Real-time monitoring and early warning method and device for lifting hooks based on multi-source data fusion

Abstract

This invention provides a real-time monitoring and early warning method and device for crane hooks based on multi-source data fusion, relating to the field of artificial intelligence technology. It adaptively segments the real-time monitoring sequence based on the velocity envelope fluctuation intensity of the lifting height sequence, avoiding feature aliasing caused by start-stop, speed change, and other operations. Based on the sampling point differential sequence, data from each channel is coupled, quantifying the intensity of synchronous changes between load, lifting, and oscillation into comparable anomaly sensitivity indicators. Based on the difference in motion trends between the coupling characteristics and the aforementioned segmented characteristics, a trend matrix and a residual matrix are decomposed. This allows the separation of the operator's active operational intent from the normal inertial motion of the load, enabling the residual matrix to centrally reflect abnormal fluctuations such as mechanical clearance and elastic oscillations. The model can effectively distinguish between normal fluctuations and early hidden dangers under strong dynamic and interference environments, achieving adaptive, high-sensitivity, and low-false-alarm-rate real-time anomaly identification, ensuring the inherent safety of lifting operations.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and device for real-time monitoring and early warning of lifting hooks based on multi-source data fusion. Background Technology

[0002] As a key load-bearing component in material handling, crane hooks frequently undergo complex operating conditions such as lifting, translation, deceleration, braking, and swinging during operation. They are subjected to alternating loads and occasional impacts over a long period of time, making them highly susceptible to plastic deformation or fatigue cracks due to continuous overload, or structural damage due to sudden stop collisions. Once the hook fails, it will directly lead to the falling of heavy objects, causing significant casualties and property losses. Therefore, real-time and accurate monitoring and early warning of the hook's operating status has important engineering value.

[0003] Currently, crane hook monitoring mainly relies on periodic non-destructive testing or setting fixed threshold alarms for single physical quantities such as load and height. In daily operations, operators often rely on experience to judge the hook's condition. This approach typically uses fixed time windows to capture sensor data and performs simple comparisons based on manually extracted statistical features (such as peak values ​​and average values). However, during the frequent lifting, translation, deceleration, braking, and swinging processes in crane operations, sensor signals are mixed with a large number of load fluctuations, height changes, and horizontal swings caused by normal operations. Fixed time windows easily overlap data from different motion stages, causing abnormal features caused by impacts and off-center loads to be submerged by normal inertial changes; while relying solely on a single threshold cannot distinguish between "normal fluctuations caused by operation" and "early signs of structural danger," easily leading to frequent false alarms at start-up and shutdown, or missing minute cracks or plastic deformations due to excessive signal smoothing. These limitations make it difficult for existing methods to achieve real-time and accurate hook condition early warning in complex and highly interfering operating environments, posing safety hazards. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time monitoring and early warning method and device for lifting hooks based on multi-source data fusion, which can effectively distinguish between normal fluctuations and early hidden dangers under strong dynamic and strong interference environments, achieve adaptive, high-sensitivity, and low false alarm rate real-time anomaly identification, and ensure the inherent safety of lifting operations.

[0005] In a first aspect, embodiments of the present invention provide a real-time monitoring and early warning method for lifting hooks based on multi-source data fusion, wherein the method includes: acquiring a real-time monitoring sequence of the lifting hook of a lifting device during the lifting process; wherein the real-time monitoring sequence includes a load sequence, a lifting height sequence, and a horizontal displacement sequence; based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in a preset local interval, dividing the real-time monitoring sequence into multiple monitoring matrices; each monitoring matrix in the multiple monitoring matrices respectively includes a first segment sequence corresponding to the load sequence, a second segment sequence corresponding to the lifting height sequence, and a third segment sequence corresponding to the horizontal displacement sequence; based on The sampling point difference sequences corresponding to the first, second, and third segment sequences are used to perform feature coupling on each monitoring matrix segment to construct corresponding motion intensity coupling features. Based on the difference between the motion intensity coupling features and the motion trend of each monitoring matrix segment, the trend matrix and residual matrix contained in the lifting hook during the lifting process are determined. Based on the trend matrix and residual matrix, the kinematic enhancement feature matrix corresponding to the real-time monitoring sequence is constructed, and the kinematic enhancement feature matrix is ​​input into the pre-constructed hook state recognition model. The hook state recognition model identifies abnormal signals in the kinematic enhancement feature matrix to determine the abnormal identification result of the lifting hook.

[0006] In conjunction with the first aspect, this embodiment of the invention also provides a first implementation of the first aspect, wherein the step of dividing the real-time monitoring sequence into a multi-segment monitoring matrix based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in a preset local interval includes: determining the lifting speed envelope of the lifting height sequence based on the absolute value of the difference between adjacent sampling points of the lifting height sequence; determining the segment boundary set of the lifting height sequence based on the fluctuation intensity of the lifting speed envelope in the preset local interval; and segmenting the load sequence, lifting height sequence, and horizontal displacement sequence of the real-time monitoring sequence according to the segment boundary set to form a multi-segment monitoring matrix.

[0007] In conjunction with the first aspect, this embodiment of the invention also provides a second implementation of the first aspect, wherein the step of performing feature coupling on each segment of the monitoring matrix based on the sampling point difference sequences corresponding to the first segment sequence, the second segment sequence, and the third segment sequence to construct the corresponding motion intensity coupling features includes: performing intra-segment local normalization on the first segment sequence, the second segment sequence, and the third segment sequence in the multi-segment monitoring matrix to obtain a normalized segmented monitoring matrix; determining the difference sequences corresponding to the first segment sequence, the second segment sequence, and the third segment sequence based on the absolute value of the first-order difference between adjacent sampling points in the normalized segmented monitoring matrix; and weighting and summing the difference sequences based on preset fusion weights to construct the motion intensity coupling features corresponding to the first segment sequence, the second segment sequence, and the third segment sequence.

[0008] In conjunction with the first aspect, this embodiment of the invention also provides a third implementation of the first aspect, wherein the step of determining the trend matrix and residual matrix contained in the lifting hook during the lifting process based on the motion intensity coupling characteristics and the difference in motion trends of each segment monitoring matrix includes: based on the segmented normalized values ​​of the motion intensity coupling characteristics, performing weighted moving averages on the first segment sequence, the second segment sequence, and the third segment sequence of each segment monitoring matrix to construct a reconstructed trend matrix corresponding to each segment monitoring matrix; and subtracting each segment monitoring matrix from the reconstructed trend matrix element by element to determine the trend matrix and residual matrix contained in each segment monitoring matrix.

[0009] In conjunction with the first aspect, this invention also provides a fourth implementation of the first aspect, wherein the step of constructing a kinematic enhancement feature matrix corresponding to a real-time monitoring sequence based on a trend matrix and a residual matrix includes: performing peak enhancement on the residual matrix to obtain an enhanced residual envelope matrix; performing feature concatenation of the enhanced residual envelope matrix with the reconstructed trend matrix of each monitoring segment to obtain a piecewise dynamic enhancement feature matrix; and concatenating the piecewise dynamic enhancement feature matrices corresponding to each monitoring segment in the multiple monitoring matrices to construct a kinematic enhancement feature matrix of the real-time monitoring sequence.

[0010] In conjunction with the first aspect, this invention also provides a fifth implementation of the first aspect, wherein the step of inputting the kinematic enhancement feature matrix into a pre-constructed hook state recognition model, and identifying abnormal signals in the kinematic enhancement feature matrix through the hook state recognition model to determine the abnormal identification result of the lifting hook includes: time-aligning the first channel data and the second channel data based on the normalized cross-correlation value between preset first channel data and preset second channel data in the kinematic enhancement feature matrix to construct a time-aligned fusion feature matrix; performing convolution operations on the time-aligned fusion feature matrix through the pre-constructed hook state recognition model to determine the continuous overload abnormal signal and short-term disturbance feature in the time-aligned fusion feature matrix; and performing abnormal category probability prediction on the kinematic enhancement feature matrix based on the continuous overload abnormal signal and short-term disturbance feature to determine the abnormal identification result corresponding to the lifting hook.

[0011] In conjunction with the first aspect, this invention also provides a sixth implementation of the first aspect, wherein the step of determining the continuous overload anomaly signal and short-term disturbance features in the time-aligned fusion feature matrix by performing convolution operations on the time-aligned fusion feature matrix through a pre-constructed hook state recognition model includes: inputting the reconstructed trend channel group data contained in the time-aligned fusion feature matrix into the trend convolution branch to extract the continuous overload anomaly signal in the time-aligned fusion feature matrix; inputting the reconstructed trend channel group data and the enhanced residual envelope channel group data contained in the time-aligned fusion feature matrix into the transient convolution branch to extract the short-term disturbance features in the time-aligned fusion feature matrix; wherein the reconstructed trend channel group data is the data corresponding to the reconstructed trend matrix of each monitoring matrix segment, and the enhanced residual envelope channel group data is the data corresponding to the residual matrix of each monitoring matrix segment.

[0012] In conjunction with the first aspect, the present invention also provides a seventh implementation of the first aspect, wherein the method further includes: determining the transient intensity sequence corresponding to the time-aligned fusion feature matrix based on the enhanced residual envelope channel data contained in the time-aligned fusion feature matrix; mapping the transient intensity sequence to a gating sequence, so as to determine the convolution output retention ratio corresponding to the convolution operation based on the gating sequence.

