Stacked carton fatigue crush online detection method, device, system, and medium

Abstract

The application relates to a stacked carton fatigue crushing online detection method, device, system and medium, the method comprising: obtaining vibration data of each target object corresponding to an acceleration time sequence; preprocessing the acceleration time sequence to generate a first acceleration time sequence corresponding to each target object, performing frequency response feature extraction and difference processing on the first acceleration time sequence to generate real-time difference frequency response data corresponding to each carton; inputting the real-time difference frequency response data into an energy modal prediction model to obtain corresponding expected frequency response data, determining a total prediction residual based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data; judging whether an EWMA value corresponding to the total prediction residual is greater than a preset anti-shake control feature energy threshold value, and determining a stacked carton crushing result according to the judgment result. Through the application, the problem that a scheme for detecting a stacked carton crushing in transportation in the related art is prone to frequent false positives or false negatives is solved.

Description

Technical Field

[0001] This application relates to the fields of intelligent manufacturing and production scheduling technology, and in particular to online detection methods, devices, systems and media for fatigue crushing of stacked cartons. Background Technology

[0002] In modern logistics, multimodal transport, including road and rail, is the primary mode of goods circulation. To maximize the space utilization of transport vehicles, packaging cartons are typically stacked tightly and bound with film or strapping to create a closed, multi-degree-of-freedom mechanical system. Under the broadband random vibration excitation from the transport vehicle chassis, the stacked cartons undergo irreversible fatigue evolution of the internal corrugated core layer under prolonged alternating stress, progressing from microfiber fracture to local buckling and macroscopic stiffness degradation. Once the bottom layer of cartons yields, a cascading collapse effect can easily occur, causing the entire stacked structure to become instantly unstable, resulting in cargo damage and enormous reverse logistics costs.

[0003] In related technologies, offline post-evaluation methods such as static compression tests before leaving the factory or standard vibration damage tests in the laboratory are often used to detect carton crushing. In-transit detection is also carried out using machine vision systems with external cameras or by placing acceleration or microfluidic color-changing sensors inside the packaging box. However, existing offline post-evaluation methods sever the physical process of continuous damage evolution and cannot provide real-time online early warning during transportation. At the same time, existing in-transit detection cannot detect the hidden yielding and minute deformation of the core pressure-bearing parts at the bottom layer, nor can it separate the "sudden increase in road surface excitation energy" from the "degradation of the carton's structural stiffness," which can easily lead to frequent false alarms or serious omissions.

[0004] Currently, there is no effective solution to the problem that the existing technologies for detecting crushing of stacked cartons during transportation are prone to frequent false alarms or missed alarms. Summary of the Invention

[0005] This application provides an online method, apparatus, system, and medium for detecting fatigue crushing of stacked cartons, which at least solves the problem that related technologies for detecting crushing of stacked cartons during transportation are prone to frequent false alarms or missed alarms.

[0006] In a first aspect, embodiments of this application provide an online fatigue crush detection method for stacked cardboard boxes, comprising: acquiring vibration data corresponding to each target object, wherein the target object includes a support platform supporting the stacked cardboard boxes and each layer of cardboard boxes, the vibration data including an acceleration time series corresponding to each target object, the acceleration time series being used to characterize the comprehensive vibration response of each target object at the current moment; preprocessing the acceleration time series to generate a first acceleration time series corresponding to each target object, and performing frequency response feature extraction and differential processing on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked cardboard box, wherein the frequency response feature is based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the first acceleration time series corresponding to the cardboard box. The frequency domain cross power spectrum corresponding to the sequence is determined; the real-time difference frequency response data is input into the trained energy mode prediction model to obtain the expected frequency response data corresponding to the real-time difference frequency response data, and the total prediction residual is determined based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data. The expected frequency response data is used to characterize the vibration energy corresponding to each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning training on the constructed initial hidden Markov state space (HMMSSM) model; it is determined whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control feature energy threshold, and the crushing result of the stacked cartons is determined based on the judgment result.

[0007] Secondly, embodiments of this application provide an online fatigue crush detection device for stacked cartons, comprising: The acquisition module is used to acquire vibration data corresponding to each target object. The target object includes a support platform for stacking cardboard boxes and each layer of cardboard boxes. The vibration data includes an acceleration time series corresponding to each target object. The acceleration time series is used to characterize the comprehensive vibration response of each target object at the current moment. The processing module is used to preprocess the acceleration time series to generate a first acceleration time series corresponding to each target object, and to extract frequency response features and perform differential processing on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked carton. The frequency response features are determined based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the carton. The prediction module is used to input the real-time difference frequency response data into the trained energy mode prediction model to obtain the expected frequency response data corresponding to the real-time difference frequency response data, and to determine the total prediction residual based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data. The expected frequency response data is used to characterize the vibration energy of each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning and training on the constructed initial hidden Markov state space (HMMSSM) model. The judgment module is used to determine whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and to determine the crushing result of the stacked cartons based on the judgment result.

