoothing, and dynamic baseline correction, to eliminate transient electromagnetic interference or mechanical vibration noise that may exist within the verification window and generate standardized verification load time series segments.

[0188] The dynamic time warping (DTW) method is used to calculate the morphological distance matrix between the standardized verification load time sequence segment and all basic mechanical behavior unit templates in the load behavior pattern dictionary, thereby quantifying the similarity between the signal to be verified and each known behavior pattern.

[0189] Based on the morphological distance matrix, the Viterbi decoding method is used to search for the globally optimal matching path, and the continuous complex load time sequence is divided into several discrete basic mechanical behavior unit sequences with clear physical meanings. Each unit corresponds to a specific term index in the dictionary.

[0190] The transition probabilities between adjacent units in the segmented basic mechanical behavior unit sequence are extracted. Combined with the context association rules pre-stored in the dictionary, a local semantic evolution map within the verification window is constructed to identify the continuity features of load changes.

[0191] The duration percentage and peak hold rate of each basic mechanical behavior unit within the verification window are calculated. If a combination of units that is 'sharp rise' followed immediately by 'rapid fall' is detected, and this combination is marked as a low-risk transient impact mode in the dictionary, it is determined to be an interference signal.

[0192] If a combination of units is detected that is 'gradually rising' followed by 'platform maintenance' or 'continuous rise', and the risk tendency weight corresponding to this combination is higher than the preset safety threshold, then it is determined to be a genuine continuation of the overload trend.

[0193] Through the above online semantic matching and evolution trend extraction processing, the load data subsequence to be reviewed in the previous step is transformed into a sequence of behavioral units with semantic labels and their risk evolution characteristic indicators, so as to realize the authenticity of isolated high-risk segments, effectively eliminate false alarms caused by instantaneous collisions or operational jitter, and ensure the accuracy and reliability of the final overload risk classification command.

[0194] S7.4: Based on the continuous basic mechanical behavior unit and its evolution trend characteristics, determine whether the load state continues to remain at a high level or shows an upward trend, and obtain the transient interference elimination judgment result or overload confirmation flag.

[0195] Obtain the sequence of continuous basic mechanical behavior units and their corresponding evolution trend feature vectors within the verification window output by step S7.3, and use them as the input data source for determining the persistence of the load state.

[0196] The temporal slope integral operation is performed on the continuous basic mechanical behavior unit sequence to calculate the cumulative increment of load change within the verification time window, so as to quantify the overall growth or maintenance of the load in a short period of time.

[0197] A linear regression fitting method was used to fit a trend line to the standardized load time series data within the verification window. The slope parameter and residual standard deviation of the fitted line were extracted to characterize the directionality and fluctuation stability of load changes.

[0198] A load status persistence assessment function is constructed, and the fitting slope parameter, cumulative increment and residual standard deviation are used as input variables. The persistence risk index of the current load status is calculated by weighted combination.

[0199] A dynamic judgment threshold range is set, and the calculated continuous risk index is compared with the preset transient interference exclusion threshold and overload confirmation threshold to distinguish between instantaneous impact and real overload behavior.

[0200] If the sustained risk index is lower than the transient interference exclusion threshold, the isolated high-risk segment is determined to be a false trigger caused by transient mechanical vibration or electromagnetic interference, and a transient interference exclusion judgment result is generated.

[0201] If the sustained risk index is higher than the overload confirmation threshold, and the fitted slope parameter shows a non-negative value or a slightly positive value, then the load status is determined to be at a sustained high level or a slow upward trend, and an overload confirmation flag is generated.

[0202] If the persistent risk index falls between the two, a second verification is performed by combining the semantic type of the preceding fragment. If the preceding fragment is of the 'rapid decline' type, it tends to be judged as transient interference; otherwise, it is marked as a state to be observed and the verification window is extended.

[0203] Through the above chain-like derivation process, the semantic matching result of the previous step is transformed into a clear qualitative conclusion of the load state (transient interference or real overload), which realizes the accurate identification of isolated high-risk signals, effectively eliminates false alarms caused by single-point transient noise, and significantly improves the anti-interference capability and alarm reliability of the system.

[0204] S7.5: Based on the transient interference elimination judgment result or the overload confirmation flag, the initial dynamic alarm level judgment result is corrected to output the final overload risk classification instruction after eliminating false alarm interference.