[0013] In conjunction with the first aspect, this embodiment of the invention also provides an eighth implementation of the first aspect, wherein the step of constructing a reconstruction trend matrix corresponding to each monitoring matrix by performing weighted moving averages on the first segment sequence, the second segment sequence, and the third segment sequence of each monitoring matrix based on the segmented normalized values ​​of the motion intensity coupling features includes: mapping the motion intensity coupling features to a local smoothing radius based on the motion intensity magnitude represented by the segmented normalized values; and using the local smoothing radius to perform weighted moving averages on the first segment sequence, the second segment sequence, and the third segment sequence to construct a reconstruction trend matrix corresponding to each monitoring matrix.

[0014] Secondly, embodiments of the present invention provide a real-time monitoring and early warning device for lifting hooks based on multi-source data fusion, wherein the device includes: a data acquisition module, used to acquire a real-time monitoring sequence of the lifting hook of the lifting equipment during the lifting process; wherein the real-time monitoring sequence includes a load sequence, a lifting height sequence, and a horizontal displacement sequence; a data segmentation module, used to divide the real-time monitoring sequence into multiple monitoring matrices based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in a preset local interval; each monitoring matrix in the multiple monitoring matrices respectively includes a first segment sequence corresponding to the load sequence, a second segment sequence corresponding to the lifting height sequence, and a third segment sequence corresponding to the horizontal displacement sequence; and a feature coupling module. The system is divided into three modules: a first module for coupling features of the monitoring matrix and an execution module for constructing motion intensity coupling features. The first module is used to perform feature coupling on each monitoring matrix segment based on the difference sequences of the sampling points corresponding to the first, second, and third segment sequences, respectively. The second module is used to determine the trend matrix and residual matrix contained in the lifting hook during the lifting process based on the difference between the motion intensity coupling features and the motion trend of each monitoring matrix segment. The third module is used to construct the kinematic enhancement feature matrix corresponding to the real-time monitoring sequence based on the trend matrix and residual matrix, and input the kinematic enhancement feature matrix into the pre-constructed hook state recognition model. The hook state recognition model is used to identify abnormal signals in the kinematic enhancement feature matrix and determine the abnormal identification result of the lifting hook.

[0015] The embodiments of this invention bring the following beneficial effects: This invention provides a real-time monitoring and early warning method and device for crane hooks based on multi-source data fusion, relating to the field of artificial intelligence technology. It utilizes the velocity envelope fluctuation intensity of the lifting height sequence in local intervals to adaptively segment the monitoring sequence, ensuring that each segment of the monitoring matrix corresponds to a subprocess with a relatively consistent physical motion state, avoiding feature aliasing caused by normal operations such as start-stop and speed changes. Furthermore, by calculating the sampling point difference sequence and performing feature coupling for the segmented sequences in the three directions of each monitoring matrix, the intensity of synchronous changes between load, lifting, and swaying can be quantified into comparable anomaly sensitivity indicators, thereby initially capturing dangerous modes such as off-center loading, impact, or abnormal swaying. Based on this, each segment matrix is ​​decomposed into a trend matrix and a residual matrix according to the motion intensity coupling characteristics and the difference in motion trend of that segment. This allows the separation of the operator's active operational intent and the main trend formed by the normal inertial motion of the load, enabling the residual matrix to centrally reflect unexpected local abnormal fluctuations such as mechanical clearance, wire rope elastic oscillation, and wind load impact. Furthermore, by inputting the kinematic enhancement feature matrix jointly constructed from the trend matrix and residual matrix into the hook state recognition model, the model can amplify the signal-to-noise ratio of weak fault signals through physical mechanisms-driven residual amplification while preserving the global motion background. This effectively solves the technical problem of traditional methods struggling to distinguish between normal fluctuations and early-stage potential hazards in highly dynamic and interference-prone environments. In summary, this invention enables adaptive, highly sensitive, and low-false-alarm-rate real-time anomaly recognition of lifting hooks, ensuring the inherent safety of lifting operations.

[0016] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

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

[0019] Figure 1 A flowchart illustrating a real-time monitoring and early warning method for lifting hooks based on multi-source data fusion, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a real-time monitoring sequence provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the lifting speed envelope provided in an embodiment of the present invention; Figure 4 A schematic diagram of the reconstruction trend and residual of a load channel provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of an enhanced residual envelope provided in an embodiment of the present invention; Figure 6 A schematic diagram of the structure of a real-time monitoring and early warning device for lifting hooks based on multi-source data fusion provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] To facilitate understanding, the real-time monitoring and early warning method for lifting hooks based on multi-source data fusion provided in this embodiment of the invention will be described first, referring to... Figure 1 The method includes the following steps: Step S102: Obtain the real-time monitoring sequence of the lifting hook of the lifting equipment during the lifting process. Based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in the preset local interval in the real-time monitoring sequence, divide the real-time monitoring sequence into a multi-segment monitoring matrix.

[0023] The real-time monitoring sequence of the lifting hook during the lifting process is used to characterize the dynamic force and motion state of the hook in three-dimensional space. Among them, the load sequence reflects the change in the vertical tension on the hook, which can determine whether it is overloaded or impacted; the lifting height sequence records the displacement trajectory of the hook along the vertical direction, which can indirectly estimate the lifting speed and acceleration; and the horizontal displacement sequence describes the swing and movement amplitude of the hook or load in the plane, which can be used to identify lateral impacts caused by off-center loading and sudden stops.

[0024] In one embodiment, Figure 2 This is a schematic diagram of the sequence for real-time monitoring. Figure 2 The data from each channel of the original hook monitoring sequence (real-time monitoring sequence) are displayed, including the original waveforms of the load channel (kN), lifting height channel (m), and horizontal displacement channel (m) changing over time. Figure 2 (a) shows the original waveform of the load channel (kN) over time. Figure 2 (b) shows the original waveform of the lifting height channel (m) changing over time. Figure 2 (c) in the figure represents the original waveform of the horizontal displacement channel (m) over time. During actual operation of the lifting hook, the original data forms of load, height, and horizontal displacement, including the continuous overload segment (approximately 10-18 seconds) and the impact vibration segment (approximately 22-24 seconds), reflect the physical coupling relationship of the three channels and the waveform characteristics of abnormal states. Coordinate axes: The horizontal axis represents time (seconds), and the vertical axes represent load (kN), lifting height (m), and horizontal displacement (m), respectively.

[0025] In actual operations, due to frequent switching between operations such as hoisting, constant speed, braking, hook stabilization, and horizontal movement, each sequence often exhibits strong non-stationary characteristics. Existing fixed time windows or equal-length truncation methods easily mix high-speed hoisting segments and low-speed swinging segments within the same analysis unit, causing the normal inertial changes of the load to be confused with abnormal impact characteristics. This makes it difficult to distinguish between normal fluctuations caused by operations and early signs of structural damage, regardless of whether threshold alarms or simple statistical features are used. To address this limitation, this invention introduces an adaptive segmentation strategy based on the fluctuation intensity of the hoisting speed envelope. Based on determining the hoisting speed of the hoisting height sequence within a preset local interval and quantifying the fluctuation intensity of its envelope signal (such as local variance or range), the boundary points of motion state switching such as start-up, speed change, braking, and hook stabilization can be dynamically identified, thereby automatically segmenting the original continuous real-time monitoring sequence into multiple monitoring matrices. Each monitoring matrix corresponds to a sub-process with a relatively consistent motion state. Each matrix segment contains a first segment sequence corresponding to the load sequence, a second segment sequence corresponding to the lifting height sequence, and a third segment sequence corresponding to the horizontal displacement sequence. Each segment not only preserves the synchronous correspondence of each physical quantity within the same time interval but also eliminates feature aliasing between different motion modes at the signal source. In summary, in the actual monitoring of crane hooks, the signals of the rapid lifting phase, the constant speed phase, and the braking phase can be automatically separated, ensuring that inertial fluctuations generated by normal speed change operations do not mix into the steady-state operation data, thereby guaranteeing the authenticity and independence of the monitoring features under each motion mode.

[0026] Step S104: Based on the sampling point difference sequences corresponding to the first segment sequence, the second segment sequence, and the third segment sequence, feature coupling is performed on each monitoring matrix segment to construct the corresponding motion intensity coupling feature.

[0027] Each segment of the monitoring matrix contains three segmented sequences: load, lifting height, and horizontal displacement. These sequences can characterize the dynamic behavior of the hook from three dimensions: vertical force, lifting motion, and horizontal sway. However, the absolute value or rate of change of a single sequence is often insufficient to independently represent hazardous modes. For example, large load fluctuations may be caused by normal lifting inertia, while a sudden increase in horizontal displacement may be the result of active translation by the operator. Therefore, it is necessary to couple the three sequences to extract the intensity of synchronous changes among them, thereby distinguishing between "multi-directional linkages caused by normal operation" and "uncoordinated movements caused by abnormal impacts or off-center loading." In this embodiment, the first-order difference between adjacent sampling points is calculated for each segmented sequence, yielding three difference sequences: load change rate, lifting speed, and horizontal sway speed. Furthermore, the three difference values ​​at the same moment can be vectorized and combined (e.g., by calculating their Euclidean norm or weighted sum of squares) to form a scalarized motion intensity coupling feature sequence, quantifying the comprehensive intensity of synchronous changes in load, lifting, and sway within a local time window. For example, when the three are linked in a normal kinematic relationship (such as lifting accompanied by a steady increase in load and minimal horizontal displacement), the coupling characteristics are at a low level; however, once a sudden change in load occurs while the height remains almost unchanged, or a drastic horizontal displacement occurs while the load remains unchanged, the coupling characteristics will increase significantly, thus allowing for the initial identification of dangerous modes such as off-center loading, impact, or abnormal swaying.