[0008] Thirdly, embodiments of this application provide an online fatigue crush detection system for stacked cartons, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the online fatigue crush detection method for stacked cartons described in the first aspect.

[0009] Fourthly, embodiments of this application provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the online fatigue crush detection method for stacked cartons described in the first aspect.

[0010] Compared to related technologies, the online fatigue crush detection method, device, system, and medium for stacked cardboard boxes provided in this application embodiment acquire vibration data corresponding to each target object. The target object includes a support platform carrying the stacked cardboard boxes and each layer of cardboard boxes. The vibration data includes an acceleration time series corresponding to each target object, which characterizes the comprehensive vibration response of each target object at the current moment. The acceleration time series is preprocessed to generate a first acceleration time series corresponding to each target object. Frequency response features are extracted and differentially processed from the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked cardboard box. The frequency response features are determined based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the cardboard box. The real-time differential frequency response data is input into a trained energy mode prediction model to obtain expected frequency response data corresponding to the real-time differential frequency response data, and based on the real-time differential frequency response data… The similarity between the frequency response data and the corresponding expected frequency response data is used to determine the total prediction residual. The expected frequency response data characterizes the current vibration energy of each carton. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy modal prediction model is generated by unsupervised parameter calibration training of the constructed initial Hidden Markov State Space (HMMSSM) model using the Expectation-Maximization (EM) algorithm. The system determines whether the EWMA value corresponding to the total prediction residual is greater than a preset anti-shake control feature energy threshold, and based on the determination result, determines the crushing result of the stacked cartons. This addresses the problem of frequent false alarms or missed alarms in related technologies for detecting crushing of stacked cartons during transportation. It preprocesses the collected acceleration time series and calculates the corresponding frequency response function to eliminate external road condition interference. The system constructs an energy modal prediction model to predict the expected frequency response data of the multimodal modes corresponding to the current moment, and then, based on the anti-shake control feature energy threshold dynamically generated using the EWMA algorithm, achieves accurate early warning before macroscopic cascading collapses.

[0011] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a hardware structure block diagram of the terminal of the online fatigue crush detection method for stacked cartons according to an embodiment of this application; Figure 2 This is a flowchart of an online fatigue crush detection method for stacked cartons according to an embodiment of this application; Figure 3 This is a structural block diagram of an online fatigue crush detection device for stacked cartons according to an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application. Furthermore, it is understood that although the effort involved in such development may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, modifications to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0014] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0015] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "a," "an," "an," "the," and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. "Multiple stages" used in this application refers to two or more stages. "And / or" describes the relationship between related objects, indicating that three relationships may exist; for example, "A and / or B" can represent: A alone, A and B simultaneously, and B alone. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of objects.

[0016] The embodiments of this application are described below with reference to the accompanying drawings:

[0017] The method embodiments provided in this example can be executed on a terminal, computer, or similar computing device. Taking running on a terminal as an example, Figure 1 This is a hardware structure block diagram of the terminal of the online fatigue crush detection method for stacked cardboard boxes according to an embodiment of this application. Figure 1 As shown, a terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. Optionally, the terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the terminal described above. For example, the terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0018] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the online fatigue crush detection method for stacked cartons in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0019] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0020] This embodiment provides an online fatigue crush detection method for stacked cardboard boxes operating on the aforementioned terminal. Figure 2 This is a flowchart of an online fatigue crush detection method for stacked cartons according to an embodiment of this application, as shown below. Figure 2 As shown, the process includes the following steps:

[0021] Step S201: Obtain vibration data corresponding to each target object. The target objects include the support platform for stacking cardboard boxes and each layer of cardboard boxes. The vibration data includes the acceleration time series corresponding to each target object. The acceleration time series is used to characterize the comprehensive vibration response of each target object at the current moment.

[0022] In this embodiment, vibration data is acquired under laboratory conditions. Non-invasive triaxial accelerometers are placed inside the force transmission nodes and on the bottom support platform of each physical layer (n layers in total) of the sealed stacked cardboard box assembly to collect data simulating the alternating stress state of the stacked cardboard boxes under transportation conditions. The corresponding acceleration is used to characterize the corresponding stress state. In the simulated laboratory environment, accelerometers are installed on multiple stacked cardboard boxes and on the support platform. Each accelerometer is connected to a DASP data acquisition instrument, which collects and preprocesses vibration signals. During simulated online testing, simulated vibration data is input to the hydraulic vibration table of the support platform during simulated transportation. After the vibration signal of the road surface vibration is collected, the comprehensive vibration signal generated by the interaction of the cardboard boxes is collected by the acceleration sensor, and the vibration process of the multi-layer cardboard boxes is recorded by the high-speed camera to observe the deformation of the cardboard boxes. In this embodiment, multiple acceleration data are collected in a continuous time period to obtain the corresponding acceleration time series. Each acceleration time series corresponds to a target object. That is, the active acceleration time series includes at least the acceleration time series of the target objects. It can be understood that each target object is associated with at least one acceleration time series. The number of acceleration time series associated with each target object is related to the setting of the acceleration sensors set around each target object.

