Multi-element data fusion and monitoring state evaluation method and device for intelligent lifting hooks

Abstract

The application provides a multi-element data fusion and monitoring state evaluation method and device for an intelligent lifting hook, relates to the technical field of artificial intelligence, and can quantize impact strength and contribution difference by calculating the instantaneous frequency mutation index of each channel of a plurality of sensors, so as to avoid the averaging of key information; short-time correlation strength is further used for local dynamic correlation, so as to capture the cross-channel coupling relationship of the impact and construct time sequence characteristics reflecting the overall structure response. Furthermore, an adaptive mask is generated according to the channel fluctuation and coupling degree in the time sequence characteristics, the instantaneous component closely related to the impact is selectively enhanced, the redundant interference can be inhibited, and the feature signal-to-noise ratio is improved. Finally, a monotonic constraint discrimination is adopted, so as to output a state prediction result conforming to the irreversible law of damage, and the result can truly reflect the safety trend of the lifting hook under the impact cumulative effect in the lifting process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for multi-data fusion and monitoring status assessment of smart hooks. Background Technology

[0002] As a core load-bearing component of lifting machinery, the operational status of intelligent hooks directly affects the safety and reliability of lifting operations. Under unstable conditions such as lifting off the ground, load swinging, emergency braking, and sudden load changes, the hook will be subjected to high-intensity impact loads within a short period. If the abnormal states corresponding to these impacts cannot be identified in a timely and accurate manner, it can easily lead to cumulative structural damage or even fracture failure. Therefore, real-time status monitoring and assessment of intelligent hooks has become a critical requirement in the field of engineering safety.

[0003] Existing monitoring methods for impact detection often rely on fixed amplitude thresholds or frequency domain energy analysis. However, the essence of impact lies in the simultaneous occurrence of rapid changes in frequency components and instantaneous jumps in amplitude. A single-dimensional judgment can easily lead to missed detections or false alarms. Furthermore, slow drifts in slowly varying channels such as ambient temperature can also be misinterpreted as impacts due to amplitude fluctuations. In addition, regarding cross-channel correlation, existing methods typically calculate the correlation matrix over the entire time window. However, the strong multi-sensor coupling relationships triggered by impacts exist only in extremely short local moments. Global statistics significantly dilute this crucial information, making it difficult to effectively extract the linkage characteristics.

[0004] Furthermore, since hoisting operations involve multiple stages such as lifting, slewing, and braking, the sensitive channels differ at each stage. Existing methods rely on static weight allocation, which suppresses the contribution of key sensors while irrelevant channels dominate. In summary, existing technologies lack the accuracy for identifying impact loads and assessing the condition of intelligent hooks under non-stationary conditions, making it difficult to meet engineering safety requirements. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for multi-data fusion and monitoring status assessment of smart hooks, which can achieve accurate identification of hook impact events under non-steady working conditions.

[0006] In a first aspect, embodiments of the present invention provide a method for multi-data fusion and monitoring status assessment of a smart hook. The method includes: acquiring multi-sensor monitoring data of the smart hook during the lifting process; determining the impact weight of the corresponding monitoring channel data based on the instantaneous frequency mutation index of each monitoring channel data in the multi-sensor monitoring data; dynamically associating the impact of each monitoring channel data within a time window according to the short-time correlation strength of the impact weight in the local time neighborhood, thereby constructing a time-series feature containing the original signal and cross-channel coupling relationships; extracting time-frequency features and channel features from the time-series feature; generating a mask based on the fluctuation degree of the channel features and the degree of cross-channel coupling; and adaptively enhancing the time-frequency features using the mask to obtain an enhanced time-frequency feature representation; performing monotonic constraint discrimination on the enhanced time-frequency feature representation; and determining the state prediction result of the smart hook during the lifting process based on the corresponding discrimination result.

[0007] In conjunction with the first aspect, the present invention also provides a first implementation of the first aspect, wherein the step of determining the impact weight of the corresponding monitoring channel data based on the instantaneous frequency mutation index of each monitoring channel data in the multi-sensor monitoring data includes: determining the instantaneous frequency ridge value of the corresponding monitoring channel data based on the phase change of the dominant frequency band of each monitoring channel data in the multi-sensor monitoring data; determining the instantaneous frequency mutation index of the monitoring channel data based on the discrete scale of the instantaneous frequency ridge value, and the frequency jump value and amplitude jump value in adjacent time steps; determining the time step mutation screening threshold based on the instantaneous frequency mutation index; and determining the impact weight of the corresponding monitoring channel data based on the impact situation characterized by the instantaneous frequency mutation index and the time step mutation screening threshold.

[0008] In conjunction with the first aspect, the present invention also provides a second implementation of the first aspect, wherein the above method further includes: performing robust standardization on the corresponding monitoring channel data based on the channel median and channel interquartile range of each monitoring channel data in the multi-sensor monitoring data; and performing multi-band analytical filtering on the standardized monitoring channel data to determine the dominant frequency band in the monitoring channel data.

[0009] In conjunction with the first aspect, this embodiment of the invention also provides a third implementation of the first aspect, wherein the step of performing impact dynamic correlation on each monitoring channel data within a time window based on the short-time correlation strength of the impact weight of each monitoring channel data in the local time neighborhood to construct a time-series feature containing the original signal and cross-channel coupling relationship includes: determining the short-time correlation strength between any two monitoring channels in the local time neighborhood based on the impact weight; determining the impact coupling result of the corresponding local time neighborhood based on the short-time correlation strength; traversing all time steps to obtain the impact dynamic correlation tensor of the sensor monitoring data at different time steps; and combining the impact dynamic correlation tensor with the multi-sensor monitoring data to obtain the time-series feature.

[0010] In conjunction with the first aspect, this invention also provides a fourth implementation of the first aspect, wherein the step of determining the short-time correlation strength between any two monitoring channels based on the impact weights includes: determining the joint weight of the corresponding local time neighborhood based on the impact weights of multiple monitoring channel data in the same local time neighborhood; determining the weighted synchronous change based on the joint weights and the local mean-reduced sampled values ​​of the corresponding monitoring channel data; and normalizing the weighted synchronous change to determine the short-time correlation strength.

[0011] In conjunction with the first aspect, this embodiment of the invention also provides a fifth implementation of the first aspect, wherein the step of performing monotonic constraint discrimination on the enhanced time-frequency feature representation and determining the state prediction result of the intelligent hook during the hoisting process based on the corresponding discrimination result includes: performing weighted pooling compression on the enhanced time-frequency feature representation based on the impact weight to obtain a compressed feature sequence; performing time-series dependency modeling on the compressed feature sequence to obtain a hidden state sequence for the compressed time step; restoring the hidden state sequence to the original time resolution to obtain the state probability distribution for each original time step; applying monotonic constraints to the state probability distribution to prohibit state back-jumps in the state probability distribution, thereby obtaining the state prediction result of the intelligent hook at each time step during the hoisting process.

[0012] In conjunction with the first aspect, this embodiment of the invention also provides a sixth implementation of the first aspect, wherein the step of performing weighted pooling compression on the enhanced time-frequency feature representation based on the impact weight to obtain a compressed feature sequence includes: determining the impact activation intensity at each time step according to the overall intensity of the impact weight at each time step; and using the impact activation intensity as the aggregation weight, performing weighted aggregation on the enhanced time-frequency feature representation in the time dimension to determine the compressed feature sequence.

[0013] Secondly, embodiments of the present invention also provide a multi-data fusion and monitoring status assessment device for intelligent hooks. The device includes: a data processing module for acquiring multi-sensor monitoring data of the intelligent hook during the lifting process, and determining the impact weight of the corresponding monitoring channel data based on the instantaneous frequency mutation index of each monitoring channel data in the multi-sensor monitoring data; an association module for performing dynamic impact association on each monitoring channel data within a time window based on the short-time correlation strength of the impact weight of each monitoring channel data in the local time neighborhood, to construct a time-series feature containing the original signal and cross-channel coupling relationships; an enhancement module for extracting time-frequency features and channel features from the time-series feature, generating a mask based on the fluctuation degree of the channel features and the degree of cross-channel coupling, and using the mask to adaptively enhance the time-frequency features to obtain an enhanced time-frequency feature representation; and an execution module for performing monotonic constraint discrimination on the enhanced time-frequency feature representation, and determining the state prediction result of the intelligent hook during the lifting process based on the corresponding discrimination result.

[0014] Thirdly, embodiments of the present invention also provide an electronic device, which includes a processor and a memory. The memory stores machine-executable instructions that can be executed by the processor. The processor executes the machine-executable instructions to implement the multi-data fusion and monitoring status assessment method for smart hooks according to any of the above embodiments.

[0015] Fourthly, embodiments of the present invention also provide a machine-readable storage medium storing machine-executable instructions. When the machine-executable instructions are invoked and executed by a processor, the machine-executable instructions cause the processor to implement the multi-data fusion and monitoring status assessment method for smart hooks according to any of the above embodiments.

[0016] The embodiments of the present invention bring the following beneficial effects: The embodiments of the present invention provide a method and device for multi-data fusion and monitoring status assessment of smart hooks. By determining the instantaneous frequency mutation index of the acquired multi-sensor monitoring data in each monitoring channel, the severity of the impact event in each channel and its influence weight on the hook state can be known, thereby quantifying the contribution differences of different sensors in the lifting process and avoiding the submergence of key impact information due to signal averaging; furthermore, based on the short-time correlation intensity of each channel, dynamic impact correlation is performed, which can capture the propagation and coupling relationship of impact events across channels in the local time neighborhood, constructing a time-series feature reflecting the dynamic response of the overall hook structure, effectively solving the problem that traditional independent channel analysis cannot identify collaborative anomalies. The correlated data includes the original signal and cross-channel coupling relationships. Time-frequency features and channel features are extracted from it, and an adaptive mask is generated based on the fluctuation degree and coupling degree of the channel. Selective enhancement processing is performed on the time-frequency features, which can adaptively highlight the instantaneous time-frequency components closely related to the impact, while suppressing the interference of environmental noise and redundant channels, significantly improving the signal-to-noise ratio and robustness of the features. Furthermore, the enhanced time-frequency features are compressed by impact weighted pooling, and monotonic constraints are applied to the compressed target features for discrimination. This yields state prediction results (such as normal, warning, danger, etc.) that conform to the monotonic and irreversible change law of hook damage or risk. Moreover, the prediction results have physical consistency, high accuracy and strong anti-interference ability, and can truly reflect the safety trend under the impact accumulation effect during hoisting.