[0028] Step S106: Based on the motion intensity coupling characteristics and the difference in motion trend of each monitoring matrix segment, determine the trend matrix and residual matrix contained in the lifting hook during the lifting process.

[0029] The motion intensity coupling characteristic is a comprehensive measure of the intensity of synchronous changes in load, lifting height, and horizontal displacement within the monitoring matrix of a given segment, reflecting the degree of coordination of the overall motion of that segment. Each monitoring matrix records the actual time-domain waveforms of three channels. When the coupling characteristic indicates that the overall motion of the segment is stable and the three axes are synchronized, short-term spikes or rapid reverse oscillations may appear in the load or horizontal displacement channels of the actual monitoring matrix; conversely, when the coupling characteristic indicates that the motion of the segment is violent, there may also be local jitters in the actual monitoring matrix that are inconsistent with the violent trend. The essence of this difference is the unexpected fluctuation in the actual trajectory that deviates from the overall coordinated law.

[0030] In summary, based on this difference in motion trends, the embodiments of the present invention can extract the main components of the signal that conform to the overall coordinated law into a trend matrix, which is used to describe the slow trend changes formed by continuous overload or normal operation; and extract the unexpected components that deviate from the overall coordinated law into a residual matrix, which is used to bear the high-frequency spike components caused by impact vibration. Through this difference-driven decomposition, the originally superimposed slow-changing signals and fast-changing disturbances can be separated, enabling the monitoring system to identify both slowly developing plastic deformation trends and accurately capture instantaneous impact anomalies. This provides clear and reliable early warning information for on-site operators, avoids false alarms or missed alarms caused by signal aliasing, and improves the monitoring reliability and intrinsic safety level under complex dynamic working conditions.

[0031] Step S108: Construct the kinematic enhancement feature matrix corresponding to the real-time monitoring sequence based on the trend matrix and residual matrix, and input the kinematic enhancement feature matrix into the pre-constructed hook state recognition model. The abnormal signals in the kinematic enhancement feature matrix are identified by the hook state recognition model to determine the abnormal identification result of the lifting hook.

[0032] The trend matrix preserves the macroscopic trajectory of the hook during lifting, dominated by active operation and inertial motion, while the residual matrix concentrates unexpected local fluctuations that deviate from this macroscopic trajectory (such as mechanical clearance, elastic oscillations, and impact spikes). The kinematic enhancement feature matrix formed by both can preserve the overall background of the working condition and highlight abnormal signal components through physical decomposition. Compared with directly using the original sequence or a single difference feature, it can significantly improve the signal-to-noise ratio of weak fault signals.

[0033] Furthermore, this feature matrix is ​​input into a pre-trained hook state recognition model. The model does not need to distinguish between normal operation and abnormal fluctuations; instead, it directly performs pattern recognition on the abnormal signals in the residual matrix within the enhanced feature space. In one implementation, the model can output whether the hook currently exhibits abnormal types such as cracks, plastic deformation, off-center loading, or impact overload, along with their confidence levels, ultimately determining the abnormality identification result of the lifting hook. In actual lifting operation scenarios, it can adaptively follow changes in the hook's motion state, accurately identifying early potential hazards in highly dynamic and interference-prone environments, providing timely and reliable safety warnings for operators.

[0034] Furthermore, based on the above embodiments, this invention also provides another real-time monitoring and early warning method for lifting hooks based on multi-source data fusion, to illustrate each step of the above embodiments. For the above real-time monitoring sequence, continuous monitoring data can be collected from load sensors, lifting height encoders, and horizontal displacement sensors at a uniform sampling interval. The sampling interval is denoted as... , This represents the time difference between two adjacent sampling points, and the unit can be seconds. In a convenient implementation method... The timeout can be set to 0.1s or 0.2s to balance transient impact capture capability and system storage overhead.

[0035] Furthermore, a sliding time window can be used to segment the continuous monitoring data. The window length is denoted as... , This represents the total number of sampling points contained in a single sample. In one implementation, A value of 256 can be chosen, meaning each sample contains 256 consecutive sampling points; the window step size can be 64 or 128, allowing for partial overlap between adjacent samples, thereby reducing information loss caused by truncation at operating condition switching points. Furthermore, the 3-channel data within each window can be reassembled into the original hook monitoring sequence in chronological order. ,in, Indicates the sample index; Indicates the first The original hook monitoring sequence corresponding to each sample has a size of [size missing]. The first column is the load channel, the second column is the lifting height channel, and the third column is the horizontal displacement channel. Original hook monitoring sequence. Each row in the table corresponds to one time point, and each column corresponds to one monitoring channel.

[0036] In one embodiment, after training the hook state recognition model, this invention can deploy it in an online monitoring system for lifting equipment to perform corresponding real-time monitoring and early warning for the lifting hook. Corresponding to the above embodiment, the same sampling interval as in the training phase can be obtained. Real-time acquisition of load channel, lifting height channel, and horizontal displacement channel data, and processing them according to the same window length. The original hook monitoring sequence to be identified is constructed. After segmentation, feature coupling, and feature decomposition of the original hook monitoring sequence, a complete dynamic enhancement feature matrix is ​​constructed. Further anomaly identification is performed using the model. Specifically, the input features can be time-aligned, and a time-aligned fusion feature matrix is ​​constructed. This matrix is ​​then used by the trained hook state identification model to identify the state category, outputting the predicted probability of the state category, such as normal operation, continuous overload anomaly, and impact vibration anomaly.

[0037] The system can use the category with the highest predicted probability as the identification result for the current window. An alarm is triggered when the predicted probability of a continuous overload or impact vibration anomaly exceeds a preset alarm threshold. In one implementation, the alarm threshold can be set to 0.7. To reduce false alarms in a single window, an alarm can be triggered only when three consecutive windows determine the same abnormal state. Furthermore, the alarm result, anomaly category, and corresponding time period can be synchronously sent to the equipment monitoring terminal or maintenance platform to prompt operators to take measures such as load reduction, shutdown, reset, or maintenance. It should be noted that maintaining the same data organization method and preprocessing process as the training phase in the online identification process avoids performance degradation caused by training-deployment inconsistencies. Simultaneously, the continuous window determination mechanism further improves alarm stability in engineering applications.

[0038] Furthermore, the steps of the above embodiments will be described in detail: A. Regarding step S102 above, the lifting speed envelope of the lifting height sequence can be determined based on the absolute value of the difference between adjacent sampling points of the lifting height sequence. Based on the fluctuation intensity of the lifting speed envelope within a preset local interval, the segment boundary set of the lifting height sequence is determined. Based on the segment boundary set, the load sequence, lifting height sequence, and horizontal displacement sequence of the real-time monitoring sequence are segmented respectively to form a multi-segment monitoring matrix.

[0039] When the hook is in the stable lifting phase, the lifting height channel changes relatively continuously, making it suitable to use longer segments to extract the stable trend. When the hook is in the starting, braking, speed change, and oscillation enhancement phases, the local changes in the lifting height channel are faster, making it suitable to use shorter segments to preserve transient details. This step uses the local rate of change of the lifting height channel to adaptively adjust the segment length, thus adaptively segmenting the original sequence. The specific steps are as follows: 1) From the original hook monitoring sequence Extract the lift height channel to obtain the lift height sequence. .in, Indicates the first The takeoff height channel sequence of each sample is of length [length missing]. It is used to characterize the change of the vertical position of the hook over time.

[0040] 2) Lifting height sequence Calculate the absolute value of the difference between adjacent sampling points to obtain the height difference sequence; then perform one-dimensional smoothing on the height difference sequence to obtain the liftoff velocity envelope. .in, Indicates the first The liftoff velocity envelope of each sample has a length of [length missing]. It is used to characterize the degree of vertical motion activity near each time point.

[0041] In practice, calculations can be performed first in chronological order. Then, a moving average of length 5 is applied to the obtained sequence, or a one-dimensional Gaussian smoothing with a kernel length of 5 and a standard deviation of 1.0 to 2.0 is performed to reduce the impact of single-point spikes on subsequent segmentation results. Indicates the first In the sample, the lifting height channel is in the first... The values ​​at each sampling point; Indicates the first In the sample, the lifting height channel is in the first... The values ​​at each sampling point. The range of values ​​is arrive .

[0042] 3) Starting from the beginning position of the original hook monitoring sequence, the fluctuation intensity of the lifting speed envelope is statistically analyzed according to the local interval, and the current segment length is calculated accordingly.