[0023] In this embodiment, it is understood that during transportation, both the cardboard boxes and the support platform will vibrate. Normal vibration will not produce abrupt vibration excitation, meaning the corresponding vibration will not cause the stacked cardboard boxes to collapse. However, abrupt changes in the transportation conditions (e.g., potholes or abnormal bumps) will generate abnormal vibrations. Under prolonged alternating stress, the corrugated core layer inside the cardboard boxes will undergo irreversible fatigue evolution, progressing from microfiber fracture to local buckling to macroscopic stiffness degradation. When the cardboard boxes yield (usually starting with the bottom layer), a cascading collapse effect will occur, causing the entire stacked structure to become instantly unstable and collapse. In this embodiment, during the vibration of the cardboard boxes and the support platform, they will interact, meaning that the acceleration corresponding to each target object has two... The factors affecting this are: first, the excitation generated by the vibration of the target object itself; and second, the mutual excitation generated by the vibration impact of other target objects located above and below the target object. For example, the gravity of other target objects located above the target object and the impact caused by the vibration of other target objects above the target object are coupled to obtain the upper mutual excitation; the upward impact of the vibration of other target objects located below the target object and the impact corresponding to the reaction force generated by the target object and other target objects pressing on other target objects below the target object are coupled to obtain the lower mutual excitation. The upper mutual excitation and the lower mutual excitation are coupled to form the corresponding mutual excitation. Combined with the excitation generated by the target object's own vibration, the two excitations are coupled to form a comprehensive vibration response, which is converted into the acceleration of the corresponding target object.

[0024] Step S202: Preprocess the acceleration time series to generate a first acceleration time series corresponding to each target object, and extract frequency response features and perform differential processing on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked carton. The frequency response features are determined based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the carton.

[0025] In this embodiment, to generate a stable signal, the acceleration time series corresponding to each target object is preprocessed, including detrending to eliminate baseline drift in the acceleration time series and data filtering. In this embodiment, the fluctuations or non-stationary fluctuations of the complex transportation environment road surface, as well as white noise from interlayer contact, are effectively removed. The first acceleration time series is processed using the corresponding frequency response function, that is, the corresponding frequency response function value is calculated based on the frequency domain cross-power spectrum of the first acceleration time series corresponding to each cardboard box and the frequency domain auto-power spectrum of the first acceleration time series corresponding to the bearing platform. The frequency power spectrum corresponding to the bearing platform characterizes the excitation generated by the bearing platform under the vibration signal at the corresponding frequency at the current moment, and is the statistical mean of the bearing platform's own energy density under the excitation of the corresponding vibration signal. In this embodiment, in order to make the frequency response characteristics tend to the true frequency response, the obtained frequency response characteristics (the frequency response function value corresponding to the frequency response function) are differentiated from the set health state benchmark (also a benchmark frequency response function value) (the difference between the frequency response function value and the benchmark frequency response function value) to obtain the corresponding real-time difference frequency response data (corresponding to a difference signal of stiffness degradation) to measure the stiffness change of the true frequency response characteristics.

[0026] In this embodiment, the use of frequency response function processing can effectively remove the white noise from the road surface fluctuations and interlayer contact in complex transportation environments, thus solving the problem of frequent false alarms when using static thresholds.

[0027] Step S203: Input the real-time difference frequency response data into the trained energy mode prediction model to obtain the expected frequency response data corresponding to the real-time difference frequency response data. Based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data, determine the total prediction residual. The expected frequency response data is used to characterize the vibration energy corresponding to each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning training on the constructed initial hidden Markov state space (HMMSSM) model.

