A telescopic forklift truck load stability prediction system based on compression bar stress

Abstract

This invention relates to the field of engineering machinery safety technology, specifically disclosing a load stability prediction system for telescopic boom forklifts based on the force exerted by the boom. The system collects motion sensing data and stress wave sensing data at the boom hinge; calculates a cross-spectral coherence function based on the motion data to generate a hinge coherence disorder index; performs time-frequency transformation on the stress wave data and automatically identifies energy impact patches, extracting their peak energy, center frequency, and duration; and generates a transient collision intensity spectrum index based on sensitive frequency band mapping and aggregation; constructs a time-series comprehensive feature vector from the two indices, inputs it into a pre-trained multi-layer gated recurrent unit structure for nonlinear dynamic system identification, and outputs a fused stability assessment state quantity; maps this state quantity to a virtual potential energy surface for multi-step forward extrapolation, calculates the dynamic stability margin value, and issues an early warning when the margin is lower than an adaptive threshold; this invention achieves early and accurate prediction of nonlinear jump instability.

Description

Technical Field

[0001] This invention relates to the field of engineering machinery safety technology, specifically to a load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar. Background Technology

[0002] Telescopic boom forklifts, as a key type of engineering machinery combining lifting, handling, and high-altitude operations, are widely used in construction, logistics, and disaster relief. Their core working device, the multi-section telescopic boom system, withstands the combined effects of axial pressure, bending moment, and dynamic loads under complex working conditions. Its load stability directly affects operational safety and efficiency. Traditional stability assessments primarily rely on static calculations based on classical column buckling theory or quasi-static analysis based on overturning moments. These methods monitor macroscopic parameters (such as load weight, boom length, and angle) and compare them with preset safety thresholds to provide early warnings of risks.

[0003] Existing technologies have the following shortcomings: they are completely unable to predict "jumping instability" caused by the nonlinear dynamics of microscopic gaps at the articulation points of multi-section booms. In actual operation, gaps between boom sections caused by manufacturing tolerances and wear can trigger transient collisions and complex frictions under dynamic excitations such as bumps and micro-movements. These nonlinear time-varying boundary conditions may cause the system to suddenly lose stability due to parametric resonance before macroscopic parameters (such as load and angle) reach static critical values, resulting in catastrophic instability without warning. Traditional static models based on the assumption of a smooth continuum fail to detect or characterize such microscopic nonlinear dynamic behavior, thus completely failing to predict this type of sudden instability, posing a serious and hidden safety hazard. Summary of the Invention

[0004] The purpose of this invention is to provide a load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A load stability prediction system for telescopic boom forklift trucks based on lever stress includes:

[0007] The multi-source sensor data synchronous acquisition module is used to collect motion sensor data and stress wave sensor data in real time during telescopic boom forklift loading operations;

[0008] The micro-articulation nonlinear characteristic calculation module calculates the cross-spectral coherence function of adjacent arm segment motion sensing data, extracts the main frequency coherence value, distribution entropy and phase stability standard deviation of the cross-spectral coherence function, and generates the articulation coherence disorder index by fusing them according to preset rules.

[0009] The stress wave feature analysis module performs time-frequency transformation on stress wave sensing data and automatically identifies energy impact patches. It extracts the peak energy, center frequency, and duration of the energy impact patches, aggregates them according to the sensitive frequency band mapping rules, and integrates the event frequencies to generate a transient collision intensity spectrum index.

[0010] The fusion state quantity generation module constructs a comprehensive feature vector from the articulated coherent disorder index and the transient collision intensity spectrum index. The comprehensive feature vector is then input into a nonlinear dynamic system identification model built based on a machine learning framework, and the fusion stability assessment state quantity is output.

[0011] The dynamic early warning decision module, based on the theoretical framework of the stress stability of the compression bar, performs forward extrapolation and prediction of the fusion stability assessment state variables to calculate the dynamic stability margin value. When the dynamic stability margin value is lower than the preset adaptive threshold, an early warning signal is generated and issued for the risk of nonlinear jump instability.

[0012] As a further aspect of the present invention: the calculation of the cross-spectral coherence function of adjacent arm segment motion sensing data specifically includes:

[0013] Acquire motion sensing data of adjacent arm segments within a specified time window;

[0014] The motion sensing data is processed by frame-by-frame windowing to obtain the time-domain signal after multi-frame windowing.

[0015] Perform Fourier transform on the windowed time-domain signal of each frame to obtain the corresponding frequency-domain complex sequence;

[0016] Calculate the cross power spectral density and the individual self power spectral density of each frame signal at the corresponding frequency point;

[0017] Based on the cross-power spectral density and the self-power spectral density, the complex form of the cross-spectral coherence function is calculated point by point.