[0043] The current segment length can be denoted as: , Indicates the first The number of sampling points contained in each segment; the standard deviation of the lift velocity envelope of the local interval is denoted as... , This represents the degree of fluctuation statistically obtained from the current segment starting point within the basic observation interval. The current segment length can be expressed as:

[0044] in, This represents the minimum segment length, used to prevent rapidly changing areas from being divided into too many fragments; a value of 16 is preferred. This represents the length of the basic observation interval, with a preferred value of 40. This represents the sensitivity coefficient to changes in lifting height, used to adjust the intensity of the influence of lifting speed envelope fluctuations on segment length; a preferred value is 0.8. This indicates the floor function; This indicates a maximum value operation. In one embodiment, as an example, when... , , At that time, if the standard deviation of the lift velocity envelope in the current local interval is... The current segment length is approximately 37; if Therefore, the current segment length is approximately 20. This shows that a longer segment is automatically used during the steady-state phase, while a shorter segment is automatically used during the variable-speed or impact-enhanced phases.

[0045] 4) Proceed sequentially according to the current segment length to generate the segment boundary set: P = p0, p1, ..., p M.in, Indicates the first The set of segment boundaries for each sample; Indicates the starting boundary position, preferably 0; Indicates the position of the end boundary, preferably taking The difference between two adjacent boundary positions corresponds to the length of one segment. In the specific implementation, the current starting point can be initially set to... Then, based on the results obtained in the previous step Calculate the current endpoint ;when Exceed At that time, directly Cut off as Then ordered As the starting point for the next segment, repeat the above operation until the entire original hook monitoring sequence has been traversed.

[0046] 5) Based on the segment boundary set From the original hook monitoring sequence Each segment is extracted sequentially to obtain the segmented monitoring matrix. .in, Indicates the first The sample at the th The segmented monitoring matrix is ​​located on each segment, with a size of [missing information]. This is used to characterize the raw data of the load channel, lifting height channel, and horizontal displacement channel within this local time period.

[0047] Hook condition switching is typically first reflected in the rate of change of the lifting height channel. This step does not mechanically cut the original hook monitoring sequence into fixed-length segments, but rather allows the segment length to be automatically adjusted according to the local changes in the lifting height channel. Based on this, more sufficient trend information can be accumulated during stable periods, and more complete transient details can be preserved during start-stop and swing periods, thus providing a clearer input basis for subsequent trend-disturbance decomposition.

[0048] In one embodiment, Figure 3 A schematic diagram of the lifting speed envelope is shown. This curve is obtained by Gaussian smoothing the absolute value of the difference between the lifting height channels. The activity level of the hook's vertical movement can be observed. Regions with larger speed envelope values ​​correspond to the starting, speed change, braking, or increased swaying stages. Adaptive windowing can be performed on the data based on these speed envelope values. The horizontal axis represents time (s), and the vertical axis represents the lifting speed envelope (m / s).

[0049] B. Regarding step S104, the first segment sequence, the second segment sequence, and the third segment sequence in the multi-segment monitoring matrix can be locally normalized within each segment to obtain a normalized segment monitoring matrix. Based on the absolute value of the first-order difference between adjacent sampling points in the normalized segment monitoring matrix, the difference sequences corresponding to the first segment sequence, the second segment sequence, and the third segment sequence are determined. Based on the preset fusion weights, the difference sequences are weighted and summed to construct the motion intensity coupling features corresponding to the first segment sequence, the second segment sequence, and the third segment sequence.

[0050] 1) Local normalization within segments of the segmented monitoring matrix can eliminate local dimensional differences and operational baseline differences between different segments.

[0051] Among them, the segmented monitoring matrix can be used. For each channel, the mean and standard deviation within the segment are calculated separately. Then, the channel value at each time point is subtracted from the corresponding channel mean and divided by the sum of the corresponding channel standard deviation and the minimum constant to obtain the normalized result.

[0052] The minimum constant can be denoted as: , To prevent the denominator from being 0, the preferred value is [value]. For each segment monitoring matrix Performing local normalization within each segment yields the normalized segmented monitoring matrix. .in, Indicates the first The sample at the th The normalized segmented monitoring matrix on each segment still has a size of [size missing]. In one implementation, if a channel fluctuates very little within the current segment, causing the standard deviation to be lower than a preset threshold, then the standard deviation is directly replaced with 1 or a preset minimum positive number to avoid abnormal amplification of the normalization result.

[0053] 2) Normalized segmented monitoring matrix The three channels are used to calculate the absolute value of the first-order difference between adjacent sampling points, yielding the load difference sequence, the lifting height difference sequence, and the horizontal displacement difference sequence. Combining these three difference sequences using a weighted sum yields the coupled motion intensity sequence. .in, Indicates the first The sample at the th The coupled motion intensity sequence on each segment has a length of [number]. This is used to characterize the local dynamic activity level reflected by the three channels in the vicinity of the current position. In one implementation, the coupled motion intensity sequence can be represented as: ; in, Indicates the load channel in the first The absolute value of the first-order difference at each position Indicates the lifting height channel at the 1st The absolute value of the first-order difference at each position Indicates the horizontal displacement channel in the first... The absolute value of the first-order difference at each position; , and These represent the fusion weights of the three channels, and In a convenient implementation method, one can take... , , This allows the load channel to have a slightly higher weight, while retaining the contributions of the lifting height channel and the horizontal displacement channel to the dynamic state.

[0054] C. Regarding step S106, based on the segmented normalized values ​​of the motion intensity coupling characteristics, weighted moving averages can be applied to the first, second, and third segment sequences of each monitoring matrix to construct a reconstruction trend matrix corresponding to each monitoring matrix. Each monitoring matrix is ​​then subtracted element-wise from the reconstruction trend matrix to determine the trend matrix and residual matrix contained in each monitoring matrix. In one embodiment, the motion intensity coupling characteristics can be mapped to a local smoothing radius based on the magnitude of the motion intensity represented by the segmented normalized values. Using the local smoothing radius, weighted moving averages can be applied to the first, second, and third segment sequences to construct a reconstruction trend matrix corresponding to each monitoring matrix.

[0055] Although the normalized piecewise monitoring matrix has eliminated local dimensional differences, persistent overload anomalies and impact vibration anomalies are still mixed in the same waveform. Persistent overload anomalies mainly exhibit slow trend changes, while impact vibration anomalies mainly exhibit local high-frequency disturbances. This invention no longer employs the complex method of training independent reconstruction networks, but directly utilizes a 3-channel coupled motion intensity adaptive control smoothing scale to extract the reconstruction trend matrix from the normalized piecewise monitoring matrix, and further generates the corresponding kinematic enhancement feature matrix. In one embodiment, this can be represented by the following steps: 1) The greater the coupled motion intensity, the more likely the current position is to be in a state of enhanced impact, acceleration, or oscillation, and the smaller the corresponding local smoothing radius; the smaller the coupled motion intensity, the more stable the current position, and the larger the corresponding local smoothing radius. This embodiment of the invention uses a coupled motion intensity sequence... Perform intra-segment normalization and map it to an adaptive smoothing radius. It can be used to control the neighborhood range used when extracting trends from the current location.

[0056] In practical implementation, the minimum value of the coupled motion intensity sequence within the current segment can be calculated first. and maximum value Then, the normalized coupled motion intensity is obtained by using the minimum-maximum normalization method within the segment; subsequently, the local smoothing radius is obtained by using a linear mapping method, for example:

[0057] in, Indicates the first The sample at the th The first segment, the first Local smoothing radius at each location; This represents the normalized coupled motion intensity, with a value ranging from 0 to 1; This represents the minimum smoothing radius, with a preferred value of 2. This represents the maximum smoothing radius, with a preferred value of 8.

[0058] In one embodiment, as an example, if the normalized coupling motion intensity at a certain location... The local smoothing radius is approximately 7; if The local smoothing radius is approximately 2. Therefore, a larger smoothing neighborhood is used for stable locations, while a smaller smoothing neighborhood is used for rapidly changing locations.

[0059] 2) Using a normalized segmented monitoring matrix Using the input as input, a weighted moving average is applied to the three channels at each position according to the corresponding local smoothing radius to obtain the reconstructed trend matrix. .in, Indicates the first The sample at the th The reconstruction trend matrix on each segment has a size of [missing information]. , is used to characterize the smooth trend component that conforms to the current continuous motion law of the hook.

[0060] In practical implementation, the current position can be used. Centered on, in Within a selected local neighborhood, sampling points closer to the center are assigned greater weight, while those farther away are assigned less weight. For example, a triangular weighting can be used, where closer points to the center receive higher weights. When a neighborhood crosses a segment boundary, only the valid sampling points within the current segment are weighted and averaged, and the valid weights are then renormalized. Based on this, a reconstructed trend matrix is ​​generated. It provides a smoother transition in stable regions and retains the necessary transition shape in impact regions, unlike fixed large-window smoothing which completely smooths out local changes. Furthermore, "effective weight renormalization" refers to renormalizing all calculated weights... Summation yields Then adjust the actual application weight of each point to This ensures that even if the neighborhood is truncated by the segment boundary, the sum of the weights of the remaining points is still 1, thus guaranteeing the numerical stability of the weighted average.

[0061] In one embodiment, as an example, suppose we are currently at the th Each segment needs to have its position calculated. The trend value at that point. The boundary of this segment is... arrive Local smoothing radius Therefore, the range of neighborhood indexes to be considered is... For any point within this neighborhood... Its weight A linearly decaying triangular weight can be used for calculation, for example... .in, and All come from the segmented boundary set , Indicates the first The starting boundary index (sampling point number) of each segment. Indicates the first End boundary index (sampling point number) of each segment; As the local smoothing radius The abbreviation for "current position" indicates the current location. The number of sampling points allowed to expand to the left and right neighborhoods at this point. These values ​​are provided as examples only. Indicates that the index is within the neighborhood. The weighting assigned to the sampling points has the physical meaning of being the distance from the current position. The closer the point, the greater its contribution to trend estimation; This represents the weight of all sampling points within the effective neighborhood. The sum of .