[0028] In this embodiment, the real-time difference frequency response data includes frequency response features in multiple latent states. Each frequency response feature corresponds to a vibration excitation at a specific frequency. In other words, the real-time difference frequency response data for a cardboard box at the current moment includes frequency response features corresponding to multiple latent states. For example, at the current moment, the real-time difference frequency response data for cardboard box A includes n frequency response features at a specific frequency f. n Frequency response data △H under vibration excitation nIn this embodiment, the energy modal prediction model is trained to first calculate the prior probability of at least one hidden state related to the input frequency response data (including the current real-time difference frequency response data and the frequency response data used for testing during training) based on the input frequency response data. Then, it predicts the expected frequency response corresponding to each hidden state. Finally, it obtains the expected frequency response data corresponding to the input frequency response data by weighting the prior probability of each hidden state and the corresponding expected frequency response. In this embodiment, it should be understood that for the real-time difference frequency response data input at the current time, the energy modal prediction model will predict the prior probability of the hidden state (associated with a vibration excitation of a certain frequency) corresponding to the real-time difference frequency response data based on the input real-time difference frequency response data. That is, it calculates the prior probability of each hidden state of at least one carton at the current time by forward filtering recursively, then predicts the expected frequency response of each hidden state, and then multiplies the prior probability with the expected frequency response of each hidden state to obtain the expected frequency response data of a hidden state. For example, for carton A at time 1, the energy modal prediction model will calculate the expected frequency response data of each hidden state based on its corresponding real-time difference frequency response data ΔH={ΔH}. 11 , △H 12 ..., △H 1n First, predict the prior probability P = {P} in the corresponding hidden state. 11 P 12 ... P 1n}, and then predict the expected frequency response Y={Y} in the corresponding hidden state. 11 Y 12 ..., Y 1n}, and based on the prior probability and expected frequency response, obtain the expected frequency response data y={P} for carton A at time 1. 11 ×Y 11 P 12 ×Y 12 ... P 1n ×Y 1n}

[0029] In some alternative implementations, using real-time difference frequency response data and a trained model, the prior probability of the system being in each hidden state at the current time is calculated recursively through forward filtering according to the following formula: Among them, C js Let be the single-step probability of transitioning from state j to state s. Here, an upper triangular constraint is also required to satisfy the irreversible law. After obtaining the discrete state prior probability distribution at time t, it needs to be mapped back to the continuous observation space to synthesize the theoretically expected response of the system at the current time. The theoretical expected value of the observed characteristics at time t (corresponding to the expected frequency response data) can be obtained through probability weighting: , where Y 1:t-1 The historical observation set represents the sequence of all observation features from time 1 to t-1; Let be the prior state probability, representing the probability of predicting the current state as s before seeing the data at time t; Let be the theoretical expected observation value, representing the ideal observation vector that should appear at time t according to the model prediction; The basis function weight matrix is ​​the weight matrix of the trained state s.

[0030] In this embodiment, the expected frequency response data is a real-valued vector with the same dimension as the real-time difference frequency response data, representing the expected vibration energy distribution at each frequency point.

[0031] In this embodiment, after predicting the corresponding expected frequency response data based on the real-time difference frequency response data corresponding to each carton, the difference calculation is performed based on the real-time difference frequency response data and the corresponding expected frequency response data corresponding to all cartons. Spatial heteroscedasticity decoupling is performed by introducing the model covariance inverse matrix, the Mahalanobis distance is calculated, and a one-dimensional one-step prediction error continuous scalar sequence (OSPE) is output, which is the corresponding determined total prediction residual.

[0032] Meanwhile, it should be understood that this embodiment uses Mahalanobis distance as the corresponding similarity, and determines the corresponding residuals based on the similarity represented by Mahalanobis distance. The larger the Mahalanobis distance, the greater the difference in the corresponding residuals and the greater the total prediction residuals.

[0033] Step S204: Determine whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and determine the crushing result of the stacked carton based on the determination result.

[0034] In some embodiments, after obtaining the corresponding total prediction residual, the following steps are used to determine the EWMA value corresponding to the total prediction residual, and to dynamically generate the stabilization control feature energy threshold adapted to the current batch operating conditions: Step 24-1: Using the exponentially weighted moving average (EWMA) algorithm, the total prediction residuals are low-pass smoothed in chronological order to obtain the EWMA statistic, where the EWMA value includes the EWMA statistic.

[0035] Step 24-2: Obtain the historical total prediction residuals corresponding to multiple detection times in a non-crushable state, and determine the EWMA mean and EWMA standard deviation of the EWMA statistic corresponding to multiple historical total prediction residuals.

[0036] Step 24-3: The sum of the EWMA mean and the dynamic value of the image stabilization control feature is used as the energy threshold of the image stabilization control feature. The dynamic value of the image stabilization control feature is determined by the product of the preset control coefficient and the standard deviation of the EWMA.

[0037] In this embodiment, after obtaining the corresponding total prediction residuals, an exponentially weighted moving average is performed on the sequence of total prediction residuals in time order to obtain the EWMA value corresponding to the total prediction residuals. Specifically, the exponentially weighted moving average (EWMA) algorithm is used to perform low-pass smoothing on the total prediction residuals (corresponding to the OSPE scalar sequence), and the corresponding calculation formula is as follows: Z t =γe t +(1-γ)Z t-1 , where e t The Mahalanobis distance prediction error is calculated at time t (the current time). γ is a smoothing parameter whose value is strictly limited to the closed interval [0, 1]. Z0 represents the initial EWMA value, which is usually initialized to the mean EWMA value during the historical healthy steady state.