[0018] As a further aspect of the present invention: the generation of the articulated coherence disorder index specifically includes:

[0019] The average modulus of the preset main frequency band is extracted from the cross-spectral coherence function in complex form as the main frequency coherence value;

[0020] The entropy of the distribution of the modulus of the cross-spectral coherence function across all analysis frequency bands is calculated as the distribution entropy;

[0021] The standard deviation of the phase angle of the cross-spectral coherence function within the preset main frequency band is used as the standard deviation of phase stability.

[0022] The coherence value of the main frequency, the distribution entropy, and the standard deviation of the phase stability are normalized.

[0023] The three normalized eigenvalues ​​are linearly combined according to preset weights, the weighted sum is calculated, and the weighted sum is output as the articulated coherence disorder index.

[0024] As a further aspect of the present invention: the step of performing time-frequency transformation on stress wave sensing data and automatically identifying energy impact patches specifically includes:

[0025] Continuous wavelet transform is performed on the acquired stress wave sensing data to generate a time-frequency distribution spectrum;

[0026] Sliding window energy statistics are performed on the time-frequency distribution map to identify local regions where the energy exceeds the dynamic background threshold;

[0027] Morphological expansion and erosion operations are performed on the identified adjacent high-energy regions to merge them into connected regions.

[0028] Each connected component is defined as an energy impact patch, and the corresponding time-frequency boundary data is output.

[0029] As a further aspect of the present invention: the generation of the transient collision intensity spectrum index specifically includes:

[0030] Based on the time-frequency boundary data of the energy impact patches, the peak energy, center frequency, and duration of each patch are extracted.

[0031] According to the preset sensitive frequency band mapping rules, the center frequency of each patch is assigned a corresponding frequency band importance weighting coefficient;

[0032] The weighted cumulative energy is obtained by multiplying the peak energy of all patches within the same analysis window by the frequency band importance weighting coefficient and summing the results.

[0033] The weighted cumulative energy is multiplied by the total number of patches within the time analysis window, and the product is output as the transient collision intensity spectrum index.

[0034] As a further aspect of the present invention: the output fusion stability evaluation state quantity specifically includes:

[0035] The articulated coherent disorder index and the transient collision intensity spectrum index are aligned and spliced ​​according to the timestamp to construct a comprehensive feature vector containing microscopic nonlinear dynamic characteristics.

[0036] The comprehensive feature vector is input into a pre-trained nonlinear dynamic system identification model based on a machine learning framework. The nonlinear dynamic system identification model includes multiple gated recurrent unit layers connected in series for multi-layer temporal abstraction of the comprehensive feature vector; and a feedforward projection layer for mapping the hidden state vector output by the last gated recurrent unit to a fused stability evaluation state quantity as the output of the nonlinear dynamic system identification model.

[0037] As a further aspect of the present invention: the training process of the nonlinear dynamic system identification model is as follows:

[0038] Construct a training dataset containing time series generated by dynamic simulations that cover normal operation and instability boundaries. Each time series point contains a corresponding integrated feature vector and a baseline value of the fused stability evaluation state quantity as the target label.

[0039] The structure composed of multiple gated recurrent unit layers and feedforward projection layers is trained in a supervised manner, using the comprehensive feature vector and the fusion stability evaluation state quantity of the previous time step as input, and the benchmark value of the fusion stability evaluation state quantity of the current time step as the target.

[0040] During training, the composite loss function is optimized. The composite loss function simultaneously constrains the approximation accuracy of the fusion stability evaluation state quantity output by the feedforward projection layer to the baseline value of the fusion stability evaluation state quantity and the temporal continuity of the instability trend prediction.

[0041] As a further aspect of the present invention: the calculation of the dynamic stability margin value specifically includes:

[0042] The state variables for the fusion stability assessment are input into the dynamic potential energy surface mapping relationship. The dynamic potential energy surface mapping relationship maps the state variables to a set of parameters describing the local geometry of the virtual potential energy surface, including the virtual potential energy curvature and the virtual gradient direction.

[0043] Starting from the fusion stability assessment state quantity of the previous moment, a multi-step forward deduction is performed on the virtual potential energy surface along the corresponding virtual gradient direction to generate a predicted state evolution trajectory.

[0044] Calculate the reciprocal of the virtual potential energy curvature corresponding to each step in the predicted state evolution trajectory, and extract the minimum value of the reciprocal as the critical curvature index;

[0045] The critical curvature index is divided by a baseline curvature reference value calculated from the current boom macroscopic geometry parameters, and the resulting quotient is output as the dynamic stability margin value.

[0046] The beneficial effects of this invention are:

[0047] (1) By introducing and integrating the articulation coherent disorder index and the transient collision intensity spectrum index, which characterize the micro-dynamic state at the articulation, the key instability induction factor of "gap nonlinear dynamics" is incorporated into the real-time calculation framework for stability assessment. This enables the prediction system to keenly perceive the precursors of system dynamic instability caused by micro-collisions, friction, and parametric resonance before macro-parameter abrupt changes, thereby completely changing the limitation of traditional methods that can only assess the safety margin of static or quasi-static loads, and solving the major safety hazard of not being able to warn of sudden and catastrophic instability accidents under complex dynamic working conditions.