[0062] 3) Normalize the segmented monitoring matrix With the reconstruction trend matrix Subtracting element by element yields the residual matrix. .in, Indicates the first The sample at the th The residual matrix over each segment has a size of . It is used to retain information on local mutations and anomalous perturbations that are not explained by the reconstructed trend matrix.

[0063] In one embodiment, for example, when the hook is under continuous overload abnormality without obvious impact, the reconstruction trend matrix can follow the continuous rise of the load channel well, and the residual matrix is ​​close to 0 in most positions; when the hook experiences abnormal impact vibration, the load channel and horizontal displacement channel will show short-term spikes or rapid reverse swings, and the residual matrix will form obvious spikes at the corresponding positions.

[0064] In summary, in the original hook monitoring sequence, continuous overload anomalies typically manifest as a prolonged upward trend, while impact vibration anomalies typically manifest as local short-term spikes or oscillating envelopes. These two types of anomalies are easily overlapped in the original waveform. Conventional fixed-window standardization or uniform smoothing can easily suppress transient anomalies or misjudge continuous trends as disturbances. However, hook anomalies are often not changes in a single channel, but rather coupled changes in the load channel, lifting height channel, and horizontal displacement channel. This embodiment of the invention first performs non-uniform time unwrapping based on the local change state of the lifting height channel, and then constructs the characteristic coupling of the three channels. The local dynamic activity level reflected by the three channels is directly used to control the trend extraction scale. Continuous overload anomalies are more stably preserved in the reconstructed trend matrix, and impact vibration anomalies are more concentratedly preserved in the residual envelope matrix, significantly reducing the overlap of the two types of anomalies in the original waveform. In one embodiment, Figure 4 This is a schematic diagram showing the reconstruction trend and residuals of the load channel, where... Figure 4 In the diagram, (a) represents the reconstruction trend obtained from the original load signal and adaptive smoothing. Figure 4 In the diagram, (b) represents the residual after subtracting the trend from the original signal. The reconstructed trend follows the rising trend of the sustained overload (approximately 10-18 seconds) while retaining the necessary transition shape during the impact phase (approximately 22-24 seconds); the residual highlights local abrupt changes and impact spikes. The horizontal axis of the coordinate system represents time (s), and the vertical axis represents either the load (kN) or the residual (kN).

[0065] D. Regarding step S108 above, peak enhancement can be performed on the residual matrix to obtain the enhanced residual envelope matrix; the enhanced residual envelope matrix and the reconstruction trend matrix of each monitoring matrix segment can be concatenated to obtain the segmented dynamic enhancement feature matrix; the segmented dynamic enhancement feature matrices corresponding to each monitoring matrix segment in the multi-segment monitoring matrix can be concatenated to construct the kinematic enhancement feature matrix of the real-time monitoring sequence.

[0066] 1) Wherein, the residual matrix can be... The absolute value of each channel is taken, and peak preservation enhancement with distance attenuation is performed in the local neighborhood to obtain the enhanced residual envelope matrix. .

[0067] in, Indicates the first The sample at the th The enhanced residual envelope matrix on each segment has a size of . , is a non-negative matrix used to characterize the intensity of anomalous disturbances near each time location. In a convenient implementation method, the The first channel in the The enhanced residual envelope value at each location can be expressed as:

[0068] in, The enhanced residual envelope matrix is ​​represented in the th... The first channel, the first The values ​​at each position; This represents the radius of the locally enhanced neighborhood, with a preferred value of 4. This represents the distance attenuation coefficient, with a preferred value of 2.0. This represents the neighborhood offset relative to the current position. Represents the natural exponential function; Indicates the offset from arrive Within the integer range, take the maximum value of the expression within the parentheses. Based on this, at the current position... Within the local neighborhood, the peak value of the absolute value of the residual after distance attenuation can be found, thus forming an "envelope" that can contain the local impact.

[0069] In one embodiment, for example, if the residual matrix of the load channel exhibits a peak with an amplitude of 1.2 at the 18th position, then in , In this case, position 18 retains a high response of 1.2, while positions 17 and 19 also obtain relatively high responses after decay. Based on this, the spikes that were originally concentrated at a single point can be expanded into local envelopes, making it easier for subsequent classification networks to identify abnormal persistence intervals.

[0070] 2) Reconstruct the trend matrix With enhanced residual envelope matrix By splicing along the channel direction, a piecewise dynamic enhancement feature matrix is ​​obtained. .in, Indicates the first The sample at the th The piecewise dynamic enhancement feature matrix on each segment has a size of [value]. The first three channels are, in order, the reconstructing trend of the load channel, the reconstructing trend of the lifting height channel, and the reconstructing trend of the horizontal displacement channel; the last three channels are, in order, the enhanced residual envelope of the load channel, the enhanced residual envelope of the lifting height channel, and the enhanced residual envelope of the horizontal displacement channel.

[0071] 3) Arrange all segmented dynamic enhancement feature matrices in chronological order. By concatenating these features sequentially, a complete dynamic enhancement feature matrix is ​​obtained. .in, Indicates the first The complete dynamic enhancement feature matrix of each sample, with size . This serves as the input for the subsequent hook state recognition model.

[0072] E. The complete dynamic enhancement feature matrix has separated trend information from disturbance information, but there is still a significant sequential relationship between the three channels. For example, the horizontal displacement channel envelope may rise first, followed by an enhancement in the load channel envelope; the lifting height channel trend may change first, followed by a rise in the load channel trend. If this sequential relationship is not addressed, the subsequent classification network may easily mistake the real coupled links for noise. This invention first performs cross-channel delay alignment based on the cross-correlation results, then further performs channel weighting and uses envelope-aware gated convolution to complete the final classification.

[0073] Regarding step S108 above, the abnormality identification result of the lifting hook can be determined through the following steps: 1) Based on the normalized cross-correlation value between the preset first channel data and the preset second channel data in the kinematic enhancement feature matrix, the first channel data and the second channel data are time-aligned to construct a time-aligned fusion feature matrix.

[0074] Complete dynamic enhancement feature matrix Although the six channels correspond one-to-one in time, the actual physical response of the hook exhibits a propagation time sequence. For example, an emergency stop will first cause a drastic change in the horizontal displacement channel, and only then will it trigger the impact oscillation in the load channel. If convolution is performed directly, the network will mistakenly identify this waveform misalignment caused by causal timing as different feature patterns, thereby reducing its robustness in recognition.

[0075] To address the aforementioned issues, this invention quantifies the time delay relationship between each auxiliary channel and the reference channel (load channel) by calculating the normalized cross-correlation function. By finding the time delay corresponding to the maximum cross-correlation value and aligning each channel, all channels achieve "causal alignment" in the time dimension. That is, in the feature matrix, the values ​​of each channel at the same time step reflect the response state on the same physical causal chain, thereby simplifying the feature space complexity that the subsequent convolutional network needs to learn.

[0076] a- Perform normalized cross-correlation calculations on other channels and their corresponding reference channels, and search for the optimal delay within a preset delay range.

[0077] The time delay range is denoted as , This represents the maximum search step size, used to limit the search range for time shifting. Preferably, A value of 8 can be used. The purpose of normalized cross-correlation calculation is to compare the consistency of the change shapes of two channels at different time offsets and to find the position offset with the strongest correlation. The calculation method of normalized cross-correlation value is expressed as follows:

[0078] in, Indicates channel With channel In time delay Normalized cross-correlation value under the following conditions; and These represent the two channels to be compared at the [number]th ...]. The values ​​at each position; and These represent the mean values ​​of the two channels respectively; This represents a very small constant, used to prevent the denominator from being 0.

[0079] In one embodiment, as an example, assume that both sequences have a length of 1. When calculating latency When, it means sequence Shift forward 5 sampling points and sequence Alignment. At this point, the effective comparison region is the sequence. The arrive points, and sequences The arrive Points. Summation operation. That is, iterate through all the indexes within this valid comparison area. and They should also be sequences. and The mean within the valid comparison region, not the mean of the entire sequence.

[0080] b - For each pair of "reference channel - channel to be aligned", select the delay with the largest absolute value of normalized cross-correlation as the optimal delay, denoted as . The corresponding normalized cross-correlation absolute value is denoted as... , This indicates the coupling strength between the channel and the reference channel.

[0081] c - Perform time shifting on the channels to be aligned according to the optimal delay, and calculate the channel weighting coefficients based on the coupling strength to obtain the time-aligned weighted channels.

[0082] In practical implementation, when the optimal delay is positive, the channel to be aligned is shifted forward by the corresponding step size; when the optimal delay is negative, the channel to be aligned is shifted backward by the corresponding step size. Boundary points are filled by copying boundary values ​​to avoid the beginning and end information being forcibly set to zero. Furthermore, Softmax normalization can be applied to all coupling strengths within the same channel group to obtain the channel weighting coefficients. This allows channels with higher coupling strength to have a greater weight during fusion.