[0038] In this embodiment, the anti-shake control characteristic energy threshold is generated dynamically and calibrated based on historical non-crushable vibration signals. Specifically, it utilizes a short-range baseline data collected during the initial period of a single transportation task (healthy state, also non-crushable state). In this embodiment, it is stipulated that the stacked carton system should be in a healthy state in terms of physical structure at the start of each vibration test of the stacked carton system. The executing entity will automatically capture the first K segments of the initial test period. baseline For each discrete time window, the mean μ of the dynamically estimated EWMA statistic is calculated. Z and standard deviation σ Z : , Among them, Z t This represents the EWMA smoothing statistic, which accumulates the final monitoring index of historical minor damage effects, and is used to calculate the characteristic energy threshold (UCL) of the image stabilization control adapted to the current operating conditions using the following formula: UCL = μ Z +L·σ Z Where L is the control coefficient, L·σ Z This represents a tolerance, or a range of tolerance, which can be selected according to the needs. If you want the alarm to be slightly less sensitive, L should be larger; if you want the alarm to be triggered at the slightest sound, L should be smaller.

[0039] In some alternative implementations, determining the crushing result of the stacked cartons based on the judgment result includes the following steps: If the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, it is determined that the carton in the stack is crushed and a carton crushing alarm is triggered.

[0040] If the EWMA value corresponding to the total prediction residual is not greater than the preset anti-shake control characteristic energy threshold, it is determined that there is no crushing of the stacked cartons.

[0041] In this embodiment, during continuous monitoring, when the EWMA value corresponding to the total prediction residual irreversibly breaks through the dynamically generated anti-shake control characteristic energy threshold in one direction, a final-state carton crushing alarm will be triggered.

[0042] Through steps S201 to S204, vibration data corresponding to each target object is acquired. The target objects include the support platform carrying the stacked cardboard boxes and each layer of cardboard boxes. The vibration data includes acceleration time series corresponding to each target object, which characterizes the comprehensive vibration response of each target object at the current moment. The acceleration time series is preprocessed to generate a first acceleration time series corresponding to each target object. Frequency response features are extracted and differentially processed from the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked cardboard box. The frequency response features are determined based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the cardboard box. The real-time differential frequency response data is input into a trained energy mode prediction model to obtain expected frequency response data corresponding to the real-time differential frequency response data. Based on the similarity between the real-time differential frequency response data and the corresponding expected frequency response data, the expected frequency response data is determined. The total prediction residual and expected frequency response data are used to characterize the vibration energy of each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy modal prediction model is generated by unsupervised parameter calibration and training of the constructed initial Hidden Markov State Space (HMMSSM) model using the Expectation-Maximization (EM) algorithm. It is determined whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control feature energy threshold, and based on the judgment result, the crushing result of stacked cartons is determined. This solves the problem that the scheme for detecting the crushing of stacked cartons in transportation is prone to frequent false alarms or missed alarms in related technologies. The collected acceleration time series is preprocessed and the corresponding frequency response function is calculated to eliminate external road condition interference. The energy modal prediction model is constructed to predict the expected frequency response data of the multi-mode corresponding to the current moment. Then, based on the anti-shake control feature energy threshold dynamically generated by the EWMA algorithm, accurate early warning before macroscopic cascading collapse is achieved.

[0043] It should be noted that the embodiments of this application use the vibration acceleration of multi-layer stacked cardboard boxes as an external stimulus, extract frequency response features using a preset frequency response function, and utilize the frequency response function and singular value decomposition of the uncentered orthogonal stage during model training to remove non-stationary fluctuations in the road input and perform feature dimensionality reduction. This allows for unsupervised state inference and online early warning using an irreversible Hidden Markov Model that considers inter-layer propagation. The state transition probability matrix of the HMMSSM model constructed in this application is an upper triangular matrix. This state transition probability matrix is ​​not an ergodic state transition, but incorporates irreversible state constraints and inter-layer propagation constraints to realize the irreversible unidirectional state transition law. Furthermore, an inter-layer propagation coefficient matrix with lower triangular characteristics is introduced into its continuous observation equation to analyze... The unidirectional upward transmission of vibration energy from the bottom layer makes the constructed HMMSSM model more consistent with physical reality, avoiding state reversal. The inclusion of prior conditions makes the model easier to converge and fit, and also solves the "zero-sample" machine learning dilemma of not being able to obtain accurate fatigue labels in actual business. In this embodiment, the prior theoretical expectation of frequency response features is synthesized through Hidden Markov Forward Filtering, the Mahalanobis distance is calculated using the covariance inverse matrix to obtain the one-step prediction error (OPSE), and time memory smoothing is performed using the Exponential Weighted Moving Average (EWMA) algorithm. Its control threshold is generated by dynamic reconstruction calibration of the controlled baseline at the start of transportation. In this embodiment, uniform truncation and linear fitting with regularization penalties are applied to the time series before the expectation-maximization algorithm iteration.