[0048] (2) This invention seamlessly integrates real-time big data acquisition and feature fusion with machine learning-based dynamic system state identification, along with stability margin calculation combining lever theory and potential energy surface derivation, into an automatically operating real-time prediction system. This system not only automates the early warning logic but also, by introducing an identification model with temporal memory capabilities and a forward derivation algorithm based on a virtual potential energy surface, endows the system with the ability to predict stability evolution trends. This reduces reliance on operator experience and enables proactive and advanced risk warnings at early stages that are imperceptible to the human eye and traditional instruments, thereby improving operational safety and equipment intelligence. Attached Figure Description

[0049] The invention will now be further described with reference to the accompanying drawings.

[0050] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Please see Figure 1 As shown, the present invention is a load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar, comprising:

[0053] The multi-source sensor data synchronous acquisition module is used to collect motion sensor data and stress wave sensor data in real time during telescopic boom forklift loading operations;

[0054] The micro-articulation nonlinear characteristic calculation module calculates the cross-spectral coherence function of adjacent arm segment motion sensing data, extracts the main frequency coherence value, distribution entropy and phase stability standard deviation of the cross-spectral coherence function, and generates the articulation coherence disorder index by fusing them according to preset rules.

[0055] The stress wave feature analysis module performs time-frequency transformation on stress wave sensing data and automatically identifies energy impact patches. It extracts the peak energy, center frequency, and duration of the energy impact patches, aggregates them according to the sensitive frequency band mapping rules, and integrates the event frequencies to generate a transient collision intensity spectrum index.

[0056] The fusion state quantity generation module constructs a comprehensive feature vector from the articulated coherent disorder index and the transient collision intensity spectrum index. The comprehensive feature vector is then input into a nonlinear dynamic system identification model built based on a machine learning framework, and the fusion stability assessment state quantity is output.

[0057] The dynamic early warning decision module, based on the theoretical framework of the stress stability of the compression bar, performs forward extrapolation and prediction of the fusion stability assessment state variables to calculate the dynamic stability margin value. When the dynamic stability margin value is lower than the preset adaptive threshold, an early warning signal is generated and issued for the risk of nonlinear jump instability.

[0058] In the multi-source sensor data synchronous acquisition module, motion sensor data is acquired by installing miniature inertial measurement units at the key hinge points of each boom segment of the boom system. These units continuously measure and output the micro-vibration signals of angular velocity and linear acceleration at the hinge points along multiple axes, thereby capturing the relative micro-motion dynamics between boom segments.

[0059] Stress wave sensing data is acquired by installing broadband acoustic emission sensors on or inside the load-bearing structure of adjacent arm sections. These sensors are sensitive to transient stress wave signals generated by microscopic collisions or friction within the structure and can capture high-frequency elastic wave signals released when transient contact events occur at the hinge interface.

[0060] In the micro-hinged nonlinear characteristic calculation module, the calculation of the cross-spectral coherence function begins with data preprocessing. A length of... A sliding time window of seconds, for example, T=0.5 seconds, is used to synchronously acquire two motion sensing data sequences from adjacent arm segments A and B within this window at a fixed sampling frequency (e.g., 1000 Hz), denoted as […]. and ,in This refers to the discrete time sequence number. Subsequently, the two data sequences are subjected to frame-by-frame windowing: each T-second data point is divided into K frames, each containing N sampling points, with a certain proportion of overlap between adjacent frames. A window function (e.g., a Hamming window) is applied to each frame to obtain the windowed time-domain signal frame. and ,in, =1,2,..., Indicates the frame number.

[0061] Perform frequency domain transformation on each windowed signal frame. For the first frame... Frame signal and Performing discrete Fourier transforms on each signal yields the corresponding complex sequences in the frequency domain. Based on these frequency domain representations, the power spectral density can be calculated. and In the Frame, frequency point The cross-power spectral density at point is calculated as follows: and The product of the conjugate complex numbers. Signal In the Frame, frequency point The self-power spectral density at that point is The product of its conjugate complex number, the signal In the Frame, frequency point The self-power spectral density at that point is The product of its conjugate complex number.

[0062] After obtaining the power spectral density of each frame, The average cross-power spectral density is obtained by averaging the cross-power spectral density across all frames. The auto-power spectral density of each frame is averaged over all frames to obtain the average auto-power spectral density. Finally, the complex form of the cross-spectral coherence function is calculated, with the following expression: ;in, It is a complex number whose magnitude (amplitude) is between 0 and 1, and the phase angle represents the phase difference between two signals on this frequency component; Represents the average cross-power spectral density. and Indicates signal and signal Average self-power spectral density.