[0205] The transient interference elimination judgment result or overload confirmation flag generated in step S7.4, and the initial dynamic alarm level judgment result output in step S5 are received as the input data objects for this step.

[0206] The logic branch judgment process is executed for the input judgment result. When the transient interference elimination judgment result is true, the urgency code in the initial dynamic alarm level judgment result is forcibly reset to the safety status code, and the corresponding area positioning mark is cleared to block the upward transmission of false alarm signals.

[0207] When the overload confirmation flag is received as true, the original risk level code in the initial dynamic alarm level determination result is retained, and the confidence correction factor in the spatial fusion risk rating generated in step S6 is extracted for subsequent signal strength enhancement processing.

[0208] An alarm level correction mapping function is constructed, which performs a weighted coupling operation on the retained risk level code and the confidence correction factor to calculate the corrected alarm priority index. This index represents the actual overload threat level after spatiotemporal dual verification.

[0209] The calculated alarm priority index is discretized and mapped to a range. Based on the preset grading threshold boundaries, the continuous values ​​are converted into standardized final overload risk grading instruction codes to ensure that the instruction format conforms to the communication protocol specifications of the alarm execution module.

[0210] The final overload risk classification instruction code, trigger timestamp, and corresponding sensor node ID are encapsulated into a standardized data structure. After adding a check bit, the data is sent to the alarm execution module via the internal bus to complete the closed-loop output of risk determination.

[0211] By using the above-mentioned logical branch judgment and weighted coupling operation processing method, the transient verification result of the previous step is transformed into a final overload risk classification instruction with high confidence, thereby achieving the expected technical effect of eliminating transient interference false alarms and accurately matching the severity of overload.

[0212] Step S8: Based on the final overload risk classification instruction, the alarm execution module is driven to activate the corresponding level of audible and visual alarm or area location prompt, and the current load behavior pattern is recorded to update the risk tendency weight in the load behavior pattern dictionary. Specifically, this includes: S8.1: Obtain the final overload risk classification instruction and the corresponding spatial fusion risk rating result. Based on the preset alarm mapping strategy, parse and process the final overload risk classification instruction to obtain a structured control signal containing alarm level code and area coordinate information.

[0213] Receive the final overload risk classification instruction and spatial fusion risk rating result output from the preceding step S7. The instruction includes the risk level identifier after confidence correction, the set of sensor node IDs that triggered the alarm, and the corresponding physical coordinate mapping information.

[0214] The protocol parsing operation is performed on the final overload risk classification instruction to extract the risk level code field and map it to a predefined alarm priority index, where the first-level emergency alarm corresponds to the highest priority index, the second-level warning corresponds to the medium priority index, and the third-level prompt corresponds to the lowest priority index, thus establishing the time urgency benchmark for alarm response.

[0215] Based on the set of sensor node IDs, the spatial topology configuration table of the distributed pressure sensor array on the workbench is queried to obtain the two-dimensional plane coordinate values ​​and functional area identifiers of each trigger node, and the discrete node IDs are converted into spatial location data clusters with geometric meaning.

[0216] A centroid calculation method is performed on the spatial location data cluster to determine the coordinates of the geometric center point of the overload distribution on the platform. This coordinate serves as the physical anchor point for the area positioning prompt, ensuring that the alarm indication can accurately point to the area where the load is concentrated rather than a single sensor point.

[0217] Combining the alarm priority index with the geometric center point coordinates, a structured control signal data packet is constructed. This data packet contains pulse width modulation parameters for driving the audible and visual alarm, address codes for driving the area positioning LED indicator, and timestamp tags for logging by the host computer.

[0218] Through the above-mentioned parsing and mapping processing, the abstract risk classification instructions are transformed into structured control signals containing specific execution parameters and spatial coordinates, enabling the alarm execution module to accurately locate and classify overload events, thus eliminating the technical defects of ambiguity and homogeneity in traditional fixed threshold alarms.

[0219] S8.2: Receive the structured control signal, use the multimodal alarm driving method to perform level conversion and pulse modulation processing on the alarm level code, obtain the graded excitation waveform that drives the sound and light alarm to work, and trigger the area positioning indicator to flash.