[0017] 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.

[0018] 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

[0019] 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.

[0020] Figure 1 A flowchart of a method for multi-data fusion and monitoring status assessment of smart hooks provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of multi-sensor monitoring data provided in an embodiment of the present invention; Figure 3 A schematic diagram of the overall impact activation intensity at each time step within the entire time window provided for embodiments of the present invention; Figure 4 This invention provides a schematic diagram illustrating the change of instantaneous time-frequency ridge value and instantaneous frequency mutation index corresponding to the impact weight over time steps, as provided in an embodiment of the invention. Figure 5 This is a schematic diagram of a multi-data fusion and monitoring status assessment device for a smart hook, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] 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.

[0022] 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.

[0023] To facilitate understanding, the multi-data fusion and monitoring status assessment method for smart hooks provided in this embodiment of the invention will be described first, referring to... Figure 1 The method includes the following steps: Step S102: Obtain multi-sensor monitoring data of the smart hook during the lifting process, and determine the impact weight of the corresponding monitoring channel data based on the instantaneous frequency change index of each monitoring channel data in the multi-sensor monitoring data.

[0024] The intelligent hook is the main body of a lifting hook, which can be monitored through multiple channels using various sensors to obtain multi-sensor monitoring data. For example, multiple sensors such as accelerometers, strain gauges, and inclinometers can be used to simultaneously monitor physical quantities such as the hook's vibration, force, and attitude. Regarding the lifting process, the corresponding data includes dynamic response signals generated by the hook at various stages such as lifting, bearing, swinging, emergency stop, and unloading. Each monitoring channel can be described as the output of an independent sensor or a signal along a specific axis of a single sensor. For example, the X-axis channel of the accelerometer reflects the impact vibration of the hook in the horizontal direction, while the strain gauge channel reflects the tensile stress fluctuations at the hook's critical section.

[0025] In this invention, different channels exhibit varying response intensities and time delays to the same impact event, making it difficult for a single channel to independently characterize the overall impact state of the hook. In this embodiment, based on the instantaneous frequency mutation index (e.g., the degree of drastic frequency jumps within a very short time, characterizing the non-stationary transient components caused by mechanical impact in the channel), the relative intensity of the impact on each channel during the current time period can be determined. This intensity is further quantified into a corresponding impact weight, allowing subsequent processing to automatically focus on the sensor channel that best reflects the abnormal force on the hook. For example, a channel with a higher instantaneous frequency mutation index has a greater impact weight.

[0026] Step S104: Based on the short-term correlation strength of the impact weight of each monitoring channel data in the local time neighborhood, dynamic impact correlation is performed on the monitoring channel data in the time window to construct a time series feature containing the original signal and cross-channel coupling relationship.

[0027] After obtaining the impact weights for each monitoring channel, this embodiment of the invention further examines the intrinsic temporal and spatial relationships between impact events in different channels. During actual lifting operations, a single mechanical impact (such as a sudden lifting or emergency braking of a heavy object) often simultaneously triggers responses from multiple sensor channels, and the responses of different channels exhibit a definite temporal sequence and intensity correlation. For example, the strain gauge channel typically senses the load change first, followed by the accelerometer channel detecting the resulting structural vibration. To capture this cross-channel propagation characteristic, this embodiment of the invention introduces the concept of a local temporal neighborhood, such as a short-time window centered on the current moment with a fixed length (e.g., tens to hundreds of milliseconds), to limit the immediate impact range of the impact event.

[0028] Within this local time neighborhood, the short-time correlation strength between the impact weights of each monitoring channel and the impact weights of other channels is calculated. This quantifies the similarity and synchronization of the two channel impact weight sequences within this window. A high short-time correlation strength indicates that the impacts experienced by these channels within the same time window originate from the same physical excitation. Channel pairs with correlation strengths exceeding a preset threshold can be dynamically correlated within a global time window to ultimately construct a feature representation of multi-channel coupling (i.e., corresponding temporal features). This feature representation not only preserves the original amplitude, frequency, and other signal content of each monitoring channel but also explicitly embeds the temporal features of cross-channel coupling relationships (e.g., the order of impacts between channels, quantitative values ​​of coupling strength, and delay or synchronization patterns), enabling a complete description of the overall dynamic response of the hook structure under impact.

[0029] Step S106: Extract time-frequency features and channel features from the time-series features respectively. Generate a mask based on the fluctuation degree of the channel features and the degree of cross-channel coupling. Use the mask to adaptively enhance the time-frequency features to obtain an enhanced time-frequency feature representation.

[0030] The aforementioned time-series characteristics are composed of the raw signals from each monitoring channel and the impact correlation information between channels. The raw signals from each channel exhibit energy distribution characteristics that vary with time and frequency, i.e., time-frequency information, which clearly shows which frequency bands and times the impact components are concentrated in. It also includes channel information, namely the fluctuation level of each channel's own data (e.g., variance, reflecting signal activity) and the degree of cross-channel coupling between different channels (i.e., short-time correlation strength, reflecting the tightness of the impact event's correlation between channels). This embodiment of the invention utilizes the fluctuation level and cross-channel coupling degree in the channel information to generate a mask (the mask's dimension is the same as the time-frequency information). Larger mask values ​​are assigned to time-frequency regions with high fluctuation and high coupling, while smaller mask values ​​are assigned to regions with stable fluctuation or low coupling. This allows time-frequency units corresponding to larger mask values ​​(i.e., regions where the actual impact energy is concentrated) to be preserved, while units corresponding to smaller mask values ​​(i.e., noise or irrelevant vibration regions) are suppressed or blocked, thereby obtaining time-frequency characteristics of impact enhancement. The entire process is data-driven and can adaptively highlight impact-related time-frequency units based on the actual distribution of channel fluctuation and coupling degree under the current hoisting conditions.

[0031] Step S108: Perform monotonic constraint discrimination on the enhanced time-frequency feature representation, and determine the state prediction result of the smart hook during the hoisting process based on the corresponding discrimination result.

[0032] Because the damage or safety risk of the hook exhibits an irreversible unidirectional evolution pattern with the accumulation of impact, the state can only gradually progress from normal to dangerous and will not reverse on its own. This invention also applies monotonic constraints to the aforementioned features to ensure that the evaluation results maintain monotonically non-decreasing consistency over time. Specifically, features can be monotonically correlated with the intensity of accumulated impact; for example, the output states under different impact intensities can be constrained to satisfy order invariance, thereby ensuring that the final state prediction strictly follows a monotonically progressive path of "normal → attention → warning → danger". This method of using physical priors to guide the judgment direction avoids logical contradictions that may arise from traditional data-driven methods (such as risk increasing first and then decreasing), ultimately outputting a physically reliable intelligent hook state prediction result.

[0033] Furthermore, based on the above embodiments, this invention also provides another method for multi-data fusion and monitoring status assessment of intelligent hooks, which elaborates on the above steps. During non-steady operation, the intelligent hook exhibits interconnected changes in triaxial acceleration, triaxial angular velocity, hook load, rope tension, motor current, and ambient temperature at different time scales. Lifting off the ground, load swaying, braking impact, and sudden load changes correspond to short-duration, high-intensity impacts, while the steady lifting phase corresponds to low-intensity, slow changes. For the aforementioned monitoring data, multi-sensor data can be collected from the intelligent hook during lifting, lowering, slewing, luffing, braking, emergency stopping, and load transfer processes.

[0034] The system can synchronously collect data from multiple sensors on the smart hook at a uniform sampling frequency and generate corresponding multi-sensor monitoring data according to a sliding time window. The specific steps are as follows: 1> A three-axis accelerometer, a three-axis gyroscope, a hook load sensor, a rope tension sensor, a motor current sensor, and an ambient temperature sensor are installed on the intelligent hook. The sensor channels for data acquisition include: acceleration... Axial channel, acceleration Axial channel, acceleration Axial channel; angular velocity Axial channel, angular velocity Axial channel, angular velocity The sensor channels include: shaft channel; hook load channel; rope tension channel; motor current channel; and ambient temperature channel. The total number of sensor channels is denoted as... , This represents the number of sensor channels included at each sampling time, and is taken from one implementation method. .

[0035] In one implementation, the unit of the acceleration x / y / z axis channels is... For example, the sampled value of the acceleration z-axis channel at the moment of lifting can be: The units for the angular velocity x / y / z axis channels are... For example, when the hook rotates, the sampled value of the angular velocity z-axis channel can be: The unit of the hook load channel is For example, when lifting a 1.5-ton load, the sampled value of the load channel can be... The unit of rope tension channel is For example, the sampled value of the tension channel during smooth hoisting can be... The unit of motor current channel is For example, the sampled value of the current channel when the hoisting motor is working can be... The unit of ambient temperature channel is For example, the sampled value of the ambient temperature channel near the motor can be... .

[0036] 2> Synchronize all sensor channels using the same sampling frequency, denoted as . , This represents the number of sampling points collected per second, measured in Hz, which can be 50Hz in one implementation. If different sensors have different original sampling frequencies, linear interpolation or nearest neighbor alignment is first performed based on the highest common time axis to ensure that multiple channels at the same time step correspond to the same physical moment.

[0037] In one embodiment, as an example, suppose the original sampling frequency of the triaxial accelerometer is... The original sampling frequency of the motor current sensor is To achieve synchronization, the highest common timeline is set to... That is, per second Based on a time step; for motor current data, the original sampling interval is... The current data is measured in seconds. Using linear interpolation, three new values ​​are inserted between every two adjacent original current sampling points, thus transforming the current data into... This allows the data to be on the same time axis as the acceleration data, thus achieving a correspondence with the "same physical moment".

[0038] 3> Extract a length of [length] from the continuous sampled data. The sliding time window is used to obtain the first... A multi-sensor time window matrix for an original intelligent hook. ,in, Indicates the first The original multi-channel time series data corresponding to each sample, with a size of [size missing]. ; This represents the length of the time window, i.e., the number of sampling moments contained in each sample, which is 128 in one implementation. This represents the sample index, with a value ranging from 1 to... ; This represents the total number of samples.