[0063] After obtaining the cross-spectral coherence function Then, three features can be extracted to generate the articulated coherence disorder index. The first feature is the dominant frequency coherence value. First, based on the structural characteristics of the boom system or historical data analysis, a primary frequency band of interest is preset. Calculate the arithmetic mean of the moduli of the cross-spectral coherence function at all discrete frequency points within the preset dominant frequency band. The calculation expression is as follows: ;in, Indicates the coherence value of the main frequency. This indicates the number of discrete frequency points included within the preset main frequency band. This indicates taking the modulus of a complex number.

[0064] The second characteristic is the distribution entropy. Used to quantize the magnitude of cross-spectral coherence function Throughout the analysis frequency band The uniformity or disorder of the distribution within (e.g., from 0 to half the sampling frequency). First, analyze the entire frequency band. Divided into Each sub-band may have equal or unequal width. Calculate the cross-spectral coherence function magnitude in each sub-band. The average energy on the surface is calculated using the following expression: ;in Representing the Sub-band This represents the number of frequency points within that sub-band. These energy values ​​are then normalized to a probability distribution: Finally, the distribution entropy is calculated according to the definition of information entropy. : ;

[0065] Distribution Entropy The larger the value, the more dispersed and uneven the distribution of cross-spectral coherence mode values ​​across different frequency bands, which may indicate the presence of multiple nonlinear coupling sources or complex frequency response characteristics.

[0066] The third characteristic is the standard deviation of phase stability. Calculate the cross-spectral coherence function. Phase angle at all frequency points within the same preset main frequency band Then, calculate the standard deviation of these phase angle values. The larger the value, the more drastic the fluctuation in the phase relationship between the two signals within the main frequency band, and the worse the consistency of motion.

[0067] After extracting the main frequency coherence value Distribution entropy and phase stability standard deviation After obtaining these three original feature values, they need to be normalized to eliminate the influence of dimensions and bring them into a similar numerical range. For example, a min-max normalization method can be used to determine the maximum value of each feature based on historical data or theoretical ranges. and minimum value Then calculate the normalized eigenvalues. : ;

[0068] in Represents the original feature values. The normalized features are denoted as follows: , and .

[0069] Finally, the three normalized eigenvalues ​​are linearly combined using preset weights, and a weighted sum is calculated. This weighted sum is the final articulated coherence disorder index. The calculation method is as follows: ;in, , and The preset weighting coefficients are used, and they satisfy the following conditions: Due to the coherence value of the main frequency A higher value indicates better synergy, therefore it is preceded by a [missing information - likely a typo]. This is converted into a quantity representing the degree of "disorder," with higher values ​​indicating poorer coordination. The final calculated... The exponent is a unitless scalar. The higher the value, the worse the coordination of the relative micro-motions at the hinge and the higher the degree of nonlinear coupling.

[0070] In the stress wave characteristic analysis module, the time-frequency transformation of the stress wave signal is achieved by performing a continuous wavelet transform on the acquired raw stress wave sensing data. The sampling frequency is set to a value (e.g., 2 MHz), and a time window of data with a length of T seconds (e.g., 0.1 seconds) is acquired. ,in The discrete-time index is used. A complex-valued Morlet wavelet with a well-defined analytical form is chosen as the mother wavelet function. Because of its excellent time-frequency localization characteristics, the continuous wavelet transform is performed by taking the inner product of the signal with a series of scaled and translated mother wavelets. Its mathematical expression is: ;

[0071] in, The wavelet transform coefficients are complex numbers. The scaling factor is inversely proportional to the frequency. The smaller the value, the higher the frequency. This is the translation factor, corresponding to time. This represents the complex conjugate of the mother wavelet. In practice, a scale sequence covering the target analysis frequency band is pre-defined and calculated on the time axis to generate a complex matrix containing M rows (corresponding to different scales / frequencies) and N columns (corresponding to different times), the square of which constitutes the time-frequency energy distribution spectrum of the signal.

[0072] In order to obtain the time-frequency energy distribution map The automatic identification of energy impact patches generated by transient collisions requires the following steps. First, the dynamic background threshold is calculated, and high-energy regions are initially screened: a sliding analysis window is defined on the time-frequency plane, for example, with a width of W time points (approximately 50 microseconds) in the time dimension and a width of WF scale points (approximately 10 kHz frequency band) in the frequency (scale) dimension. For each location in the spectrum... Calculate the average energy value of all points within a local window centered at that location. and standard deviation Dynamic background threshold Defined as: :in The sensitivity coefficient is set to 3. Then, the energy value of each point in the time-frequency energy distribution map is compared with the corresponding dynamic background threshold. If... If so, the point is marked as a potential high-energy point, and a binarized initial high-energy region mask is obtained.