[0220] S8.3: Extract the set of basic mechanical behavior units and its triplet structure matching record when the alarm is triggered this time, and encapsulate the set of basic mechanical behavior units and the triplet structure matching record into a sample of the load behavior pattern for this time to obtain the historical behavior log data to be updated.

[0221] S8.4: Retrieve the term records in the load behavior pattern dictionary that match the current load behavior pattern sample, and perform deviation correction calculation on the risk tendency weight of the term records based on the actual confirmation result of this alarm to obtain the updated risk tendency weight value.

[0222] Receive the historical behavior log data to be updated generated in step S8.3. This data includes the set of basic mechanical behavior units when the alarm was triggered, the triplet structure matching record, and the confirmation status flag in the final overload risk classification instruction output in step S7.

[0223] The historical behavior log data to be updated is parsed, and the key index fields corresponding to the unique identifiers (IDs) of the terms in the load behavior pattern dictionary are extracted, including the type encoding of the current semantic triple, the spatial topological clustering feature value, and the initial risk tendency weight value.

[0224] Read existing term records in the load behavior pattern dictionary that completely match the key index field, and obtain the historical risk tendency weight value, historical cumulative trigger count, and historical false alarm / missed alarm statistics for the term in the current version.

[0225] Based on the confirmation status flag in the final overload risk classification instruction output in step S7, determine the true nature of this load behavior. If it is confirmed as a real overload event, mark it as a positive sample feedback; if it is confirmed as a transient disturbance or false alarm after delayed review, mark it as a negative sample feedback.

[0226] A risk propensity weight bias correction model based on online gradient descent is constructed. A loss function is defined to quantify the difference between the current weight prediction value and the actual feedback label, aiming to minimize the long-term accumulated classification error.

[0227] Boundary constraint checks are performed on the updated risk propensity weight values ​​to ensure that their values ​​fall within the preset range of [0,1]. If they exceed the upper limit, they are truncated to 1; if they fall below the lower limit, they are truncated to 0, in order to ensure the rationality of the physical meaning.

[0228] The risk tendency weight values, after boundary constraint correction, are temporarily stored in a memory buffer and associated with the corresponding term record ID, forming a weight update queue to be written.

[0229] By using the above-mentioned deviation correction calculation method based on online gradient descent, the historical behavior log data from the previous step is transformed into risk tendency weight values ​​with self-correction capabilities, thereby achieving the expected technical effect of adaptive optimization of the load behavior pattern dictionary to dynamic operating condition changes.

[0230] S8.5: Replace the old weight parameters in the load behavior pattern dictionary with the updated risk propensity weight values, and perform a persistent storage write operation on the load behavior pattern dictionary to complete the iterative update of the risk assessment model with self-learning capabilities.

[0231] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0232] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0233] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for load-bearing monitoring and overload alarm of a stainless steel workbench, characterized in that, Specifically, it includes: S1: Acquire the raw load time sequence signal collected by the distributed pressure sensor array of the stainless steel workbench, and process the raw load time sequence signal to obtain a standardized load time sequence. S2: Divide the standardized load time sequence into data segments to obtain a set of basic mechanical behavior units; S3: Extract morphological features and contextual relationships from the set of basic mechanical behavior units, and calculate risk tendency weights by combining historical false alarm and false alarm feedback data, so as to construct a load behavior pattern dictionary; S4: Use the load behavior pattern dictionary to parse the semantic structure combination of the current load sequence; S5: Query the pre-stored triplet risk evolution map to map the load evolution trend implied by the semantic structure combination, and initially generate dynamic alarm level determination results; S6: Based on the analysis of the spatial topology information of the workbench, the semantic consistency and geometric clustering characteristics of multiple distributed pressure sensors are analyzed, and the confidence level of the dynamic alarm level determination result is corrected to obtain the spatial fusion risk rating. S7: Determine whether the spatial fusion risk rating corresponds to an isolated high-risk segment. If the condition is met, execute the delayed review logic; otherwise, directly output the final overload risk classification instruction. S8: Based on the final overload risk classification instruction, drive the alarm execution module to activate the corresponding level of audible and visual alarm or area positioning prompt, and record the current load behavior pattern to update the risk tendency weight in the load behavior pattern dictionary.

2. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 1, characterized in that, The process of obtaining a standardized load time series by processing the original load time series signal involves denoising and normalizing the original load time series signal to obtain the standardized load time series.

3. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 1, characterized in that, The process of segmenting the standardized load time series to obtain a set of basic mechanical behavior units includes: Based on the standardized load time sequence, a change point detection method is executed to identify abrupt change points and steady-state intervals, and continuous data is divided into physically meaningful independent segments to obtain a set of basic mechanical behavior units.

4. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 1, characterized in that, The process of parsing the semantic structure combination of the current load sequence using a load behavior pattern dictionary includes: The load behavior pattern dictionary is used to perform online semantic segmentation and triple structure matching on the latest load data in the real-time sliding window to parse out the semantic structure combination of the current load sequence.

5. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 2, characterized in that, Step S1 specifically includes: The original load time-series signal containing high-frequency electromagnetic interference and mechanical vibration noise is acquired from the distributed pressure sensor array of the stainless steel workbench. The original load time-series signal is then processed by pulse noise removal using the sliding window midpoint filtering method to obtain the pulse noise-free load time-series signal. Based on the aforementioned de-impulse noise load time-series signal, an adaptive wavelet threshold denoising method is used to perform multi-scale decomposition and coefficient shrinkage on the residual high-frequency random noise to obtain a smoothed and denoised load time-series signal. For the smoothed and denoised load timing signal, a dynamic baseline drift correction method is performed to identify and subtract the zero drift component caused by temperature changes or sensor creep, so as to obtain a zero baseline aligned load timing signal. Using the maximum absolute value of the zero-baseline aligned load time series signal as a reference, the maximum and minimum value normalization mapping process is performed on the data of each channel to generate a dimensionless normalized load time series signal. Based on the dimensionless normalized load timing signal, resampling and format encapsulation are performed according to preset timestamp alignment rules and a unified sampling frequency to obtain the final standardized load timing sequence.

6. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 3, characterized in that, Step S2 specifically includes: The standardized load time series is segmented by a sliding window to obtain a set of short-time load data frames containing local time-domain features; The mean drift and variance volatility of each frame are calculated based on the short-time load data frame set to obtain a dynamic statistical feature vector sequence characterizing the stability of the local signal. An adaptive threshold comparison logic is executed using a sequence of dynamic statistical feature vectors to identify the set of potential change point coordinates where the load signal undergoes a significant state transition. Based on the set of potential change point location coordinates, the standardized load time series is segmented and truncated to obtain the original load segment sequence composed of multiple independent steady-state intervals and transition intervals. The original load segment sequence is subjected to physical validity verification and noise segment removal to obtain the final set of basic mechanical behavior units with clear mechanical behavior semantics.

7. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 5, characterized in that, The sliding window median filtering method specifically involves extracting the local median point by point using an odd-length sliding window for the original load time-series signal of each channel, and replacing the amplitude outlier point with the median of the current window for each sampling point to achieve robust removal of amplitude abrupt peaks.

8. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 5, characterized in that, The adaptive wavelet thresholding denoising method specifically employs discrete wavelet decomposition, selects the Dobesie wavelet family as the basis function, adaptively updates the residual detail coefficient thresholds online according to the Stein unbiased risk estimation rule, and applies soft thresholding to each level of detail coefficients. Finally, the denoised signal is synthesized through inverse transform.

9. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 6, characterized in that, The physical validity verification and noise segment removal process includes performing generalized likelihood ratio detection on adjacent time-domain windows based on the mean, variance and waveform coefficients within the sliding window, taking time points exceeding the threshold as change points, segmenting steady-state and transition behavior segments, and marking the start and end positions of the segments, the main rate of change and the interval category attributes.

10. The method for load-bearing monitoring and overload alarm of a stainless steel workbench according to claim 1, characterized in that, The confidence correction of the dynamic alarm level determination result includes: By leveraging sensor spatial distribution, node semantic consistency, and spatial clustering characteristics, combined with spatial coverage area ratio and collaborative indicators of similar nodes, a nonlinear coupling correction function is used to weight the risk probability of all dynamic alarm level determination results, significantly suppressing isolated false alarms and improving the reliable response to concentrated overload risks.

Images