[0039] Specifically, for state prediction, training samples can be constructed based on the aforementioned data to obtain a corresponding discrimination model. Specifically, for the... A multi-sensor time window matrix for an original intelligent hook Each sampling time point in the dataset is labeled with a state category, resulting in the first... A sequence of frame-level state labels ,in, , Indicates the first The training sample at the th ... The true state category of a time step. This represents the time step index, with a value range of 1 to... The number of state categories is denoted as , This represents the total number of identifiable state categories, and is taken from one implementation. Category 1 indicates normal, Category 2 indicates warning, and Category 3 indicates danger.

[0040] Furthermore, the frame-level status label sequence can be determined based on hoisting operation records, manual inspection records, load alarm records, emergency stop records, and video verification results. Verification is performed. If a short-term load increase occurs at the moment of lifting off the ground but there is no abnormal swing or continuous impact, the corresponding time step can be marked as normal; if the load changes abruptly, the hook swings, or the braking impact exceeds the warning threshold but does not reach the danger threshold, the corresponding time step is marked as warning; if there is violent swinging, abnormal load drop, abnormal peak rope tension, or continuous vibration after emergency stop, the corresponding time step is marked as dangerous.

[0041] In one implementation, the entire original multi-sensor time window matrix of the smart hook and the corresponding frame-level state label sequence can be divided into a training set, a validation set, and a test set. During the division, it is ensured that all three states—normal, warning, and danger—appear in both the training and validation sets, preventing the model from only learning normal operating conditions and failing to identify a few abnormal states.

[0042] During lifting, emergency braking, load swinging, and sudden load changes, the sensor signals of a smart crane hook are typically not stable periodic signals. Instead, they simultaneously exhibit amplitude abrupt changes, frequency component jumps, and enhanced cross-channel coupling within a short period. Traditional amplitude thresholding methods easily misjudge single-point noise as impacts, while traditional full-window correlation matrices average out the transient coupling relationships at the moment of impact. In this embodiment of the invention, for steps S102-S106, the aim is to extract the instantaneous frequency change index of each channel and calculate the cross-channel dynamic correlation relationship near the impact abrupt change time step to obtain the enhanced smart crane hook time-series tensor.

[0043] Regarding step S102 above, the process involves converting sensor channels with different dimensions to a comparable scale, and combining frequency jumps and amplitude jumps to locate the impact time step, thereby identifying early impacts where "the main frequency suddenly changes but the amplitude has not yet reached a fixed threshold," while suppressing false triggering caused by slowly changing channels such as ambient temperature. The impact weight can be determined through the following steps: 1) Based on the phase change of the dominant frequency band of each monitoring channel data in the multi-sensor monitoring data, determine the instantaneous frequency ridge value of the corresponding monitoring channel data.

[0044] For the dominant frequency band, the first... A multi-sensor time window matrix for an original intelligent hook And perform robust normalization on each sensor channel to obtain the first... A standardized intelligent hook multi-sensor time window matrix ,in, This represents multi-channel time series data after eliminating channel dimension differences, with the size still being [size missing]. For the first A standardized intelligent hook multi-sensor time window matrix Each channel in the process performs multi-band analytical filtering to obtain analytical responses in multiple frequency bands.

[0045] Among them, robust standardization can be performed on the corresponding monitoring channel data based on the channel median and channel interquartile range of each monitoring channel data in the multi-sensor monitoring data; multi-band analytical filtering is then performed on the standardized monitoring channel data to determine the dominant frequency band in the monitoring channel data.

[0046] In one implementation, for the first Each sensor channel can be used to calculate the median of that channel. and channel interquartile range , This represents the channel index, with values ​​ranging from 1 to... . Indicates the first Robust central values ​​of each sensor channel in the dataset; Indicates the first The robustness of each sensor channel's discreteness in the dataset is calculated by subtracting the first quartile from the third quartile. Then, for the... A multi-sensor time window matrix for an original intelligent hook The Middle Time step The sampled values ​​of the channel are standardized and represented as follows: ;in, Indicates the first In the standardized intelligent hook multi-sensor time window matrix, the first... Time step The standardized sampled values ​​of the channel; Indicates the first In the original intelligent hook multi-sensor time window matrix, the first... Time step The original sampled value of the channel; This represents a zero-preserving minimum constant, used to avoid a denominator of 0. In one implementation, it is taken as... .

[0047] For multi-band analytical filtering, the first band can be used to... The standardized time series of the channel first passes through A bandpass filter, This indicates the number of bandpass filters, for example, 16. Each bandpass filter covers a frequency range: the low-frequency band describes hook sway and slow-varying vibrations during hoisting, while the high-frequency band describes the rapid response caused by braking shocks, motor vibrations, and structural collisions. After bandpass filtering, a Hilbert transform is performed on the output of each frequency band to obtain an analytical response containing amplitude and phase information. Finite impulse response (FIR) filters can be used for the bandpass filters, with reflection-filled filter boundaries to maintain the length of the filtered sequence. Furthermore, a dominant frequency band can be selected at each time step, and the instantaneous frequency ridge value can be calculated based on the phase change of the dominant frequency band.

[0048] Specifically, for the first Channel and the Time step, comparable The amplitude of the analytical response of each frequency band is selected, and the frequency band with the largest amplitude is chosen as the first frequency band. The dominant frequency band of the time step is determined; then the phase angle of this dominant frequency band at adjacent time steps is read, and the phase angle is expanded along the time direction to avoid phase angle fluctuations. arrive The jump between time steps caused an error in the frequency calculation; finally, the instantaneous frequency ridge value was converted based on the phase difference between adjacent time steps.

[0049] Among them, for the first The sample, the first Channel, First The instantaneous frequency ridge value of the time step can be denoted as: , Indicates the first The channel is in The instantaneous frequency of the dominant vibrational component near the time step, expressed in Hz. Furthermore, for The boundary position can make equal .

[0050] In one embodiment, as an example, for the acceleration z-axis channel (the first... Analyze the channel, assuming it is in the first channel. The time step is obtained after multi-band analytical filtering. The analytical signals of each frequency band; comparing the instantaneous amplitudes of these 16 frequency bands, it was found that the 5th frequency band (center frequency approximately...) amplitude If the maximum value is found, then the frequency band is the [number]. The dominant frequency band at the time step; read the dominant frequency band at the 1st... The phase angle of the analytical signal at the time step is In the The phase angle of the time step is After phase expansion, if the phase difference... Then the first Instantaneous frequency ridge value of time step From phase difference and time interval The conversion yields, i.e. .

[0051] 2) Based on the discrete scale of the instantaneous frequency ridge value, and the frequency jump value and amplitude jump value in adjacent time steps, determine the instantaneous frequency mutation index of the monitoring channel data.

[0052] Calculate the first Robust discrete scaling of the instantaneous frequency ridge value of the channel yields the channel frequency scale value. ,in, Indicates the first In the nth sample The normal fluctuation range of the instantaneous frequency ridge value of a channel is used to convert the frequency jumps of different channels to a similar scale. In specific implementation, the first step is to calculate the... The median of the instantaneous frequency ridge values ​​for all time steps of the channel is used. Then, the absolute deviation between the instantaneous frequency ridge value for each time step and this median is calculated. Finally, the median of the absolute deviations is taken as the channel frequency scale value. .like Too small, for example, less than Then Set as This avoids the abnormal amplification of the mutation index due to an excessively small denominator in the stable channel.

[0053] In one embodiment, for example, for a certain channel The sequence of instantaneous frequency ridge values ​​for all time steps is as follows: First, calculate the median of the sequence. Then, the absolute deviation at each time step is calculated as follows: Finally, the median of these absolute deviations is taken to obtain the channel frequency scale value. Based on this, it indicates that the normal fluctuation of the instantaneous frequency of this channel is approximately... about, The significant deviation caused by this point will not be used to calculate the normal fluctuation scale.

[0054] Further, the instantaneous frequency jump index is calculated based on the instantaneous frequency jump and the normalized sample value jump, to obtain the first... A matrix of instantaneous frequency mutations ,in, Indicates the first The shock mutation intensity of each sample across all time steps and all channels, with dimensions of [value missing]. ; Indicates the first Time step The instantaneous frequency mutation index of the channel.

[0055] In the actual implementation, the absolute value of the difference between the instantaneous frequency ridge values ​​of adjacent time steps is first calculated, and then divided by the channel frequency scale value. The sum of the normalized frequency jump value and the amplitude jump value is obtained by combining the normalized frequency jump value with the amplitude jump value. Simultaneously, the absolute value of the difference between the normalized sampled values ​​of adjacent time steps is calculated to obtain the amplitude jump value. Finally, the normalized frequency jump value is multiplied by the amplitude jump value to obtain the instantaneous frequency jump exponent, expressed as:

[0056] in, Represents the zero-limit constant, which is taken in one implementation. ; Indicates the first In the standardized intelligent hook multi-sensor time window matrix, the first... Time step The standardized sampled values ​​of the channel; Indicates the first The sample, the first Channel, First The instantaneous frequency ridge value at the time step. For The boundary position will Set to 0.

[0057] Furthermore, embodiments of the present invention also address the third... A matrix of instantaneous frequency mutations A light smoothing process is performed along the time direction to reduce spikes caused by noise at single sampling points. Specifically, a one-dimensional median filter of length 3 or 5 is used on the instantaneous frequency change exponential sequence of each channel to obtain the smoothed [then...]. A matrix of instantaneous frequency mutation indexes. In one embodiment, as an example, the instantaneous frequency mutation index sequence before smoothing of a certain channel is: Use length is One-dimensional median filtering, for the third time step, takes the value within the window. The median is After smoothing, the original point The isolated spikes were replaced with If the stationary values ​​remain unchanged before and after, then the sequence becomes .

[0058] It should be noted that median filtering only replaces isolated spikes and does not significantly change the location of the impact, making it suitable for short-term impact positioning of hooks. In one embodiment, as an example, when the hook load channel is lifted off the ground, the normalized sampled value increases from 0.1 to 0.9, while the instantaneous frequency ridge value jumps from around 3Hz to around 12Hz. In this case, both the normalized frequency jump value and the amplitude jump value are relatively large, and the instantaneous frequency mutation index of the corresponding time step is significantly increased. The ambient temperature channel usually only experiences slow drift, and even if there are small changes in the sampled value, the instantaneous frequency ridge value will not show a significant jump, and the corresponding instantaneous frequency mutation index remains at a low level.