[0073] Morphological operations are performed on the initial high-energy region mask to merge connected regions: the directly obtained binary mask may form multiple discrete, adjacent small regions due to the time-frequency diffusion of noise or the energy of a collision event. Therefore, dilation and erosion operations from mathematical morphology are used for post-processing. First, a structuring element is defined, for example, an element with sizes in the time and frequency dimensions of [missing information]. (e.g., corresponding to 5 microseconds) and (e.g., a rectangle corresponding to 2000 Hz). First, a dilation operation is performed on the binary mask, which expands the high-energy region outward, thereby connecting adjacent isolated small regions. Then, an erosion operation is performed on the dilated result to restore the approximate original boundaries of the regions and smooth the contours. After this series of operations, multiple originally adjacent small regions may be merged into a single connected region. Each ultimately formed connected region, where all points correspond to a continuous region with significantly higher energy than the surrounding background in the original time-frequency energy spectrum, is defined as an "energy impact patch". The time-frequency boundary data of each energy impact patch is output, typically including its row index range (corresponding to the frequency range) and column index range (corresponding to the time range) in the time-frequency matrix.

[0074] After obtaining the time-frequency boundary data of all energy impact patches, three key physical features need to be extracted from each patch. First, the peak energy. ,in This refers to the patch number. The extraction method is as follows: from the original time-frequency energy distribution map... Up, locate to the All points covered by the time-frequency boundary of each patch Find the maximum energy value among these points; this maximum value is the peak energy of the patch. Second, center frequency. Calculate the energy-weighted average of all points within the patch along the frequency (scale) dimension. First, index the scale. Converting the scale-frequency relationship using wavelet transform to the actual frequency value Then calculate: Third, duration Based on the starting column index of the patch on the timeline. and terminating column index Combined with sampling time interval (For example, 0.5 microseconds), calculate the duration: ;

[0075] Based on the preset sensitive frequency band mapping rules, the center frequency of each patch is... Assign a frequency band importance weighting coefficient The mapping rule is constructed based on historical data analysis or structural dynamics characteristics. For example, the entire analysis frequency band is divided into... A continuous frequency band: And assign a preset weighting coefficient to each frequency band. , A positive integer greater than 1. Weighting coefficient. The setting principle is as follows: for frequency bands known to be closely related to structural resonance or harmful impacts, assign higher weights (e.g., 1.5); for frequency bands related to normal friction or background noise and with low harmfulness, assign lower weights (e.g., 0.5); for general frequency bands, assign standard weights (e.g., 1.0). For the [missing information], [missing information] Determine the center frequency of each patch. Which frequency band does it fall into? Then the weight corresponding to the frequency band Assigned to the plaque, that is Finally, within the same time analysis window (e.g., the same 0.1-second time window as the aforementioned time-frequency transformation), a transient collision intensity spectrum index is generated based on the characteristic values ​​of all identified energy impact patches. The calculation process consists of two steps. First, the weighted cumulative energy within the time window is calculated. The formula is as follows: ;in, To calculate the total number of energy impact patches identified within the time window, summation is performed across all patches. This step weights the impact intensity (peak energy) of each patch according to the importance of its frequency. Then, the weighted cumulative energy is calculated. The total number of patches within the window Multiplying these together yields the final transient collision intensity spectrum index: Multiply by the number of patches The purpose is to introduce the "event frequency" dimension, so that the index reflects not only the weighted strength of a single shock, but also the density of shock events. Therefore, The higher the index value, the more frequent, stronger, and more sensitive the transient collision activity at the articulated interface during the analysis time. Its value can be used to directly assess the overall intensity level of micro-collisions.

[0076] In the fusion state variable generation module, the construction of the comprehensive feature vector is a data alignment and concatenation process. Within the same processing time window, the calculated articulated coherent disorder index and transient collision intensity spectrum index are precisely aligned according to their corresponding timestamps. The alignment result at each time point is a feature set containing two elements. To construct a sequence that can be used for time series analysis, the feature sets from multiple consecutive time points are concatenated in chronological order to form a comprehensive feature vector with a dimension of 2. This vector integrates the evolutionary information of the microscopic nonlinear dynamic characteristics over a past period in the time dimension.

[0077] The nonlinear dynamic mapping relationship used to process the comprehensive feature vector and output the fused stability evaluation state quantity consists of multiple recurrent layers with specific internal computational logic connected in series, followed by a fully connected layer. Each recurrent layer employs a gated recurrent unit design, the core of which is to selectively retain historical information or forget irrelevant information through update gates and reset gates, thereby achieving long-term and short-term dependency modeling of the input time-series data. Multiple such gated recurrent unit layers are stacked sequentially, with the hidden state sequence output by the previous layer serving as the input to the next layer. This multi-layer structure can abstract the input time-series features layer by layer, capturing complex features from low-level dynamics to high-level dynamic patterns. The hidden state vector output by the last gated recurrent unit layer at the final time step is regarded as a condensed representation of the dynamics of the entire input sequence. Subsequently, the hidden state vector is fed into a fully connected projection layer, which consists of multiple neurons with nonlinear activation functions. Its function is to map (or decode) the high-dimensional hidden state vector into a lower-dimensional output vector with a clearer physical meaning. This output vector is defined as the fusion stability evaluation state quantity, which comprehensively expresses the coupling state of the system's macroscopic stability and microscopic nonlinear dynamics at the current moment.