[0044] In some embodiments, the acceleration time series is preprocessed to generate a first acceleration time series corresponding to each target object, which is achieved through the following steps: Step 21: Using the least squares method, detrendization processing is performed on the acceleration time series corresponding to each target object to generate a time-domain series.

[0045] In this embodiment, the physical layers (including the carton layer and the support platform layer) corresponding to the target object corresponding to the vibration data are set to N layers. To eliminate low-frequency baseline drift, ordinary least squares method is used for detrending processing. Let the original discrete acceleration sequence containing drift corresponding to the vibration data (corresponding to the acceleration time sequence at the current moment) be x[n] (n=0, 1, ..., N-1). Assuming that the baseline drift follows a linear trend p[n]=c1n+c0, the trend line is fitted by minimizing the residual sum of squares J(c1, c0) using the ordinary least squares method: After solving for the optimal fitting coefficients, they are separated from the original observations to obtain the corresponding time-domain sequence (corresponding to the detrended acceleration sequence). ,in, Corrected time-domain sequence The signal is strictly constrained to be a stationary signal with a constant mean of zero, thus eliminating low-frequency artifact interference.

[0046] In this embodiment, n is a scalar representing the total number of physical layers corresponding to each target object. The bottom layer is the first layer, represented by n=0, which is also the support platform layer. The top layer is the nth layer, represented by n=N-1. x[n] represents the original discrete acceleration sequence, corresponding to the acceleration value of the unprocessed nth sampling point collected by the accelerometer. N is a scalar representing the total number of sampling points, which also represents the number of discrete data points collected within a time window. p[n] is the linear trend term, representing the low-frequency baseline drift signal fitted due to temperature drift or gravity factors of the accelerometer. c1 represents the slope coefficient of the trend line, representing the growth rate of the baseline drift. c0 is the intercept coefficient of the trend line, representing the initial bias of the baseline drift. J(c1, c0) is the residual sum of squares, a loss function used to measure the deviation between the fitted trend line and the real data. This represents the detrended acceleration sequence.

[0047] Step 22: Using a preset target digital low-pass filter, the time-domain sequence is subjected to zero-phase forward and reverse digital filtering to obtain the first acceleration time sequence. The target digital low-pass filter includes one of the following: Butterworth digital low-pass filter or Gaussian low-pass filter.

[0048] In this embodiment, the time-domain sequence is then... A Butterworth digital low-pass filter is used, with the filter order set to M=6 and the cutoff frequency to ω. c =100Hz. To eliminate the inherent nonlinear phase delay of conventional infinite impulse response (IIR) filters, which would disrupt the time correspondence between the transmission responses of multiple cardboard layers, zero-phase forward and reverse digital filtering is employed; where M represents the filter order, ω c This indicates the highest frequency that the low-pass filter allows to pass through, which is also the cutoff frequency.

[0049] Through steps 21 to 22 above, the original discrete acceleration time series is preprocessed to generate a stationary detrended acceleration series (corresponding to the first acceleration time series).

[0050] In some embodiments, frequency response feature extraction and differential processing are performed on the first acceleration time series to generate real-time difference frequency response data corresponding to each stacked carton, which is achieved through the following steps: Step 31: Perform time-frequency conversion on the first acceleration time series to obtain frequency domain power spectrum data corresponding to each target object. The frequency domain power spectrum data is used to characterize the power spectrum generated in each time window associated with the frequency domain power spectrum data after the external excitation of the preset frequency is input. The frequency domain power spectrum data includes the external excitation power spectrum and the response power spectrum.

[0051] Step 32: Based on the external excitation power spectrum and the response power spectrum, determine the frequency domain self-power spectrum corresponding to the carrier platform and the frequency domain cross-power spectrum generated by the carrier platform for each layer of cardboard. Based on the ratio of the frequency domain cross-power spectrum to the frequency domain self-power spectrum, generate the frequency response function corresponding to each cardboard box. The frequency response function is used to characterize the frequency response features corresponding to the first acceleration time series.

[0052] Step 33: Linearize the frequency response function corresponding to each carton, and perform differential processing on the full life cycle frequency response data obtained by linearization and the preset reference frequency response data to generate the corresponding real-time difference frequency response data. The linearization process includes logarithmic processing.