[0059] 3) Determine the time step mutation screening threshold based on the instantaneous frequency mutation index.

[0060] 4) Determine the impact weight of the corresponding monitoring channel data based on the impact characteristics represented by the instantaneous frequency mutation index and the time step mutation screening threshold.

[0061] It is important to note that the real risk of intelligent hooks is usually not an isolated change in a single channel, but rather the coordinated interaction of acceleration, angular velocity, load, tension, and motor current within a short period. If a fixed correlation matrix is ​​directly calculated for the entire time window, the impact time step accounts for only a small proportion, and the key coupling relationships are diluted by a large number of stable time steps. This embodiment of the invention also calculates the short-term coupling relationships between sensor channels near the moment of impact, assigning impact weights to different time steps using an instantaneous frequency mutation exponent, ensuring that only impact-related time steps make a major contribution to the cross-channel correlation calculation.

[0062] First, according to the A matrix of instantaneous frequency mutations Determine the mutation screening threshold ,in, Indicates the first The threshold used to distinguish significant impact time steps from stationary time steps in the samples. In one implementation, the threshold is... A matrix of instantaneous frequency mutations The 80th percentile of all elements in the spectrum is used as the mutation screening threshold. If the noise level is high, the 85th or 90th percentile can be used; if you want to identify minor warnings earlier, the 70th or 75th percentile can be used.

[0063] Then, based on the mutation screening threshold Calculate the first Impact weight matrix .in, This represents the effective shock intensity after all time steps and all channels exceed the mutation screening threshold, with dimensions of [size missing]. ; Indicates the first Time step The impact weight of the channel. In the specific implementation, the impact weight of the channel will be... A matrix of instantaneous frequency mutations The mutation screening threshold was not exceeded. Setting the element to 0 exceeds the mutation screening threshold. The element that extends beyond the specified portion is represented as:

[0064] in, This means that results less than 0 will be set to 0. Based on this, stationary time steps and weakly noisy time steps will not dominate cross-channel correlation calculations, while strong-impact time steps will receive greater weight.

[0065] Furthermore, regarding step S104 above, the corresponding timing characteristics can be determined through the following steps: 1) Determine the short-term correlation strength between any two monitoring channels in the local time neighborhood based on the impact weight.

[0066] Among them, the first can be used as the first Establish a local time neighborhood centered on the time step Including participation in the The set of time steps dynamically correlated across channels. The local time neighborhood radius is denoted as... , This represents the number of time steps taken forward and backward, typically ranging from 8 to 16 in one implementation. When... Local time neighborhood when near the start or end of the time window Automatically truncate to 1 to Within the effective range.

[0067] The joint weight of a corresponding local time neighborhood can be determined based on the impact weights of data from multiple monitoring channels within the same local time neighborhood. Based on the joint weight and the locally mean-reduced sampled values ​​of the corresponding monitoring channel data, a weighted synchronous change is determined. The weighted synchronous change is then normalized to determine the short-time correlation strength. In one implementation, any two sensor channels can be... and In the local time neighborhood The short-time correlation strength is calculated by internally weighting the impact, and the first... The sample at the th Cross-channel dynamic correlation matrix of time steps .in, Indicates the first The short-term coupling relationship between the sensor channels at each time step, with a size of ; Indicates the first Channel and the The channel is in The short-term correlation strength near the time step is a cross-channel dynamic correlation matrix. The element in the i-th row and j-th column. In one implementation, the element in the i-th row and j-th column can be... Channel and the The impact weights of the channels at the same local time step are multiplied together to form the joint weight for participating in the relevant calculations at that local time step. This joint weight is then multiplied by the local mean-reduced sample values ​​of the two channels to obtain the weighted synchronization change. Finally, the weighted vector length is used for normalization to obtain a value close to... The short-term correlation strength up to 1. The calculation method is expressed as:

[0068] in, Represents the local temporal neighborhood Time step index within; Indicates the first In the impact weight matrix, the first... Time step Impact weight of the channel; Indicates the first In the impact weight matrix, the first... Time step Impact weight of the channel; Indicates the first In the local mean-reducing matrix, the th Time step Local mean-reduced sampled values ​​of the channel; Indicates the first In the local mean-reducing matrix, the th Time step Local mean-reduced sampled values ​​of the channel; Represents the zero-limit constant, which is taken in one implementation. .

[0069] When the local time neighborhood Inner Channel and the When the impact weights of the channels are all close to 0, Setting it to 0 indicates that there is no reliable evidence of shock coupling within the current local time neighborhood. For diagonal elements... If the first Channel in local time neighborhood If a valid impact weight exists in memory, set it to 1; otherwise, set it to 0.

[0070] Among them, it is also possible to be in the local time neighborhood. Internal to the first A standardized intelligent hook multi-sensor time window matrix Perform local mean removal to obtain the first... Local mean-reducing matrix Based on this, the impact of hook load baseline, sensor zero-point offset, and motor current operating condition baseline on correlation calculations can be reduced, allowing the correlation calculations to focus more on whether the changes before and after the impact are synchronized. Among these, Indicates surrounding the first The local mean-removed result obtained from the time step calculation has a size of ; Represents the local temporal neighborhood The number of valid time steps contained within. In one implementation, for the first... Channels can compute local temporal neighborhoods. The average value of the standardized sampled values ​​for that channel is then calculated by subtracting this average value from the standardized sampled values ​​for each local time step.

[0071] 2) Based on the short-time correlation strength, determine the impact coupling results of the corresponding local time neighborhood.

[0072] In the specific implementation, the processing rules for diagonal elements are defined to represent the state of a single channel during the impact, including the following two cases: Scenario 1: Channel exist There is an effective impact weight in memory; for example, All time steps within If not all values ​​are 0, it indicates that the channel has experienced an impact in the current local neighborhood. At this point, [the value will be...]. Forced to be set , indicating channel It is in a reliable state of impact or change; Scenario 2: Channel exist There are no effective impact weights within; for example, Impact weights at all time steps All This indicates that the channel was very stable during this period, at which point... Forced to be set This indicates that there is no evidence of impact coupling in the channel itself.

[0073] 3) Traverse all time steps to obtain the impact dynamic correlation tensor of sensor monitoring data at different time steps.

[0074] Iterate through all time steps to obtain the first step. The dynamic correlation tensor of the impact ,in, Indicates the first A sample cross-channel short-time coupling structure at different time steps, with a size of [missing information]. ; Indicates the first Time step Channel and the The impact-weighted short-time correlation strength between channels, and by The assignment is obtained. In the specific implementation, starting from the first... The time step to the Each time step, for each time step Build one Cross-channel dynamic correlation matrix ; Complete all After looping for n time steps, these matrices are stacked along the time dimension to obtain a final size of Impact dynamic correlation tensor .

[0075] 4) Combine the impact dynamic correlation tensor with multi-sensor monitoring data to obtain time-series characteristics.

[0076] Among them, the first A standardized intelligent hook multi-sensor time window matrix (i.e., multi-sensor monitoring data) and the first The dynamic correlation tensor of the impact Combine to obtain the first The timing tensor of a smart hook (i.e., time sequence features). Among them, This represents network input data that simultaneously contains both channel-specific normalized sample values ​​and cross-channel impulse coupling characteristics, with a size of [size missing]. .

[0077] In specific implementation, the first A time-series tensor of an enhanced smart hook The first feature layer stores the first A standardized intelligent hook multi-sensor time window matrix That is, each time step and each channel retains its own standardized sampled value; subsequently The first feature layer stores the first feature layer. The dynamic correlation tensor of the impact The dynamic correlation value between the current channel and other channels. Based on this, each time step and each channel not only contains its own signal, but also the linkage relationship with other sensor channels at the moment of impact.

[0078] For example, when emergency braking occurs, acceleration Axial channels and angular velocities If both shaft channels exhibit high impact weights and show consistent positive and negative fluctuations in their local time neighborhoods, the corresponding short-term correlation strength is high. If the ambient temperature channel has no impact weight, even if it slowly changes along with other channels over long time scales, it will not form a high correlation value in the impact dynamic correlation tensor. It should be noted that in this embodiment of the invention, the cross-channel relationship is dynamically calculated at each time step based on the local time neighborhood and impact weights. This preserves the physical coupling of multiple sensors during transient conditions such as lifting off the ground, emergency braking, and sudden load changes, while avoiding the dilution of key transient features in stable time steps. Figure 2 A schematic diagram of multi-sensor monitoring data according to an embodiment of the present invention is shown. Figure 2 A multi-sensor time window matrix heatmap is used to represent the complete multi-channel time-series data after robust normalization. Figure 2 In the figure, the horizontal axis represents the time step (unit: sampling point, 128 steps in total), and the vertical axis represents the sensor channel index (0 to 9 corresponding to ten types of physical quantities). The color mapping represents the standardized sampled value of each channel at each time step (dimensionless, obtained by subtracting the median and dividing by the interquartile range). The warm-colored areas in the figure represent impact moments that deviate positively from the normal level, while the cool-colored areas represent relatively stable or negatively abrupt changes. It can be seen that the acceleration Z, hook load, rope tension, and other channels show obvious synchronous high / low value clusters around steps 40-46 and 95-100, while the ambient temperature channel (index 9) shows a gradual color change throughout the time window, reflecting the non-stationary linkage characteristics of multi-sensor signals under impact conditions. This proves that the standardization process can successfully eliminate the dimensional differences between channels while preserving the cross-channel comparability of impact events.

[0079] The aforementioned smart hook time-series tensor (i.e., time-series features) simultaneously contains the sensor's own time-series signal and impact dynamic correlation information. The sensor channel dimension reflects different physical quantities and their linkage relationships, the time dimension reflects the impact generation, attenuation, and recovery process, and the frequency structure reflects vibration mode changes. If a common sequence model is directly used to mix all dimensions, it is easy for a large number of stationary time steps to suppress a small number of critical impact time steps. Regarding step S106 above, this embodiment of the invention can also construct a state classification network to extract time-frequency feature representations and channel feature representations respectively, and then enhance critical channels based on channel importance masks. In specific implementation, feature enhancement can be performed through the following steps: 1) Parallel encoding of channel features and time-frequency features This step involves separately encoding the channel coupling information and time-frequency vibration information in the aforementioned time-series features (smart hook time-series tensor). The channel feature representation focuses on describing "which sensors change together under the current operating conditions," while the time-frequency feature representation focuses on describing "what kind of impact vibration mode appears in the sensor signal."