[0078] To ensure accurate mapping capabilities, supervised training is required. The training dataset was constructed through high-fidelity dynamic simulation. The simulation simulated the movement of the telescopic boom under various working conditions, including normal operation and scenarios induced by human intervention to the instability boundary, generating a large amount of time-series data. For each sampling moment in the simulation, a corresponding comprehensive feature vector was synchronously calculated based on its state. Using the precisely obtainable system state parameters from the simulation (such as the actual stress, strain energy, and Lyapunov exponential components of each part of the boom), a calibrated state vector was calculated through a pre-defined physical relationship. This calibrated state vector served as the benchmark value for the fused stability assessment state quantity at that moment, i.e., the training target label. Thus, the dataset consists of many time-series sequences, each containing multiple consecutive time points, with each time point having a comprehensive feature vector and a corresponding benchmark value for the fused stability assessment state quantity.

[0079] The model is trained sequentially. For a sequence in the training data, at any given training time step, the model's input consists of two parts: the current integrated feature vector and the predicted value of the fusion stability evaluation state variable output by the model at the previous time step (initialized to a zero vector at the beginning of training or at the start of each sequence). The model operates as described above, ultimately outputting the predicted value of the fusion stability evaluation state variable at the current time step. The training objective is to make this predicted value as close as possible to the known baseline value of the fusion stability evaluation state variable at the current time step. By providing the model with a large amount of such sequence data and continuously adjusting all parameters within the model (including the weights and biases in each gated recurrent unit layer and fully connected projection layer), its predictive ability is optimized.

[0080] The criterion used to guide parameter adjustment during training is a composite loss function. This function consists of the sum of two main parts. The first part is the mean squared error loss, which calculates the average of the squares of the differences between the predicted values ​​of the fusion stability assessment state variables output by the model and the true baseline values ​​in each corresponding dimension. This part directly constrains the static accuracy of the model output. The second part is the temporal smoothness loss, which calculates the norm of the difference between the predicted values ​​of the fusion stability assessment state variables output by the model at two adjacent time points. This part encourages the model output to change smoothly over time, avoiding drastic fluctuations, thereby enhancing its continuity in trend prediction. The total composite loss function is the sum of these two loss parts multiplied by a preset weight coefficient. By using the backpropagation algorithm and gradient descent optimizer, the value of this composite loss function is iteratively minimized, thus completing the effective training of the model parameters. The trained model then possesses the ability to dynamically and accurately infer the current system's fusion stability assessment state variables based on the real-time input micro-nonlinear dynamic feature sequence.

[0081] In the dynamic early warning decision module, a dynamic potential energy surface mapping relationship is defined. This relationship is a pre-defined mathematical function whose input is the current fusion stability assessment state quantity (a multi-dimensional vector). This function calculates through a series of nonlinear transformations and outputs two key geometric parameters. The first parameter is the virtual potential energy curvature, a scalar value. It is calculated by taking the square root of the weighted sum of the squares of the input state quantity's dimensions, then adjusting it with an exponential decay function. This value intuitively reflects the steepness of the curvature of the local region of the potential energy surface where the state quantity is located at its current "position." A larger curvature value indicates a steeper local surface and a stronger tendency for the system to restore equilibrium. The second parameter is the virtual gradient direction, a vector with the same dimension as the input state quantity. It is calculated by taking the partial derivatives of a virtual potential energy function (which is a weighted sum of the squares of the input state quantity's dimensions), inverting them, and then normalizing the result. This vector indicates the direction of the steepest descent along the potential energy surface at the current "position."

[0082] A multi-step forward inference is performed to generate the predicted state evolution trajectory. The fused stability assessment state quantity output at the previous calculation time is used as the starting point for the inference. A fixed, small inference step size is set based on the virtual gradient direction calculated from the current state quantity. Starting from the starting point, the system moves one step along the current virtual gradient direction to obtain the coordinates of a new point, which is considered as the possible state of the system a very short time interval in the future. Then, using this new point as input, the virtual gradient direction at the new point is recalculated through the aforementioned dynamic potential energy surface mapping relationship. Next, the system moves one step along this new direction to obtain the next point. This process is repeated several times (e.g., 10 times), and the series of state points obtained in chronological order are connected to form a future short-term state evolution trajectory predicted from the current state based on the local geometry of the current potential energy surface.