[0053] In this embodiment, the preprocessed multi-channel acceleration sequence (corresponding to the first acceleration time sequence) is transformed into K time windows using a Short Time Fourier Transform (STFT) with a Hanning Window. Simultaneously, the time-domain signal is converted into a frequency-domain power spectrum, yielding the power spectrum Y(f,t) of the external excitation at frequency f in the t-th time window and the power spectrum X of the response at each layer. (i) (f,t), calculate the expected power spectrum G of the bottom layer input respectively. xx (f,t)=|X(f,t)|², and the cross-power spectrum between the bottom layer input and the i-th layer response. (X) * (representing a conjugate complex number), using the calculated power spectrum Y(f,t), the frequency response function is calculated, which is defined as the ratio of the expectation of the cross power spectrum to the expectation of the self power spectrum: Next, it is mapped to a logarithmic scale (decibels), which is a linearization process to linearize the energy decay: Simultaneously extract the top K systems that are in good condition. base Calculate the expected health baseline frequency response within a time window. : Finally, the obtained full lifecycle frequency response data H dB,i (f,t) and reference frequency response data The difference signal characterizing stiffness degradation is obtained by subtraction (corresponding to real-time difference frequency response data). Where K is the total number of time windows, representing the number of segments into which the total duration is divided by the Hanning window using the short-time Fourier transform; t is the index of the current time window, t∈{1,2,…,K}; f is the frequency point, representing the discrete frequency value in the frequency domain analysis; Y(f,t) is the power spectrum of the excitation source, representing the energy distribution of the vibration input to the platform at frequency f in time period t; X （i） (f,t) is the power spectrum of the i-th layer response, representing the energy distribution of the i-th layer of the cardboard box at frequency f during time period t; G xx (f,t) is the expected value of the input power spectrum, which represents the statistical mean of the energy density of the underlying excitation signal. Let X be the cross-power spectrum between the input and the i-th layer, representing the frequency correlation between the vibration of the bearing platform and the response of the i-th layer; * (f,t) represents the complex conjugate of the bottom-level input power spectrum; H 1,i (f) is the frequency response function of the i-th layer; H dB,i (f,t) represents the logarithmic frequency response in decibels, indicating that the amplitude of the frequency response function is taken as the logarithm, making the energy decay change more linearly visible; K base The number of health baseline windows represents the number of early windows used to calculate the health template when the detection has just started and the structure is intact; Let H be the reference frequency response for the i-th layer, representing the average transmission characteristics measured when the carton is brand new and intact; ΔH i (f,t) is the frequency response difference signal of the i-th layer, which represents the frequency response at the current moment minus the healthy baseline. A larger value indicates that the stiffness degradation is more severe. This is a mathematical expectation operator used to average multiple time windows.

[0054] Through steps 31 to 33 above, the preprocessed first acceleration time series is converted into time and frequency data to generate real-time difference frequency response data for adapting to the energy mode prediction model.

[0055] In some embodiments, the total prediction residual is determined based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data, through the following steps: Step 41: Using the preset covariance inverse matrix, spatial heteroscedasticity decoupling is performed on the real-time difference frequency response data and the expected frequency response data to generate standard difference frequency response data and standard expected frequency response data respectively.

[0056] In this embodiment, the real-time difference frequency response data obtained at the current moment is differentially analyzed with the expected frequency response data obtained by processing using the energy mode prediction model. Spatial heteroscedasticity is decoupled by introducing the model covariance inverse matrix to obtain standard difference frequency response data and standard expected frequency response data.

[0057] Step 42: Calculate the Mahalanobis distance between the standard difference frequency response data and the standard expected frequency response data, and average the generated Mahalanobis distance to generate the one-step prediction residual OSPE, wherein the total prediction residual includes the one-step prediction residual OSPE.

[0058] In this embodiment, the Mahalanobis distance is calculated using the following formula, and a one-dimensional continuous scalar sequence of one-step prediction error (OSPE) is output: Where O is the total number of discrete frequency points; y t,ω This represents a slice of the actual observed column vector at time t and frequency ω. This is a slice of the column vector of the theoretical prediction at time t and frequency ω at frequency ω. The inverse covariance matrix obtained by fitting the offline data to the model.

[0059] Through steps 41 to 42 above, the total prediction residual is calculated, providing a data basis for subsequent determination of whether the image stabilization control feature energy threshold has been exceeded and crushing has been detected.

[0060] In some embodiments, generating an energy mode prediction model includes the following steps: Step 51: Construct a frequency response difference matrix based on the real-time difference frequency response data corresponding to the preset sample vibration data, perform centerless truncated singular value decomposition on the frequency response difference matrix to extract a preset number of orthogonal basis functions, and perform sign alignment constraint processing on the orthogonal basis functions to obtain the corresponding orthogonal basis.

[0061] In this embodiment, the sample vibration data is first preprocessed according to steps 21 to 22, and then time-frequency transformed according to steps 31 to 33 to obtain the real-time difference frequency response data corresponding to the sample vibration data; then, the corresponding frequency response difference matrix is ​​constructed; in this embodiment, the frequency response difference matrix ΔH1 is subjected to feature dimensionality reduction using the centerless truncated singular value decomposition (Truncated SVD) algorithm. During the decomposition process, mean centering is not performed to retain the zero point representing the physical health benchmark, and m orthogonal basis functions are extracted. At the same time, the matrix corresponding to the orthogonal basis functions is processed according to the following formula: By imposing sign alignment constraints, the frequency domain peak with the largest absolute value in each basis function is forced to always be positive, ensuring that the characteristic projection coefficients are unique and strictly correspond to the damage aggravation process, thus obtaining the corresponding orthogonal basis; where ΔH1 is the frequency response difference matrix, containing ΔH at all times and all frequency points. i A two-dimensional data matrix of differences; m is the truncation order, used to determine the number of orthogonal basis functions (principal components) retained; φ kis the Kth orthogonal basis function, representing the most representative frequency domain waveform feature extracted from the frequency response data; sgn() represents the sign function, used to force the peak directions of the basis functions to be consistent, preventing positive and negative cancellation during feature projection; argmax is the maximum position operator, indicating the search for the frequency index with the largest absolute value in the orthogonal basis function vector.