[0083] In one implementation, the weighting coefficients obtained for the reconstructing trends of the lifting height channel and the horizontal displacement channel can be used to adjust the degree to which they supplement information to the reconstructing trend of the load channel; the weighting coefficients obtained for the enhanced residual envelopes of the lifting height channel and the horizontal displacement channel can be used to adjust the degree to which they supplement information to the enhanced residual envelopes of the load channel.

[0084] d- The reference channel is retained as the main channel, and the time-aligned weighted channel is combined with the main channel to form a time-aligned fusion feature matrix. .in, Indicates the first The time-aligned fused feature matrix of each sample has a size of [size missing]. Its number of channels is consistent with the complete dynamic enhancement feature matrix, but each auxiliary channel has completed time alignment and coupling strength weighting, which is more conducive to subsequent convolution operations to extract the real cross-channel coupling mode.

[0085] It should be noted that the mechanical response of the hook itself has a propagation order. This embodiment of the invention does not simply splice the channels together at the same time, but first estimates "which changes first, which changes later, and how many sampling points are different," and then completes channel alignment and weighting. This allows the subsequent hook state recognition model to more accurately distinguish between "continuous trend coupling corresponding to continuous overload anomalies" and "short-term leader disturbance coupling corresponding to impact vibration anomalies." In one implementation, the absolute value of the residual can be enhanced by a peak-holding operation with distance attenuation within a local neighborhood (radius 4) to obtain the enhanced residual envelope. (Refer to...) Figure 5 The diagram shows the enhanced residual envelope. The enhanced envelope expands the spikes at a single point into a local envelope, which can more clearly identify the duration of abnormal disturbances and facilitate classification network identification. The horizontal axis of the coordinate system represents time (s), and the vertical axis represents the amplitude (kN, consistent with the load dimension).

[0086] In one implementation, the complete dynamic enhancement feature matrix can be... The features are divided into two groups based on the channels. The first three channels constitute the reconstruction trend channel group, namely the load channel reconstruction trend, the lifting height channel reconstruction trend, and the horizontal displacement channel reconstruction trend; the latter three channels constitute the enhanced residual envelope channel group, namely the load channel enhanced residual envelope, the lifting height channel enhanced residual envelope, and the horizontal displacement channel enhanced residual envelope. In the reconstruction trend channel group, the load channel reconstruction trend can be used as the reference channel; in the enhanced residual envelope channel group, the load channel enhanced residual envelope is used as the reference channel.

[0087] Correspondingly, the complete dynamic enhancement feature matrix It was split into two groups: a> Reconstruction trend channel group (first 3 channels): Channel 1: Load channel reconstruction trend (reference channel); Channel 2: Lifting height channel reconstruction trend (channel to be aligned); Channel 3: Horizontal displacement channel reconstruction trend (channel to be aligned).

[0088] b>Enhanced residual envelope channel group (last 3 channels): Channel 4: Load channel enhanced residual envelope (reference channel); Channel 5: Lifting height channel enhanced residual envelope (channel to be aligned); Channel 6: Horizontal displacement channel enhanced residual envelope (channel to be aligned).

[0089] Within each group, taking the sequence corresponding to the load channel as the reference channel (i.e., assuming the load response is the final observed object in the physical causal chain), the other two channels within the group are taken as channels to be aligned, and their optimal time delays relative to the reference channel are calculated. and coupling strength The specific correspondences are listed below: Table 1: Correspondence between reference channels and channels to be aligned in cross-channel timing alignment

[0090] Referring to Table 1, the correspondence between the channels to be aligned, with the load channel as the reference, is shown in the reconstruction trend channel group and the enhanced residual envelope channel group. For example, in the reconstruction trend channel group, the "reference channel - channel to be aligned" pair includes: ① load reconstruction trend - lifting reconstruction trend; ② load reconstruction trend - horizontal reconstruction trend. In the enhanced residual envelope group, the "reference channel - channel to be aligned" pair includes: ① load envelope - lifting envelope; ② load envelope - horizontal envelope. Based on this, channel data that were originally misaligned due to physical propagation delay can achieve "causal alignment" on the time axis, that is, the channel values ​​at the same time index reflect the synchronous response state triggered by the same physical event.

[0091] In one embodiment, for example, if the normalized cross-correlation absolute value is maximized when the horizontal displacement channel enhanced residual envelope is sampled two points ahead of the load channel enhanced residual envelope, and the maximum value is 0.73, then the optimal time delay is denoted as... The coupling strength was denoted as 0.73. This result indicates that, in the current sample, the anomalous disturbance in the horizontal displacement channel occurs earlier than the anomalous disturbance in the load channel.

[0092] 2) By performing convolution operations on the time-aligned fusion feature matrix using a pre-built hook state recognition model, the continuous overload abnormal signal and short-term disturbance features in the time-aligned fusion feature matrix are determined.

[0093] Specifically, the reconstructed trend channel group data contained in the time-aligned fusion feature matrix can be input into the trend convolution branch to extract the continuous overload anomaly signal from the time-aligned fusion feature matrix; the reconstructed trend channel group data and the enhanced residual envelope channel group data contained in the time-aligned fusion feature matrix can be input into the transient convolution branch to extract the short-term perturbation features from the time-aligned fusion feature matrix. The reconstructed trend channel group data is the data corresponding to the reconstructed trend matrix of each monitoring matrix segment, and the enhanced residual envelope channel group data is the data corresponding to the residual matrix of each monitoring matrix segment.

[0094] a- Time-aligned fusion of feature matrices Input the trend convolution branch to extract persistent trend features.

[0095] In its implementation, the trend convolution branch takes the first three reconstructed trend channels of the time-aligned fused feature matrix as input, performs a one-dimensional convolution operation, with a kernel length of up to 9 and an output channel count of up to 32, followed by a non-linear activation function. This branch is primarily used to extract the slow rise, persistent bias, and trend continuation patterns corresponding to sustained overload anomalies. Since the enhanced residual envelope channels mainly consist of high-frequency noise and impulses and do not contain slow trend information, they are not used as input to this branch to prevent the introduction of irrelevant feature interference.

[0096] b- Time-aligned fusion of feature matrices Input transient convolutional branches, extract short-term perturbation features, and use gated sequences. Adjust the position of the transient convolution branch output time-by-time.

[0097] In practical implementation, the transient convolution branch can use a one-dimensional convolution kernel of length 3 or 5 to convolve all 6 channels to extract local edge, short-term impact, and fast rebound features; after the convolution output, the gated sequence is... It is broadcast along the channel direction and multiplied element-wise with the convolution output, thereby suppressing the interference of the stable background on the transient convolution branch.

[0098] In one embodiment, for example, when the transient intensity near a certain position is low and the gate value is close to 0.15, even if there is a certain response in the transient convolution branch, the response will be significantly suppressed; when the transient intensity near a certain position is high and the gate value is increased to above 0.8, the abnormal response of the transient convolution branch will be effectively preserved.

[0099] 3) Based on the continuous overload abnormal signal and short-term disturbance characteristics, the probability of abnormal category is predicted on the kinematic enhancement feature matrix to determine the abnormal identification result corresponding to the lifting hook.

[0100] Specifically, the outputs of the trend convolutional branch and the transient convolutional branch can be concatenated along the channel direction to obtain a fused convolutional feature sequence; then, a one-dimensional convolution of length 5 is performed for cross-branch fusion to output high-level temporal features. Furthermore, to ensure that the output time length is consistent with the input time length, symmetrical padding can be used in the convolutional layer; to improve training stability, batch normalization or layer normalization can be applied after the convolutional layer.

[0101] Furthermore, global average pooling can be performed on the high-level temporal features along the time axis to obtain sample-level feature vectors. .in, Indicates the first The sample-level feature vector, whose dimension is determined by the number of output channels of the last convolutional layer, is used to characterize the overall state of the entire hook operation window. Further, the sample-level feature vector... The input is a fully connected classification layer, which outputs predicted probabilities for three states through a softmax transform. These three states correspond to normal operation, continuous overload anomaly, and shock / vibration anomaly, respectively. The class with the highest probability is used as the identification result for the current sample.

[0102] One approach is to train the hook state recognition model using class-weighted cross-entropy loss to mitigate the class imbalance problem caused by the fact that normal operation samples are usually far more numerous than abnormal samples from continuous overload and impact vibration. In specific implementation, the first... The loss weights for each class of samples are denoted as... You can press: Calculate, where, Indicates the first Number of class samples This represents the number of samples in the category with the largest sample size. In one embodiment, for example, if the number of samples for normal operation is 6000, the number of samples for continuous overload anomaly is 1500, and the number of samples for impact vibration anomaly is 1000, then the weights for the three categories can be 1, 4, and 6, respectively. This allows continuous overload anomaly and impact vibration anomaly to receive higher gradient weights during the training phase.

[0103] In this approach, an optimizer can be used to update the parameters of the hook state recognition model, and the model parameters with the best performance on the validation set can be selected as the final output model. One implementation can employ an adaptive moment estimation optimizer; the initial learning rate can be set to... The batch size can be 32; the number of training cycles can be 80 to 150. Furthermore, during the model validation phase, the F1 scores for normal operation, continuous overload anomaly, and impact vibration anomaly can be calculated separately. Then, the arithmetic mean of the three types of F1 scores is calculated to obtain the macro-average F1 score, and the model parameters corresponding to the highest macro-average F1 score are saved.