[0083] Calculate the critical curvature index. On the generated state evolution trajectory, for each state point, calculate the virtual potential energy curvature value corresponding to that point using the dynamic potential energy surface mapping relationship. Then, calculate the reciprocal of each curvature value (i.e., divide the curvature value by a constant 1). Traverse the entire trajectory and find the minimum value among these reciprocals. This minimum value is defined as the critical curvature index. The significance of this index lies in its identification of the weakest restoring force corresponding to the flattest potential energy surface region (i.e., the point of minimum curvature) that the system may experience along the predicted evolution path; this point is considered the most vulnerable link in stability.

[0084] Calculate the dynamic stability margin. This calculation requires a reference curvature value as the denominator. The reference curvature value is calculated in real-time from the current macroscopic geometric parameters of the boom. The specific calculation formula is as follows: taking the current working length, current elevation angle, and rated load of the boom system as inputs, a scalar value is calculated through an analytical expression containing polynomial and trigonometric function terms. This value theoretically represents the characteristic curvature corresponding to the linear buckling critical state of an ideal rigid column under the current macroscopic configuration. The dynamic stability margin value is the quotient obtained by dividing the critical curvature index calculated above by this reference curvature value. This margin value is a dimensionless number. When its value is greater than 1, it indicates that the curvature of the predicted weakest link in the system is greater than the linear buckling reference, and the system is stable; when the value is less than 1, it indicates that there is a weak point on the predicted path that is flatter than linear buckling. Set a preset adaptive threshold (e.g., 0.8). When the dynamic stability margin value calculated in real time is lower than the threshold, it is determined that the system has a high risk of entering a jump instability state dominated by nonlinear dynamics. Then, a specific warning command signal is generated and prompted through the human-computer interaction interface.

[0085] The working principle of this invention is as follows: Motion and stress wave data are synchronously acquired through sensors deployed at the boom articulation points. First, a cross-spectral coherence function is calculated based on the motion data, and the dominant frequency coherence value, distribution entropy, and phase stability standard deviation are extracted and fused to generate an articulation coherence disorder index characterizing the articulation coordination and nonlinearity. Second, continuous wavelet transform is performed on the stress wave data to obtain a time-frequency spectrum, automatically identifying energy impact patches and extracting their peak energy, center frequency, and duration. Event frequencies are aggregated and synthesized according to sensitive frequency band mapping rules to generate a transient collision intensity spectrum index characterizing the collision intensity. Then… The two indices mentioned above are constructed as a time-series comprehensive feature vector, which is input into a pre-trained multi-layer gated recurrent unit structure for nonlinear dynamic system identification. The output is a fusion stability assessment state quantity that integrates macroscopic and microscopic dynamics. Finally, this state quantity is mapped to a virtual potential energy surface for multi-step forward extrapolation. By calculating the ratio of the inverse of the curvature of the weakest link in the trajectory to the linear buckling reference curvature, the dynamic stability margin value is obtained. When this value is lower than the adaptive threshold, an early warning signal for nonlinear jump instability is issued, thereby achieving real-time and accurate prediction of macroscopic stability risks from microscopic nonlinear dynamic characteristics.

[0086] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A load stability prediction system for telescopic boom forklifts based on lever stress, characterized in that, include: The multi-source sensor data synchronous acquisition module is used to collect motion sensor data and stress wave sensor data in real time during telescopic boom forklift loading operations; The micro-articulation nonlinear characteristic calculation module calculates the cross-spectral coherence function of adjacent arm segment motion sensing data, extracts the main frequency coherence value, distribution entropy and phase stability standard deviation of the cross-spectral coherence function, and generates the articulation coherence disorder index by fusing them according to preset rules. The stress wave feature analysis module performs time-frequency transformation on stress wave sensing data and automatically identifies energy impact patches. It extracts the peak energy, center frequency, and duration of the energy impact patches, aggregates them according to the sensitive frequency band mapping rules, and integrates the event frequencies to generate a transient collision intensity spectrum index. The fusion state quantity generation module constructs a comprehensive feature vector from the articulated coherent disorder index and the transient collision intensity spectrum index. The comprehensive feature vector is then input into a nonlinear dynamic system identification model built based on a machine learning framework, and the fusion stability assessment state quantity is output. The dynamic early warning decision module, based on the theoretical framework of the stress stability of the compression bar, performs forward extrapolation and prediction of the state variables of the integrated stability assessment, and calculates the dynamic stability margin value. When the dynamic stability margin value is lower than the preset adaptive threshold, an early warning signal is generated and issued for the risk of nonlinear jump instability.

2. The load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar as described in claim 1, characterized in that, The calculation of the cross-spectral coherence function for adjacent arm segment motion sensing data specifically includes: Acquire motion sensing data of adjacent arm segments within a specified time window; The motion sensing data is processed by frame-by-frame windowing to obtain the time-domain signal after multi-frame windowing. Perform Fourier transform on the windowed time-domain signal of each frame to obtain the corresponding frequency-domain complex sequence; Calculate the cross power spectral density and the individual self power spectral density of each frame signal at the corresponding frequency point; Based on the cross-power spectral density and the self-power spectral density, the complex form of the cross-spectral coherence function is calculated point by point.