[0080] a- can make the first The timing tensor of a smart hook As the input to the state classification network, let it be denoted as the . Network input tensor ,in, This represents the input data received by the state classification network, with a size of [size missing]. . This represents the input feature dimension for each time step and each channel. ; Indicates the first Time step The standardized sampled values ​​of the channel; Indicates the first Time step Impact dynamic correlation characteristics between channels.

[0081] b-From the first Network input tensor The first feature layer is taken as the input to the time-frequency coding branch. The time-frequency coding branch performs a one-dimensional depthwise convolution on each sensor channel to obtain the... Each time-frequency feature represents ,in, This represents the impact vibration features extracted by multi-scale temporal convolution, with a size of [size missing]. ; This represents the dimension of the time-frequency feature, which is either 32 or 64 in one implementation.

[0082] In its implementation, the time-frequency coding branch contains multiple sets of one-dimensional convolutional kernels. The kernel lengths can be 5, 11, and 21, respectively used to capture short-term spikes, moderate sustained vibrations, and longer oscillation trends. Each convolutional kernel slides along the time axis and processes each sensor channel independently, preventing premature mixing of physical quantities from different channels in the initial stage. The convolutional boundaries are zero-padding or reflection-padding to maintain the output time length. The convolutional outputs, after batch normalization and a nonlinear activation function, are concatenated along the feature dimension to form the ... Each time-frequency feature represents In one implementation, three sets of one-dimensional convolutional kernels are used, each with a length of... , , The one-dimensional convolution kernel is as follows: Short-term spikes: The convolution kernel length is 5, and it is initialized to... This shape, similar to that of a difference operator, responds strongly to sudden changes in data (such as braking shocks); Medium-duration vibration: The convolution kernel length is 11, which is initialized as a damped oscillating waveform with a high center frequency (such as a 50Hz sine wave with 5 cycles, the envelope first increases and then decreases), which is suitable for capturing structural responses such as motor vibration. Longer oscillation tendency: The convolution kernel length is 21, which is initialized as a low-frequency smooth oscillating waveform (such as a 2Hz sine wave with 1.5 cycles), which is sensitive to the slow oscillation of the hook and the load transfer process.

[0083] In one implementation, the convolutional kernels in the time-frequency coding branch can be initialized using a bandpass filter shape. Specifically, short convolutional kernels are initialized with a strong high-frequency response differential shape to detect sudden stop impacts and structural collisions; long convolutional kernels are initialized with a low-frequency smooth oscillation shape to detect hook sway and slow load changes. During training, the convolutional kernels are updated as learnable parameters to adapt to the actual vibration modes under different hook models and sampling frequencies. In one embodiment, as an example, a short-time spike convolutional kernel (length 5) can be initialized with a high-frequency differential shape, for example... The kernel outputs close to 0 during stable signal periods, but at points such as braking or impact, the normalized values ​​of adjacent sampling points abruptly change from near 0 to large positive / negative values, causing the convolution response to immediately produce a high-amplitude output. A moderately sustained vibration convolution kernel (length 11) can be initialized as a damped oscillating waveform with a high center frequency. For example, centered at 50Hz, at a 50Hz sampling rate, each cycle has one sampling point, for a total of five sampling points across five cycles. The waveform with an envelope that first increases and then decreases can be set as follows: It is suitable for matching the structural response of motor vibration and other components in the vicinity of this frequency range; the longer swing trend convolution kernel (length 21) can be initialized as a low-frequency smooth oscillation waveform, for example, for detecting hook swing of about 2Hz. Its shape can be a segment of a low-frequency sine wave, such as [0.1, 0.3, 0.5, ...], which is sensitive to low-frequency drift during slow hook swing and load transfer process, and is used to detect low-frequency drift during slow hook swing and load transfer process.

[0084] c- runs in parallel with the time-frequency coding branch, for the th Network input tensor Perform channel feature projection to obtain the first Each channel feature representation ,in, This represents the channel structure formed by standardized sampled values ​​and dynamic correlation features of the impact, with a size of [missing information]. ; This represents the channel feature dimension, which is 32 in one implementation.

[0085] In the actual implementation, for each time step and each channel ,Will Inputting the same linear projection layer, we get a length of The channel feature vectors are then normalized and activated by the ReLU function to obtain... .because Simultaneously includes the first Channel's own sampled value and the first The dynamic correlation value of the impact between the channel and other channels, therefore the first Each channel feature representation It can express the relative importance of the current channel in a multi-sensor system.

[0086] In one embodiment, it is assumed that in the... Time step, number Network input feature vectors for channels (such as hook load channels) It is a length of The vector, which contains not only the normalized sampled values ​​of the load channel itself (e.g., It also includes its impact dynamic correlation values ​​with nine other channels, such as acceleration and tension; this vector is input to a shared linear projection layer (a (weight matrix), output a length of The channel feature vector; if the correlation value is high, the projected vector will encode the information that "the load is the core of the current impact linkage".

[0087] It should be noted that the embodiment of the present invention sets up time-frequency encoding and channel encoding in parallel, rather than directly mixing all sensor channels at the input layer. This allows the state classification network to learn "vibration mode" and "sensor coupling mode" separately, improving the interpretability of the model for the source of hook impact.

[0088] 2) Time-frequency feature enhancement guided by channel importance mask Different sensor channels contribute differently to hook state classification. During lifting off the ground, the hook load channel and rope tension channel are typically more sensitive; during emergency braking, the triaxial accelerometer channel and triaxial gyroscope channel are typically more sensitive; the ambient temperature channel mainly provides slowly varying background information and should not be overemphasized in impact identification. This step involves generating a channel importance mask based on the channel characteristic fluctuation level and cross-channel coupling degree, and using this mask to adjust the time-frequency feature representation.

[0089] To adjust the time-frequency feature representation using importance masks, refer to the following steps: 1. Regarding the first Each channel feature representation The Middle The time average of the feature vectors of all time steps of the channel is calculated to obtain the first... The average channel eigenvector of each channel is calculated; then the Euclidean distance between the channel eigenvector and the average channel eigenvector at each time step is calculated; finally, the average of the Euclidean distances over all time steps is taken to obtain the first... In the nth sample Channel time fluctuation score .in, Indicates the first The intensity of characteristic changes of the channel within the current time window; the larger the value, the more likely the channel is to participate in shocks or state changes.

[0090] In one implementation, for the first The time fluctuation score of a channel (such as a rope tension channel). The calculation process can be represented as follows: Take the eigenvectors of the channel at all time steps, calculate the average, and obtain the "average channel eigenvector". For each time step, calculate the Euclidean distance between its eigenvector and the "average eigenvector"; if the eigenvector of the tension channel deviates significantly from the average value at the moment of lifting off the ground, then the distance at that moment is large. Average the distances over all time steps. If the deviation at moments such as lifting off the ground is significant, the final result will be... A larger value indicates that the characteristics of the channel change drastically within that time window.

[0091] 2. From the first The dynamic correlation tensor of the impact Extract the first The correlation value between the channel and other channels is calculated. Cross-channel coupling score .in, Indicates the first The extent to which a channel participates in multi-channel shock linkage within the current time window.

[0092] In the specific implementation, All time steps and all channels Summing the squares of the above and taking the square root, we get the first square root. The correlation strength of the channels; then divided by the sum of the maximum correlation strength of all channels and the zero-limit constant, so that... Normalized to approximately 0 to 1. The zero-limit constant can be taken as... .

[0093] In one implementation, the calculation of the first... Cross-channel coupling score of the channel (acceleration x-axis) The method can be expressed as: 1. In Extract all time steps Other channels With the Channel correlation value Calculate the sum of squares: 2. Repeat this operation for all channels to obtain... And find the maximum value. ,in, Indicates the first The sum of squares of correlation of the channels, Indicates all 3. Calculate the maximum value among the correlated sums of squares of each channel; ,if near This indicates that the x-axis acceleration channel is closely linked to the impact of other channels within this window, and is the core of the coupling.

[0094] 3. According to the first Channel time fluctuation score and cross-channel coupling score Calculate the first Channel importance score .in, Indicates the first The overall contribution of a channel to state classification within the current time window is calculated as follows: .in, Indicates to The time fluctuation score is normalized across all channels. Based on this, isolated noise with only time fluctuation but not coupled with other channels will not be over-amplified, while key channels with strong time fluctuation and strong cross-channel coupling will receive higher importance scores.

[0095] Furthermore, the importance scores of all channels can be combined to form the first... Importance score vectors ,in, This represents the set of importance for all sensor channels within the current time window, with dimension 1. .

[0096] 4. In this embodiment of the invention, the channel gating vector is used as the corresponding channel importance mask. Each element is generated based on the importance score of the corresponding sensor channel, with a value between 0 and 1. Furthermore, the time-frequency features can be weighted channel-by-channel based on this mask to enhance important channels. Specifically, the weighting can be applied to the ... Importance score vectors Performing a linear transformation and sigmoid activation yields the first... Channel gate vectors . Represents the gating response of all channels, with dimension . ; Indicates the first The gating response of the channel has a value range from 0 to 1.

[0097] In one embodiment, for example, suppose there are Each channel yields an importance score vector. After passing through a linear layer and a sigmoid function, the output channel-gated vector The value is Among them, the gating responses of acceleration, angular velocity, load, and tension channels are close to... The gating response of the ambient temperature channel is close to This indicates that the model adaptively focuses its attention on the core physical quantities related to the impact.

[0098] Furthermore, it can also be based on the first Channel gate vectors Calculate the first Channel enhancement factor Furthermore, the channel enhancement coefficient is applied channel by channel to the first... Each time-frequency feature represents , obtained the An enhanced time-frequency feature representation.

[0099] in, Indicates the first The amplification or compression ratio of the channel's time-frequency characteristics is calculated as follows:

[0100] in, This represents the minimum enhancement factor, which is 0.5 in one implementation. This represents the maximum enhancement factor, which is 2.0 in one implementation. When When approaching 1, the first The channel's time-frequency characteristics are amplified; when When approaching 0, the first The channel time-frequency characteristics are compressed. The enhanced time-frequency characteristic representation can be expressed as: , representing the time-frequency vibration characteristics after channel importance masking, with a size of .