[0104] To ensure a uniform input format for the model's input samples (such as training samples), this invention also performs windowed segmentation of the continuous monitoring stream to form labeled original hook monitoring sequences. Status labels can be assigned to each original hook monitoring sequence by combining rated load information, equipment controller alarm records, mechanical maintenance records, and manual verification results. .in, Indicates the first The status category of each sample. In one implementation, the status label may include three categories, such as normal operation, continuous overload anomaly, and shock vibration anomaly.

[0105] Among them, normal operation corresponds to samples where the load channel does not continuously exceed the rated load threshold and there is no obvious impact waveform; continuous overload abnormality corresponds to samples where the load channel continuously exceeds the rated load threshold within the window, such as continuously exceeding 1.05 times the rated load for more than 2 seconds; impact vibration abnormality corresponds to samples where short-term spikes, rapid rebounds or high-frequency swing envelopes appear in the load channel or horizontal displacement channel, and are corresponding to records of sudden stops, collisions, strong swings or braking impacts.

[0106] Furthermore, all samples can be divided into training, validation, and test sets for subsequent training and evaluation of the hook state recognition model. In one implementation, the ratio of the number of samples in the training, validation, and test sets can be 7:2:1.

[0107] It should be noted that shock vibration anomalies usually only occur in a few short time intervals, rather than persisting throughout the entire time period. This step directly transforms the local intensity of the enhanced residual envelope channel into a gating sequence to adjust the effective response of the transient convolution branch, rather than using the same convolution preservation strategy for all time positions. Based on this, the redundant background in the stable region is actively suppressed, and the local response in the abnormally active region is actively enhanced, thereby enabling the hook condition recognition model to form a clearer discrimination boundary between normal operation, continuous overload anomaly, and shock vibration anomaly.

[0108] Furthermore, in this embodiment of the invention, the transient intensity sequence corresponding to the temporally aligned fused feature matrix is ​​determined based on the enhanced residual envelope channel data contained in the temporally aligned fused feature matrix; the transient intensity sequence is mapped to a gating sequence, and the convolution output retention ratio corresponding to the convolution operation is determined based on the gating sequence.

[0109] In the aforementioned time-aligned fusion feature matrix, the reconstructed trend channel is more suitable for describing sustained overload anomalies, while the enhanced residual envelope channel is more suitable for describing shock vibration anomalies. If convolution calculations of the same intensity are used at all time points, the stationary region will consume a large amount of computation and dilute the local response of the anomaly. This embodiment of the invention also utilizes the local intensity of the enhanced residual envelope channel to generate a gated sequence, allowing the convolutional network to maintain a high response in the anomaly active region and automatically suppress redundant background in the stationary region. The specific steps are as follows: 1) Fuse feature matrices based on time alignment The last three enhanced residual envelope channels are used to calculate the transient intensity sequence. .in, Indicates the first The transient intensity sequence of samples, with a length of It is used to characterize the level of activity of anomalous disturbances near each time location.

[0110] In practical implementation, the average amplitude of the three enhanced residual envelope channels can be calculated within a preset local window, centered on the current position. Then, the average of the results from the three channels is taken to obtain the transient intensity at the current position. The local window radius can be set to 2, corresponding to a window length of 5.

[0111] In one embodiment, for example, when all three enhanced residual envelope channels within a certain local window are close to 0, the transient intensity will be close to 0; when the enhanced residual envelope of the load channel and the enhanced residual envelope of the horizontal displacement channel both show significant bulges within a certain local window, the transient intensity will increase significantly, for example, reaching 0.3, 0.5 or higher.

[0112] 2) Transient intensity sequence Mapped to gated sequence .

[0113] in, Indicates the first The gating sequence of each sample has a length of [number]. The value ranges from 0 to 1 and is used to adjust the convolution output retention ratio at each time point.

[0114] In one implementation, the gating sequence can be obtained using the Sigmoid function. When local anomalies are weak, the gating value is low; when local anomalies are strong, the gating value is high, thus ensuring a higher retention rate of abnormally active regions in subsequent convolutions. The calculation method is expressed as follows:

[0115] in, Indicates the first The gating value for each position; This represents the gating gain coefficient, used to control the sensitivity of transient intensity to gating changes; a value of 6 is preferred. This indicates the gating bias, with a preferred value of -2.

[0116] In summary, the hook state recognition model of this invention uses the aforementioned time-aligned fusion feature matrix. (size Taking the input as input, classification is performed through a two-branch heterogeneous receptive field convolutional network and an envelope-aware gating mechanism. Its overall architecture and data flow are as follows: a> Input layer. Input data: dimension is of The first three columns are the reconstructed trend channels after time alignment, and the last three columns are the enhanced residual envelope channels after time alignment.

[0117] b> Dual-branch feature extraction layer. Referring to Table 2 below, the input channels, core operations, output dimensions, and design intent of the trend convolutional branch and the transient convolutional branch are shown.

[0118] Table 2: Configuration Table of Dual-Branch Feature Extraction Layer for Hook Status Recognition Model

[0119] c> Envelope-aware gating module (operates on transient convolution branches).

[0120] Gated sequence generation: using The transient intensity is calculated using the three enhanced residual envelope channels in the middle and rear. The sequence is mapped to a gated sequence by the Sigmoid function. .

[0121] Gated weighting: Multiplying the convolution output of the transient convolution branch by time step by time. (Broadcast to all channels) to achieve adaptive adjustment of "high retention of abnormally active areas and low suppression of stable background areas".

[0122] d> Cross-branch fusion and classification layers. Referring to Table 3 below, the operational descriptions and output dimensions of channel concatenation, cross-branch fusion convolution, global average pooling, and fully connected classification layers are shown.

[0123] Table 3: Cross-branch fusion and classification layer configuration table for hook status recognition model

[0124] In summary, the real-time monitoring and early warning method for lifting hooks based on multi-source data fusion provided by this invention is innovative in the following aspects compared to the prior art: 1. An adaptive segmented normalization method based on the rate of change of lifting height is proposed. The segment length is dynamically adjusted by utilizing the intensity of height difference fluctuations, so that the window of the stable segment is long and the window of the impact segment is short, thus avoiding the loss of transient details or trend truncation caused by fixed windows.

[0125] 2. Construct a trend residual decomposition mechanism driven by coupled motion intensity. By fusing three-channel differential calculations to calculate local dynamic activity and mapping it to an adaptive smoothing radius, a large radius is used in the stable region to preserve the trend, and a small radius is used in the impact region to preserve abrupt changes. The single signal is separated into a reconstructed trend matrix and an enhanced residual envelope matrix, and the two types of anomalies, continuous overload and impact vibration, are explicitly decoupled.

[0126] 3. A cross-channel causal alignment strategy based on cross-correlation delay estimation is adopted. By calculating the optimal delay and coupling strength between the lifting height, horizontal displacement and load channel, the misaligned waveform is time-shifted and weighted to make each channel on the same time index reflect the synchronous response of the same physical causal chain.

[0127] 4. An envelope-aware gated dual-branch convolutional network is adopted. The gate sequence generated by enhancing the residual envelope is used to adjust the transient convolutional branches position by position. High response is retained in abnormally active areas and background interference is suppressed in stable areas. At the same time, class-weighted loss is used to alleviate the sample imbalance problem.

[0128] Based on the above embodiments, this invention also provides a real-time monitoring and early warning device for lifting hooks based on multi-source data fusion, referring to... Figure 6The device includes: a data acquisition module 10, used to acquire real-time monitoring sequences of the lifting hook of the lifting equipment during the lifting process; wherein the real-time monitoring sequences include load sequences, lifting height sequences, and horizontal displacement sequences; a data segmentation module 20, used to divide the real-time monitoring sequences into multiple monitoring matrices based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in a preset local interval; each monitoring matrix in the multiple monitoring matrices contains a first segment sequence corresponding to the load sequence, a second segment sequence corresponding to the lifting height sequence, and a third segment sequence corresponding to the horizontal displacement sequence; and a feature coupling module 30, used to perform data coupling based on the first segment sequence, the second segment sequence, and the third segment sequence. The sampling point difference sequences corresponding to the third segment sequence are used to perform feature coupling on each monitoring matrix segment to construct the corresponding motion intensity coupling features; the data processing module 40 is used to determine the trend matrix and residual matrix contained in the lifting hook during the lifting process based on the difference between the motion intensity coupling features and the motion trend of each monitoring matrix segment; the execution module 50 is used to construct the kinematic enhancement feature matrix corresponding to the real-time monitoring sequence based on the trend matrix and residual matrix, and input the kinematic enhancement feature matrix into the pre-constructed hook state recognition model, and identify the abnormal signals in the kinematic enhancement feature matrix through the hook state recognition model to determine the abnormal identification result of the lifting hook.

[0129] The real-time monitoring and early warning device for lifting hooks based on multi-source data fusion provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0130] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figure 1 The steps of the method are shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figure 1 The steps of the method are shown. This invention also provides a schematic diagram of the structure of an electronic device, as shown. Figure 7 The diagram shows the structure of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101. The processor 101 executes the computer-executable instructions to implement the above-mentioned... Figure 1 The method shown.

[0131] exist Figure 7In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102. The memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced eXtensible Interface). Bus 102 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7The diagram uses only a single bidirectional arrow, but this does not imply a single bus or a single type of bus. Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor 101 reads information from the memory and, in conjunction with its hardware, completes the aforementioned tasks. Figure 1 The method shown.