3. The load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar as described in claim 1, characterized in that, The generation of the articulated coherence disorder index specifically includes: The average modulus of the preset main frequency band is extracted from the cross-spectral coherence function in complex form as the main frequency coherence value; The entropy of the distribution of the modulus of the cross-spectral coherence function across all analysis frequency bands is calculated as the distribution entropy; The standard deviation of the phase angle of the cross-spectral coherence function within the preset main frequency band is used as the standard deviation of phase stability. The coherence value of the main frequency, the distribution entropy, and the standard deviation of the phase stability are normalized. The three normalized eigenvalues ​​are linearly combined according to preset weights, the weighted sum is calculated, and the weighted sum is output as the articulated coherence disorder index.

4. The load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar as described in claim 1, characterized in that, The process of performing time-frequency transformation on stress wave sensing data and automatically identifying energy impact patches specifically includes: Continuous wavelet transform is performed on the acquired stress wave sensing data to generate a time-frequency distribution spectrum; Sliding window energy statistics are performed on the time-frequency distribution map to identify local regions where the energy exceeds the dynamic background threshold. Morphological expansion and erosion operations are performed on the identified adjacent high-energy regions to merge them into connected regions. Each connected component is defined as an energy impact patch, and the corresponding time-frequency boundary data is output.

5. The load stability prediction system for telescopic boom forklifts based on lever force as described in claim 1, characterized in that, The generation of the transient collision intensity spectrum index specifically includes: Based on the time-frequency boundary data of the energy impact patches, the peak energy, center frequency, and duration of each patch are extracted. According to the preset sensitive frequency band mapping rules, the center frequency of each patch is assigned a corresponding frequency band importance weighting coefficient; The weighted cumulative energy is obtained by multiplying the peak energy of all patches within the same analysis window by the frequency band importance weighting coefficient and summing the results. The weighted cumulative energy is multiplied by the total number of patches within the time analysis window, and the product is output as the transient collision intensity spectrum index.

6. The load stability prediction system for telescopic boom forklifts based on lever force as described in claim 1, characterized in that, The output fusion stability evaluation state variables specifically include: The articulated coherent disorder index and the transient collision intensity spectrum index are aligned and spliced ​​according to the timestamp to construct a comprehensive feature vector containing microscopic nonlinear dynamic characteristics. The comprehensive feature vector is input into a pre-trained nonlinear dynamic system identification model based on a machine learning framework. The nonlinear dynamic system identification model includes multiple gated recurrent unit layers connected in series for multi-layer temporal abstraction of the comprehensive feature vector; and a feedforward projection layer for mapping the hidden state vector output by the last gated recurrent unit to a fused stability evaluation state quantity as the output of the nonlinear dynamic system identification model.

7. The load stability prediction system for telescopic boom forklifts based on the force exerted by the compression bar as described in claim 1, characterized in that, The training process of the nonlinear dynamic system identification model is as follows: Construct a training dataset containing time series generated by dynamic simulations that cover normal operation and instability boundaries. Each time series point contains a corresponding integrated feature vector and a baseline value of the fused stability evaluation state quantity as the target label. The structure composed of multiple gated recurrent unit layers and feedforward projection layers is trained in a supervised manner, using the comprehensive feature vector and the fusion stability evaluation state quantity of the previous time step as input, and the benchmark value of the fusion stability evaluation state quantity of the current time step as the target. During training, the composite loss function is optimized. The composite loss function simultaneously constrains the approximation accuracy of the fusion stability evaluation state quantity output by the feedforward projection layer to the baseline value of the fusion stability evaluation state quantity and the temporal continuity of the instability trend prediction.

8. The load stability prediction system for telescopic boom forklifts based on lever force as described in claim 1, characterized in that, The calculation of the dynamic stability margin value specifically includes: The state variables for the fusion stability assessment are input into the dynamic potential energy surface mapping relationship. The dynamic potential energy surface mapping relationship maps the state variables to a set of parameters describing the local geometry of the virtual potential energy surface, including the virtual potential energy curvature and the virtual gradient direction. Starting from the fusion stability assessment state quantity of the previous moment, a multi-step forward deduction is performed on the virtual potential energy surface along the corresponding virtual gradient direction to generate a predicted state evolution trajectory. Calculate the reciprocal of the virtual potential energy curvature corresponding to each step in the predicted state evolution trajectory, and extract the minimum value of the reciprocal as the critical curvature index; The critical curvature index is divided by a baseline curvature reference value calculated from the current boom macroscopic geometry parameters, and the resulting quotient is output as the dynamic stability margin value.

Images