[0101] In the actual implementation, for each time step and each channel can Multiply ,get This multiplication is performed broadcast along the time-frequency characteristic dimension, without changing the number of time steps or channels. In one embodiment, for example, when the hook load channel and the rope tension channel simultaneously experience abrupt changes at the moment of lifting off the ground, and the two channels are at the [missing information - likely a time step or channel number]... The dynamic correlation tensor of the impact If two channels have a high correlation strength, their importance scores are high, and the channel enhancement coefficient can approach 2.0. When the ambient temperature channel changes only slowly and there is no significant impact coupling with other channels, the channel enhancement coefficient can approach 0.5. Therefore, the state classification network focuses more on sensor channels directly related to hook impact in subsequent calculations. It should be noted that this step does not fixate on certain sensors being always important, but rather adaptively generates a channel importance mask based on channel fluctuations and cross-channel coupling relationships within each time window, thus adapting to changes in risk sensors during different operational phases such as hoisting, slewing, braking, and swinging.

[0102] Furthermore, the enhanced time-frequency features of the impact clearly highlight the impact component on the time-frequency plane. To avoid introducing a large amount of redundant information in state discrimination and reduce the computational burden, this embodiment of the invention further compresses the above features in step S108 to obtain the most critical impact information. The impact weight itself reflects the relative intensity of the impact on each monitoring channel. Using it as a pooling weight coefficient allows the time-frequency region corresponding to the channel with the larger impact weight to contribute a higher proportion of information during the compression process, while the contribution of the region with the smaller weight is relatively weaker. Specifically, the impact activation intensity at each time step can be determined based on the overall intensity of the impact weight at each time step; using the impact activation intensity as the aggregation weight, the enhanced time-frequency features are weighted and aggregated in the time dimension to obtain the compressed feature sequence.

[0103] In the time window of the intelligent hook, a large number of sampling moments occur during the stable operation phase. The evidence that truly determines the state category is usually concentrated in a short period before and after the impact. Directly inputting the complete time series into the cyclic model would increase the computational load and easily lead to the averaging of key impact time steps. This step involves generating an impact activation sequence based on the instantaneous frequency mutation exponent matrix and performing time compression with the impact activation intensity as the weight. The specific steps are as follows: 1> According to the first Impact weight matrix Calculate the first shock activation sequence ,in, Represents the total impact intensity at each time step, with a length of [missing information]. ; Indicates the first The sample at the th The overall impact intensity at each time step. In specific implementations, for the first... The impact weights of all sensor channels at each time step are summed to obtain... If multiple sensor channels simultaneously exceed the mutation screening threshold at the same time step, then A clear peak is formed; if the first If the time step is in a stable phase, then Approaching 0. Furthermore, the first... shock activation sequence Divide it by the sum of its maximum value and the zero-limit constant to normalize it to the range of 0 to 1.

[0104] For example, if at a certain time step There are three channels for impact weights. Exceeding the threshold, respectively , , The remaining channels are The shock activation intensity at that time step If this value is the maximum value within this window, then the normalized shock activation intensity at that time step is... This forms an extremely strong peak; if at a certain time step all channel impact weights are equal... Then the time step .

[0105] 2> According to compression ratio The original timeline is divided into multiple consecutive compression blocks to obtain the number of compression time steps. ,in, This indicates the number of raw time steps contained in each compressed block, which is 4 in one implementation. This represents the number of time steps after time compression, and can be taken as... ; This indicates rounding up to the nearest integer. Each compressed block is denoted as ,in, This represents the compression time step index, with a value ranging from 1 to... ; Including the first in the original timeline The time step to the At one time step, the last compressed block is insufficient. Each time step is processed according to the actual remaining time steps.

[0106] In one embodiment, for example, the original time window length Compression ratio Then the number of compression time steps Including: the One compressed block Includes raw time index ;No. One compressed block Includes raw time index ;...;No. One compressed block Includes raw time index .

[0107] 3> In each compressed block Inside, according to the first shock activation sequence Computation time compression weights ,in, Indicates the first Within the compressed block, the first... The weights retained at each time step are calculated as follows:

[0108] in, This represents the impact focusing coefficient, used to control the sensitivity of the time compression weight to the impact activation intensity; in one implementation, it is set to 3.0. Represents compressed blocks The time step index within. For those not belonging to a compressed block. The time step, not participating in the first The weighted calculation of each compressed block. If the compressed block If all the impact activation intensities are close to 0, then the time compression weights are nearly uniformly distributed, allowing stable samples to retain average operating characteristics.

[0109] 4> Utilize time compression weights For the Each enhanced time-frequency feature represents Perform weighted pooling to obtain the first Each compressed time-frequency feature represents . Represents the time-frequency characteristics after time compression, with a size of In its implementation, weighted pooling employs a weighted pooling operation based on impact weights. For the ... The first compressed block, the... The channels and each time-frequency feature dimension are used to weight the enhanced time-frequency features of each original time step within the compressed block according to the time compression weights. We get the weighted sum. .

[0110] In one embodiment, for example, in the compressed block Inside, there is The four time steps correspond to the normalized shock activation intensities as follows: ;when When, their nonnormalized weights are The calculated final time compression weight Approximately ; Enhanced time-frequency characteristics within the compressed block When performing weighted summation, The characteristics of time occupy This contributes to ensuring that the peak impact information is not lost after time compression.

[0111] Furthermore, the same amount of time can be used to compress the weights. For the Each channel feature representation Perform weighted pooling to obtain the first Each compressed channel feature representation . This represents the channel structure features after time compression, with dimensions of [size missing]. Due to the first Each compressed time-frequency feature represents and the Each compressed channel feature representation Using the same time compression weights, the two features remain aligned at the compression time step.

[0112] In one embodiment, following the example above, the first... The time compression weight calculated for each compressed block is: This weight vector is related to the enhanced time-frequency features. While performing weighted pooling, it also involves channel feature representation. Same time step We perform a weighted summation of the eigenvectors to obtain... Based on this, it is ensured that the compressed time-frequency features and channel features are strictly aligned in time, both reflecting the principle of... Information convergence driven by constant impact.

[0113] 5> Furthermore, the first Each compressed time-frequency feature represents and the Each compressed channel feature representation The features are concatenated along the feature dimension, and the concatenated features of all channels within each compressed time step are expanded to obtain the first... A fusion compressed feature sequence .in, This represents the compressed feature sequence input to the subsequent sequence modeling module, with a size of [size missing]. ; This represents the fusion feature dimension at each compressed time step. ; Indicates the first The fused feature vectors of each compressed time step.

[0114] In one embodiment, for example, when the compression ratio At that time, a certain compression block contains four consecutive time steps, with normalized impact activation intensities of 0.1, 0.9, 0.2, and 0.0, respectively. When the impact focusing coefficient... When the second time step has the highest time compression weight, the compressed feature mainly retains the shock response of the second time step; if the shock activation intensity of all four time steps is close to 0, then the weight of the four time steps is close to 0.25, and the compression result is close to ordinary average pooling.

[0115] It should be noted that this step determines the contribution of each time step in the compression result based on the intensity of the shock activation. This can shorten the sequence length while preserving the shock peak, pre-shock changes, and post-shock decay patterns, and reduce the interference of stationary time steps on state classification.

[0116] Furthermore, embodiments of the present invention also analyze the shock activation sequence, referring to... Figure 3 The diagram illustrates the overall impact activation intensity at each time step within the entire time window. The horizontal axis represents the time step, and the vertical axis represents the normalized impact activation intensity (001, dimensionless). The activation intensity forms two significant peaks at approximately steps 40 and 95, corresponding to the lifting-off impact and braking impact, respectively, while the intensity approaches zero during the remaining time periods. The red dashed line in the diagram indicates the impact jump threshold of 0.7, used for rapid state transitions in sudden hazardous conditions.

[0117] Furthermore, in the state prediction part of step S108 above, time-series dependency modeling can be performed on the compressed feature sequence to obtain the hidden state sequence of the compressed time step; the hidden state sequence is restored to the original time resolution to obtain the state probability distribution of each original time step; monotonic constraints are applied to the state probability distribution to prohibit state back-jump of the state probability distribution, thereby obtaining the state prediction result of the smart hook at each time step in the hoisting process.

[0118] The state of a smart hook exhibits risk monotonicity in an engineering sense, meaning it can evolve from normal to warning, and then to danger. Within the same short time window, if evidence of danger has already appeared, it should not immediately jump back to normal due to probability fluctuations at a single time step. Simultaneously, the hook may experience sudden danger; therefore, the decoding process needs to allow for a direct transition from normal to danger under strong evidence conditions. In this embodiment of the invention, the compressed temporal changes can be modeled using a gated cyclic unit, the original temporal resolution can be restored through time backfilling decoding, and finally, the final state sequence can be obtained through monotonic constraint decoding.

[0119] 1> The first A fusion compressed feature sequence Input the compressed time steps sequentially into the gated loop unit to obtain the first... A sequence of compressed hidden states .in, Represents the state evolution characteristics on the compressed time axis, with a size of ; This represents the hidden state dimension of the gated loop unit, which is 128 in one implementation. Indicates the first The hidden state vector of each compressed time step.

[0120] In its implementation, the gated loop unit controls the retention ratio of the historical state from the previous compressed time step by updating the gates, and controls the coverage ratio of the historical state by the current impact feature by resetting the gates. The initial hidden state can be set to an all-zero vector. Compared to ordinary loop units, the gated loop unit can retain the lingering effects of the impact in a shorter sequence, making it suitable for describing the process of hook impact from generation to decay.

[0121] 2> For each original time step Determine the corresponding compressed block index ,in, , Indicates the first The time step belongs to which compressed block? Then, read the... A sequence of compressed hidden states In and with the Normalized shock activation intensity at time step The relative position codes within the block are concatenated to obtain the first... Time-filled decoding vector ,in, Indicates the use of restoring the first Decoding input for time step state probabilities.

[0122] In practical implementation, the intra-block relative position encoding can be taken as... , used to indicate the first The relative position of a time step within the current compressed block ranges from 0 to approximately 1. Adding intra-block relative position encoding ensures that different original time steps within the same compressed block will not obtain completely identical state probabilities, helping to recover fine-grained differences between the pre-impact, peak-impact, and impact decay stages.