[0132] The computer program product of the real-time monitoring and early warning method and device for crane hooks based on multi-source data fusion provided in this invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, which will not be repeated here. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage media include: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks, and other media that can store program code.

[0133] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of this invention, used to illustrate the technical solutions of this invention, and not to limit it. The scope of protection of this invention is not limited thereto. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in this invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A real-time monitoring and early warning method for lifting hooks based on multi-source data fusion, characterized in that, The method includes: The real-time monitoring sequence of the lifting hook of the lifting equipment during the lifting process is obtained; wherein, the real-time monitoring sequence includes the load sequence, the lifting height sequence, and the horizontal displacement sequence; Based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in the preset local interval in the real-time monitoring sequence, the real-time monitoring sequence is divided into multiple monitoring matrices; each monitoring matrix in the multiple monitoring matrices contains a first segment sequence corresponding to the load sequence, a second segment sequence corresponding to the lifting height sequence, and a third segment sequence corresponding to the horizontal displacement sequence; Based on the sampling point difference sequences corresponding to the first segmented sequence, the second segmented sequence, and the third segmented sequence, feature coupling is performed on each segment of the monitoring matrix to construct the corresponding motion intensity coupling feature; Based on the difference between the motion intensity coupling characteristics and the motion trend of each monitoring matrix segment, the trend matrix and residual matrix contained in the lifting hook during the lifting process are determined. Based on the trend matrix and residual matrix, a kinematic enhancement feature matrix corresponding to the real-time monitoring sequence is constructed, and the kinematic enhancement feature matrix is ​​input into a pre-constructed hook state recognition model. The hook state recognition model is used to identify abnormal signals in the kinematic enhancement feature matrix to determine the abnormal identification result of the lifting hook. The step of dividing the real-time monitoring sequence into multiple monitoring matrices based on the fluctuation intensity of the lift velocity envelope within a preset local interval in the lift height sequence includes: Based on the absolute value of the difference between adjacent sampling points of the lifting height sequence, the lifting speed envelope of the lifting height sequence is determined; based on the fluctuation intensity of the lifting speed envelope in a preset local interval, the segment boundary set of the lifting height sequence is determined; based on the segment boundary set, the load sequence, lifting height sequence and horizontal displacement sequence of the real-time monitoring sequence are segmented respectively to form a multi-segment monitoring matrix; The step of determining the trend matrix and residual matrix contained in the lifting hook during the lifting process based on the difference between the motion intensity coupling characteristics and the motion trend of each monitoring matrix segment includes: Based on the segmented normalized values ​​of the motion intensity coupling characteristics, a weighted moving average is applied to the first segment sequence, the second segment sequence, and the third segment sequence of each monitoring matrix to construct a reconstructed trend matrix corresponding to each monitoring matrix. The reconstructed trend matrix is ​​then subtracted element-wise from each monitoring matrix to determine the trend matrix and residual matrix contained in each monitoring matrix.

2. The method according to claim 1, characterized in that, The steps of constructing corresponding motion intensity coupling features by performing feature coupling on each monitoring matrix segment based on the sampling point difference sequences corresponding to the first segment sequence, the second segment sequence, and the third segment sequence include: The first segment sequence, the second segment sequence, and the third segment sequence in the multi-segment monitoring matrix are respectively subjected to local normalization within each segment to obtain a normalized segment monitoring matrix; Based on the absolute value of the first-order difference between adjacent sampling points in the normalized segmented monitoring matrix, the difference sequences corresponding to the first segmented sequence, the second segmented sequence, and the third segmented sequence are determined respectively. The differential sequences are weighted and summed based on preset fusion weights to construct motion intensity coupling features corresponding to the first segmented sequence, the second segmented sequence, and the third segmented sequence.

3. The method according to claim 1, characterized in that, The steps of constructing the kinematic enhancement feature matrix corresponding to the real-time monitoring sequence based on the trend matrix and residual matrix include: Peak enhancement is performed on the residual matrix to obtain the enhanced residual envelope matrix; The enhanced residual envelope matrix and the reconstruction trend matrix of each monitoring segment are concatenated to obtain the piecewise dynamic enhancement feature matrix; The segmented dynamic enhancement feature matrices corresponding to each segment of the multi-segment monitoring matrix are spliced ​​together to construct the kinematic enhancement feature matrix of the real-time monitoring sequence.

4. The method according to claim 1, characterized in that, The steps of inputting the kinematic enhancement feature matrix into a pre-constructed hook state recognition model, identifying abnormal signals in the kinematic enhancement feature matrix through the hook state recognition model, and determining the abnormal identification result of the lifting hook include: Based on the normalized cross-correlation value between the preset first channel data and the preset second channel data in the kinematic enhancement feature matrix, the first channel data and the second channel data are time-aligned to construct a time-aligned fusion feature matrix; By performing convolution operations on the time-aligned fusion feature matrix using a pre-built hook state recognition model, the continuous overload abnormal signal and short-term disturbance features in the time-aligned fusion feature matrix are determined. Based on the continuous overload anomaly signal and the short-term disturbance features, the probability of anomaly category is predicted for the kinematic enhancement feature matrix to determine the anomaly identification result corresponding to the lifting hook.

5. The method according to claim 4, characterized in that, The steps of determining the continuous overload anomaly signal and short-term disturbance features in the time-aligned fusion feature matrix by performing convolution operations on the time-aligned fusion feature matrix using a pre-built hook state recognition model include: The reconstructed trend channel group data contained in the time-aligned fusion feature matrix is ​​input into the trend convolution branch to extract the continuous overload abnormal signal in the time-aligned fusion feature matrix. The reconstruction trend channel group data and the enhanced residual envelope channel group data contained in the time-aligned fusion feature matrix are input into the transient convolution branch to extract the short-term perturbation features in the time-aligned fusion feature matrix; wherein, the reconstruction trend channel group data is the data corresponding to the reconstruction trend matrix of each monitoring matrix segment, and the enhanced residual envelope channel group data is the data corresponding to the residual matrix of each monitoring matrix segment.

6. The method according to claim 4, characterized in that, The method further includes: Based on the enhanced residual envelope channel data contained in the time-aligned fusion feature matrix, the transient intensity sequence corresponding to the time-aligned fusion feature matrix is ​​determined; The transient intensity sequence is mapped to a gating sequence, and the convolution output retention ratio corresponding to the convolution operation is determined based on the gating sequence.

7. The method according to claim 1, characterized in that, Based on the segmented normalized values ​​of the motion intensity coupling characteristics, the step of constructing the reconstructed trend matrix corresponding to each segment of the monitoring matrix by performing weighted moving averages on the first segment sequence, the second segment sequence, and the third segment sequence of each segment of the monitoring matrix includes: Based on the motion intensity magnitude represented by the normalized value within the segment, the motion intensity coupling feature is mapped to a local smoothing radius; Using the local smoothing radius, a weighted moving average is applied to the first segmented sequence, the second segmented sequence, and the third segmented sequence to construct a reconstructed trend matrix corresponding to each segment monitoring matrix.

8. A real-time monitoring and early warning device for lifting hooks based on multi-source data fusion, characterized in that, The device includes: The data acquisition module is used to acquire the real-time monitoring sequence of the lifting hook of the lifting equipment during the lifting process; wherein, the real-time monitoring sequence includes the load sequence, the lifting height sequence, and the horizontal displacement sequence; The data segmentation module is used to divide the real-time monitoring sequence into multiple monitoring matrices based on the fluctuation intensity of the lifting speed envelope of the lifting height sequence in a preset local interval. Each monitoring matrix in the multiple monitoring matrices contains a first segment sequence corresponding to the load sequence, a second segment sequence corresponding to the lifting height sequence, and a third segment sequence corresponding to the horizontal displacement sequence. The feature coupling module is used to perform feature coupling on each monitoring matrix segment based on the sampling point difference sequences corresponding to the first segment sequence, the second segment sequence, and the third segment sequence, respectively, to construct corresponding motion intensity coupling features; The data processing module is used to determine the trend matrix and residual matrix contained in the lifting hook during the lifting process based on the difference between the motion intensity coupling characteristics and the motion trend of each monitoring matrix segment. The execution module is used to construct the kinematic enhancement feature matrix corresponding to the real-time monitoring sequence based on the trend matrix and the residual matrix, and input the kinematic enhancement feature matrix into the pre-constructed hook state recognition model. The hook state recognition model is used to identify abnormal signals in the kinematic enhancement feature matrix and determine the abnormal identification result of the lifting hook. The data segmentation module is further configured to: Based on the absolute value of the difference between adjacent sampling points of the lifting height sequence, the lifting speed envelope of the lifting height sequence is determined; based on the fluctuation intensity of the lifting speed envelope in a preset local interval, the segment boundary set of the lifting height sequence is determined; based on the segment boundary set, the load sequence, lifting height sequence and horizontal displacement sequence of the real-time monitoring sequence are segmented respectively to form a multi-segment monitoring matrix; The data processing module is further configured to: Based on the segmented normalized values ​​of the motion intensity coupling characteristics, a weighted moving average is applied to the first segment sequence, the second segment sequence, and the third segment sequence of each monitoring matrix to construct a reconstructed trend matrix corresponding to each monitoring matrix. The reconstructed trend matrix is ​​then subtracted element-wise from each monitoring matrix to determine the trend matrix and residual matrix contained in each monitoring matrix.

Images