[0123] In one embodiment, for example, assume the compression ratio Original time step It belongs to the category of compressed blocks. ( (This indicates a rounding up operation), the time-backfilled decoding vector for this time step. The composition is as follows: a>Hidden state: Reading That is, the hidden state vector of the second compressed block after being processed by the gated loop unit; b> Impact strength: splicing That is, the normalized shock activation intensity of the 6th original time step; c> Position encoding: Relative position within the splicing block ; This spliced ​​vector It also includes the macroscopic state evolution, the instantaneous impact intensity, and its relative position during the impact process (located in the middle of the compression block), which can more accurately decode the state at that moment.

[0124] 3> Furthermore, the first Time-filled decoding vector Inputting the multilayer perceptron classifier, we obtain the first... The state logic value vector of the time step .in, Indicates the first The unnormalized score of each state category at each time step, with dimensions as follows: Then, consider the state logic value vector. Perform a Softmax transformation to obtain the first... State probability vector at time step .in, Indicates the first The probability distribution of time steps belonging to the categories of normal, warning, and danger has a dimension of .

[0125] In one implementation, the multilayer perceptron classifier consists of two fully connected layers connected in series, as follows: First layer: Input dimension is... (Right now (dimension of the output dimension) (For example ), followed by batch normalization and ReLU activation function; second layer (output layer): input dimension is The output dimension is (Number of state categories, for example) The output is then converted into a state probability vector by the Softmax function. , Indicates the first The sample at the th The state probability vector at each time step.

[0126] 4> Further, traverse all original time steps to obtain the first... A sequence of state probabilities ,in, , indicating the first The state probability output of the nth sample at the original time resolution. During the inference phase, for the nth sample... A sequence of state probabilities Perform monotonic constraint decoding to obtain the first... A monotonic constraint state sequence ,in, , Indicates the first The final output state category for each time step. In the specific implementation, the state categories are sorted by risk level into Normal, Warning, and Danger. Regular legal transitions include Normal to Normal, Normal to Warning, Warning to Warning, Warning to Danger, and Danger to Danger; Warning to Normal, Danger to Warning, and Danger to Normal are prohibited. To avoid excessive delays for sudden dangers, when the time step... The probability of the hazard category at the time step is greater than the hazard jump threshold. And the first Time-step normalized shock activation intensity Greater than the impact jump threshold At this time, a direct transition from normal to dangerous is allowed. Danger direct jump threshold. In one implementation, the threshold value is set to 0.8, representing the impact jump threshold. In one implementation, the value is 0.7.

[0127] In one embodiment, as an example, suppose the state probability vector output by Softmax is... as follows: [Normal: 0.9, Warning: 0.1, Danger: 0.0]; [Normal: 0.1, Warning: 0.8, Danger: 0.1]; [Normal: 0.2, Warning: 0.1, Danger: 0.7]; If monotonic constraints are not used, the decoding result based on the maximum probability is "Normal -> Warning -> Danger", which conforms to the monotonic law.

[0128] In this embodiment, if we assume The probability becomes [Normal: 0.6, Warning: 0.2, Danger: 0.2]. Normal decoding will result in "Normal -> Warning -> Normal"; monotonic constraint decoding will prohibit illegal transitions from warning back to normal, thus forcing the search for a legal path, and the final result may be "Normal -> Warning -> Warning". In this embodiment, if... The probability is [Normal: 0.0, Warning: 0.1, Danger: 0.9], and the condition for a direct jump to danger is met at this time (danger probability). Impact strength Decoding will allow from Normally, you can directly jump to... The danger is reflected in the result of "normal -> normal -> danger", enabling a rapid response to sudden dangers.

[0129] 5> Furthermore, unlike existing technologies, this embodiment of the invention also uses dynamic programming to search for the maximum probability state path that satisfies monotonic constraints, in order to automatically avoid illegal transitions and find the legal state path with the highest global score. Specifically, for each time step and each state category, the path score is accumulated. The path score is obtained by adding the logarithm of the state probability at the current time step to the legal state path score at the previous time step. If a state transition is illegal, the score of that transition is set to a minimum value and does not participate in the optimal path selection; traversing to the ... After a time step, backtracking begins from the state with the highest path score to obtain the th... A monotonic constraint state sequence In one embodiment, for example, suppose there are 3 time steps ( ), 3 state categories ( ): Normal (1), Warning (2), Danger (3), the probability vectors for each time step are known as follows: , , ,and If the dangerous direct jump condition is not met, the legal transition rules are: 1→1, 1→2, 2→2, 2→3, 3→3 are allowed; 2→1, 3→2, 3→1 are prohibited. The dynamic programming process is as follows: a> The path score for each state is the logarithm of its probability. , , ; b> : Calculate the maximum score to reach each state; for example, there are 2 paths to state 2: from It comes from or from state 1. It comes from state 2; take the higher score of the two: ; The only path to state 1 is from... The transition comes from state 1; the path to state 3 is from... It transitions from state 2 or state 3; c> Similarly, calculate the maximum score to reach each state; for example, the only legal path to state 1 is from... The transition comes from state 1; the legal path to state 2 is from... The transition is from state 1 or state 2; calculate and compare scores; the legal path to state 3 is from... It transitions from state 2 or state 3.

[0130] d> Retrospection: In At any given moment, select the state with the highest score, let's say state 2; backtrack forward to find the state at the previous moment that led to that maximum score, and so on. , The state constitutes the optimal path.

[0131] It should be noted that the state classification network adopts an architecture of "parallel encoding - importance enhancement - time compression - sequence decoding". First, the input tensor... The data are fed into two parallel branches: the time-frequency coding branch uses multi-scale one-dimensional convolution to extract impact vibration features. The channel feature projection branch uses linear projection to extract channel coupling features. Then, by and Calculate the channel importance mask, for Enhancement is performed to obtain Then, based on the shock activation sequence... and Performing impact-driven time compression yields shorter compressed feature sequences. Finally, The input gated recurrent unit models the time-series dependency, and through time backfill decoding and monotonic constraint decoding layers, it outputs the final state sequence at the original time resolution. It should also be noted that the embodiments of the present invention combine data-driven state probability output with the monotonic law of engineering risk. The gated loop unit is used to learn the continuity of the impact response over time, the time backfill decoding is used to restore the original time resolution, and the monotonic constraint decoding is used to eliminate state bounces that do not conform to the evolution law of hook risk, while retaining the ability to respond quickly to sudden dangers.

[0132] Furthermore, this embodiment of the invention also analyzes the instantaneous frequency change exponent and impact weight of the acceleration Z-channel, referring to... Figure 4 The diagram shows the change of the instantaneous time-frequency ridge value and the instantaneous frequency mutation index corresponding to the impact weight with time step. Figure 4 It consists of three subplots, sharing the same horizontal time step (unit: sampling point). Figure 4In the diagram, (a) shows the instantaneous frequency ridge value of the acceleration Z channel, in Hz, indicating a sharp frequency jump from approximately 3Hz to 22Hz at the moment of impact. (b) compares the original instantaneous frequency jump index (red) with the jump index after 3-point median filtering (black), overlaid with the 80th percentile jump threshold (gray dashed line). In b, isolated spikes in the original signal (such as the glitch in step 42) are successfully smoothed, while the actual impact location still maintains a high jump value. (c) shows the impact weight (dimensionless) calculated based on the portion exceeding the threshold, with non-zero weights strictly concentrated near the impact location. Experimental results show that frequency jump detection has high sensitivity and a certain degree of noise resistance for impact localization, demonstrating that the dual judgment mechanism of "main frequency jump + amplitude jump" can effectively distinguish between real impacts and random noise.

[0133] Furthermore, embodiments of the present invention also include a parameter determination step for the prediction model, so as to construct the final model using the optimal parameters. Specifically, frame-level state label sequences can be used to supervise the state classification network during the training phase, enabling the state classification network to simultaneously learn sensor impact characteristics, cross-channel coupling characteristics, and state evolution patterns. In one implementation, a class-weighted training and validation set optimal model preservation strategy is employed.

[0134] First, the median of each sensor channel can be calculated on the training set. and channel interquartile range These parameters are fixed and used as robust normalized parameters for the training, validation, and online inference phases. The current sample statistics must not be reused during the validation and online inference phases. and To avoid data leakage and online scale drift.

[0135] Furthermore, initialize the parameters of the state classification network. The multi-scale one-dimensional convolutional kernels of the time-frequency coding branch are initialized according to the trends of short-time difference, medium-time oscillation, and long-time smoothing; the channel feature projection layer, gated recurrent unit, and multilayer perceptron classifier can be initialized using He; the linear transformation parameters in the channel importance mask can be initialized with small random numbers to make the channel enhancement coefficient close to 1 in the early stage of training, so as to avoid over-amplifying some channels at the beginning of training.

[0136] Furthermore, the multi-sensor time window matrix of each original smart hook in the training set can be sequentially input into the steps of the above-described embodiments of the invention to obtain the corresponding state probability sequence. For the first... There are training samples, and the state probability sequence is... The frame-level state tag sequence is .

[0137] Furthermore, class weights are calculated based on the number of frame-level samples for each state category in the training set, and then class-weighted cross-entropy loss is used to train the state classification network. The number of frame-level samples for each state category is denoted as . ,in, This represents the state category index, with values ​​ranging from 1 to... The maximum number of frame-level samples across all state categories is denoted as . ;No. The class weights of each state category are denoted as: It is acceptable The fewer the number of state categories at the frame level, the larger their corresponding category weights. For each time step, the cross-entropy loss at that time step can be amplified based on the category weights corresponding to the true state categories, giving higher attention to warning and dangerous states during backpropagation. If dangerous state samples are extremely few, an upper limit can be set on the dangerous state category weights, for example, not exceeding 10 times the weights of normal states, to avoid excessive oscillations during training.

[0138] It should be noted that this invention employs a standard class-weighted cross-entropy loss. Specifically, when calculating the cross-entropy loss for each sample, it is multiplied by a weight coefficient corresponding to the sample's true class to alleviate the class imbalance problem. Optionally, a lightweight temporal smoothing loss can be added in addition to the class-weighted cross-entropy loss. The temporal smoothing loss is used to suppress drastic and meaningless jittering of state probabilities between adjacent time steps, but does not restrict rapid changes at the actual shock. In specific implementation, the difference in adjacent state probabilities is constrained only between stationary time steps where the shock activation intensity is below 0.2, and this constraint is not applied to time steps with high shock activation intensity. Based on this, the sensitivity to shock state changes can be maintained while reducing output jitter in the stationary phase.

[0139] In one implementation, the time smoothing loss only occurs in... and The intensity of the impact activation is always lower than Only when the probability difference between the two is considered is the penalty imposed, thereby suppressing meaningless fluctuations in the steady phase without affecting the rapid changes during the impact phase. The calculation method is expressed as follows:

[0140] in, This indicates an indicator function that takes the value when the condition inside the parentheses is true. Otherwise ; Represents the square of the L2 norm; and They represent the first The training sample at the th ... Time step and the The overall impact intensity at time step is determined by the first... A shock activation sequence is given; and They represent the first The training sample at the th ... and the The state probability vector at each time step.

[0141] Based on this, the total loss function Class-weighted cross-entropy loss and time smoothing loss The weighted summation is expressed as:

[0142] in, Indicates the first Class-weighted cross-entropy loss for each training sample; Indicates the first Temporal smoothing loss for each training sample; This represents the time smoothing loss coefficient, used to balance the main classification loss and the smoothing constraint. In one implementation, it is chosen as... .

[0143] In this process, an adaptive moment estimation optimizer can be used to update the parameters of the state classification network, and the initial learning rate can be set to... The batch size can be 16 or 32, and the total number of training epochs can be 100 to 200. If the total loss on the validation set does not decrease for several consecutive training epochs, the learning rate is multiplied by 0.5, or training is stopped early. Furthermore, after each training epoch, the state classification network is evaluated using the validation set. Validation metrics include the F1 score for the three states (normal, warning, and dangerous), the macro-average F1 score, and the recall rate for dangerous states. The state classification network parameters that achieve the highest macro-average F1 score and meet the preset recall rate for dangerous states are saved as the final deployment model.

[0144] In summary, compared with the prior art, the embodiments of the present invention are innovative in the following aspects: 1. A sudden change detection method that integrates instantaneous frequency jumps and amplitude jumps is adopted. Instantaneous frequency ridges are extracted through multi-band analytical filtering, and the product of the normalized frequency jump value and the amplitude jump value is calculated as the instantaneous frequency change index. This enables precise positioning of brief impacts such as lifting off the ground and emergency braking, while automatically suppressing false triggering of gradual changes in ambient temperature and other channels.

[0145] 2. An impact-weighted local cross-channel dynamic correlation mechanism is adopted. The short-term correlation strength between sensor channels is calculated only in the local time neighborhood at the moment of impact, using the catastrophe exponent to generate weights. This enables the model to capture the multi-sensor linkage coupling relationship at the moment of impact, avoiding dilution by a large number of stationary time steps in the global time window.

[0146] 3. An adaptive enhancement strategy guided by channel importance masks is adopted. Based on the temporal fluctuation of each channel and the cross-channel coupling score, channel importance masks are dynamically generated to differentially enhance or compress time-frequency features, enabling the model to automatically focus on the key sensors of the current stage according to different operating conditions.

[0147] 4. A sequence processing method combining impact-driven time compression and monotonic constraint state decoding is adopted. The time steps are weighted and pooled with the impact activation intensity as the weight, which shortens the sequence length while retaining the impact peak and decay pattern; then the monotonic evolution law of risk is embedded into the dynamic programming decoding process to prevent state back-jumps while retaining the ability to prevent dangerous straight jumps under strong evidence.

[0148] Based on the above embodiments, this invention also provides a multi-data fusion and monitoring status assessment device for smart hooks, referring to... Figure 5 The device includes: a data processing module 10, used to acquire multi-sensor monitoring data of the intelligent hook during the lifting process, and determine the impact weight of the corresponding monitoring channel data based on the instantaneous frequency mutation index of each monitoring channel data in the multi-sensor monitoring data; an association module 20, used to perform dynamic impact association of each monitoring channel data in a time window according to the short-time correlation strength of the impact weight of each monitoring channel data in the local time neighborhood, so as to construct a time-series feature containing the original signal and cross-channel coupling relationship; an enhancement module 30, used to extract time-frequency features and channel features from the time-series feature respectively, generate a mask according to the fluctuation degree of the channel feature and the degree of cross-channel coupling, and use the mask to adaptively enhance the time-frequency feature to obtain an enhanced time-frequency feature representation; and an execution module 40, used to perform monotonic constraint discrimination on the enhanced time-frequency feature representation, and determine the state prediction result of the intelligent hook during the lifting process according to the corresponding discrimination result. It should be noted that the implementation principle and technical effects of the multi-data fusion and monitoring state evaluation device for intelligent hooks provided in this embodiment of the invention are the same as those in 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.

[0149] 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 1The 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 6 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.

[0150] exist Figure 6 In 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 6The diagram uses only a single double-headed 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.

[0151] The present invention provides a computer program product for a method and apparatus for multi-data fusion and monitoring status assessment of intelligent hooks, comprising 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. Specific implementations can be found in the method embodiments and will not be repeated here. Those skilled in the art will understand that, for convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, and will not be repeated here. If the function is implemented as 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 the present 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, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] 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 method for multi-data fusion and monitoring status assessment of smart hooks, characterized in that, The method includes: Acquire multi-sensor monitoring data of the smart hook during the lifting process, and determine the impact weight of the corresponding monitoring channel data based on the instantaneous frequency change index of each monitoring channel data in the multi-sensor monitoring data; Based on the short-term correlation strength of the impact weight of each monitoring channel data in the local time neighborhood, dynamic impact correlation is performed on each monitoring channel data in the time window to construct a time series feature containing the original signal and cross-channel coupling relationship; Time-frequency features and channel features are extracted from the time-series features respectively. A mask is generated based on the fluctuation degree and cross-channel coupling degree of the channel features. The mask is then used to adaptively enhance the time-frequency features to obtain an enhanced time-frequency feature representation. The enhanced time-frequency feature representation is subjected to monotonic constraint discrimination, and the state prediction result of the smart hook during the lifting process is determined based on the corresponding discrimination result.

2. The method according to claim 1, characterized in that, The step of determining the impact weight of the corresponding monitoring channel data based on the instantaneous frequency mutation index of each monitoring channel data in the multi-sensor monitoring data includes: Based on the phase change of the dominant frequency band of each monitoring channel data in the multi-sensor monitoring data, the instantaneous frequency ridge value of the corresponding monitoring channel data is determined; Based on the discrete scale of the instantaneous frequency ridge value, and the frequency jump value and amplitude jump value in adjacent time steps, the instantaneous frequency mutation index of the monitoring channel data is determined; The time step mutation screening threshold is determined based on the instantaneous frequency mutation index; Based on the impact characteristics represented by the instantaneous frequency mutation index and the time step mutation screening threshold, the impact weight of the corresponding monitoring channel data is determined.

3. The method according to claim 2, characterized in that, The method further includes: Based on the median and interquartile range of each monitoring channel in the multi-sensor monitoring data, robust standardization is performed on the corresponding monitoring channel data. Multi-band analytical filtering is performed on the standardized monitoring channel data to determine the dominant frequency band in the monitoring channel data.

4. The method according to claim 1, characterized in that, Based on the short-time correlation strength of the impact weight of each monitoring channel data in the local time neighborhood, the step of dynamically associating the impacts of each monitoring channel data within a time window to construct time-series features containing the original signal and cross-channel coupling relationships includes: Based on the impact weight, determine the short-term correlation strength between any two monitoring channels in the local time neighborhood; Based on the short-time correlation intensity, the impact coupling result of the corresponding local time neighborhood is determined; By traversing all time steps, the impact dynamic correlation tensor of the sensor monitoring data at different time steps is obtained; The time-series features are obtained by combining the impact dynamic correlation tensor with the multi-sensor monitoring data.

5. The method according to claim 4, characterized in that, The step of determining the short-time correlation strength between any two monitoring channels based on the impact weight includes: Based on the impact weights of data from multiple monitoring channels in the same local time neighborhood, the joint weights of the corresponding local time neighborhoods are determined. Based on the joint weights and the local mean-reduced sampled values ​​of the corresponding monitoring channel data, the weighted synchronous change is determined; The weighted synchronous change is normalized to determine the short-term correlation strength.

6. The method according to claim 1, characterized in that, The steps of performing monotonic constraint discrimination on the enhanced time-frequency feature representation and determining the state prediction result of the smart hook during the lifting process based on the corresponding discrimination result include: Based on the impact weights, the enhanced time-frequency feature representation is weighted and compressed by pooling to obtain a compressed feature sequence; Temporal dependency modeling is performed on the compressed feature sequence to obtain the hidden state sequence of the compressed time step; The hidden state sequence is restored to the original temporal resolution to obtain the state probability distribution for each original time step; A monotonic constraint is applied to the state probability distribution to prevent state reversal, thereby obtaining the state prediction result of the smart hook at each time step in the hoisting process.

7. The method according to claim 6, characterized in that, The step of performing weighted pooling compression on the enhanced time-frequency feature representation based on the impact weights to obtain a compressed feature sequence includes: The impact activation intensity at each time step is determined based on the overall intensity of the impact weight at each time step. Using the impact activation intensity as the aggregation weight, the enhanced time-frequency feature representation is weighted and aggregated in the time dimension to determine the compressed feature sequence.

8. A multi-data fusion and monitoring status assessment device for intelligent hooks, characterized in that, The device includes: The data processing module is used to acquire multi-sensor monitoring data of the smart hook during the hoisting process, and determine the impact weight of the corresponding monitoring channel data based on the instantaneous frequency change index of each monitoring channel data in the multi-sensor monitoring data. The correlation module is used to perform dynamic correlation of each monitoring channel data in a time window based on the short-term correlation strength of the impact weight of each monitoring channel data in the local time neighborhood, so as to construct a time series feature containing the original signal and cross-channel coupling relationship; An enhancement module is used to extract time-frequency features and channel features from the time-series features, generate a mask based on the fluctuation degree and cross-channel coupling degree of the channel features, and use the mask to adaptively enhance the time-frequency features to obtain an enhanced time-frequency feature representation. The execution module is used to perform monotonic constraint discrimination on the enhanced time-frequency feature representation, and determine the state prediction result of the smart hook in the hoisting process based on the corresponding discrimination result.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the multi-data fusion and monitoring status assessment method for smart hooks as described in any one of claims 1 to 7.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the multi-data fusion and monitoring status assessment method for intelligent hooks as described in any one of claims 1 to 7.

Images