[0062] Step 52: Using orthogonal basis as observation basis, construct initial hidden Markov state space model HMMSSM. The initial HMMSSM is associated with pre-set irreversible damage constraints, inter-layer unidirectional transmission decoupling, and diagonal noise covariance. The state transition matrix of the initial HMMSSM includes an upper triangular matrix, and the initial state probability of the initial HMMSSM includes a Dirac distribution.

[0063] Step 53: Using piecewise least squares and ridge regression, the initial parameters of the initial HMMSSM are calibrated to obtain the corresponding state basis function weight matrix and noise covariance matrix.

[0064] In this embodiment, a state transition probability matrix C is defined, C∈S×S, and its elements C jk =0, ∀j>k represents the probability of transitioning from state j to state k. To satisfy the physical irreversibility constraint, the state transition probability matrix C is set as an upper triangular matrix, that is, reverse state transition is strictly prohibited: C jk =0, ∀j>k; In this embodiment, the state probability distribution vector is defined as π∈S×1, π=[1,0,…,0] T Let y i,t,f ∈n×1 represents the observation column vector of the i-th sample at frequency point f within time period t, and the observation equation is defined as: Where A∈n×n is the interlayer transfer coefficient matrix; W∈n×n is the preset constraint matrix. Since the platform receives unidirectional input from bottom to top, the lower layer response constitutes the upper layer excitation, and the upper layer response is assumed not to have backflow affecting the lower layer, therefore it is defined as a strictly lower triangular matrix. This is the Hadamard product, which is the product of each element. This is the column vector of the m-th order orthogonal basis functions extracted by SVD dimensionality reduction at frequency f; For the current implicit state z i，t The basis function weight matrix z i，t This represents the hidden state of the i-th sample at time t; e i,t,f This represents the high-frequency noise from the sensor and environmental interference during the measurement process, assumed to follow a multivariate normal distribution with a mean of 0. To reduce model complexity, the noise covariance matrix ∑ f∈n×n is constrained to be a diagonal matrix, that is, it is assumed that the noise conditions of each layer after the propagation effect is decoupled are independent; let (where I) n (where n is the identity matrix), we can obtain Assuming that the observations at consecutive O frequency points are independent, then in the time period t, the complete observation matrix y is... i,t emission probability density function B s (y i,t It can be given by the product of the multivariate Gaussian likelihoods at each frequency point: .

[0065] Step 54: Based on the EM algorithm, the posterior probability of the hidden state is inferred through the forward-backward algorithm corresponding to the E step. In the M step, the posterior probability of the hidden state is used as the weight, and the corresponding state transition matrix, basis function weights, inter-layer transfer coefficients and noise variance are updated by weighted least squares. The E step and M step are iterated alternately until the parameters converge, and the energy mode prediction model is generated.

[0066] In this embodiment, in the absence of state labels, offline data from the early stages of testing is used to assign initial values ​​through a "piecewise least squares initialization strategy" that integrates temporal priors. During the initialization of the observation equation parameters, the frequency response observation sequence of total length T is forcibly divided into S local segments along the time axis (where S is the set total number of hidden states). For the local observation subsequence divided to the s-th hidden state (of length T), s At each discrete frequency point ω, the excitation interference of the carrier platform is first stripped: ,in, , This is the original observation matrix at frequency point ω in state s. This is the frequency response sequence after eliminating interlayer propagation. Subsequently, the basis function characteristic weight matrix of each state core is solved. A multivariate linear regression model is constructed. Under the assumption of state s, the observation matrices of all frequency points ω are vertically stacked in the time dimension to construct the response matrix Y. stack ∈ (T) s ·O)×n, synchronously, the orthogonal basis vector φ corresponding to each frequency point is... ω ∈m×1 stacked along the time axis T s Then, all frequency points are stacked to construct the design matrix X. stack ∈ (T) s ·O)×m, at this time, the linear expectation equation of the system can be expressed as: Among them, E stack ∈ (T) s ·O)×n represents Gaussian noise. To robustly solve the above equations and avoid problems caused by multicollinearity, To address the problems arising from inversion, Tikhonov regularization (i.e., Ridge Regression) is introduced. A small regularization penalty constant λ is applied along the diagonal direction, and the target damage function is reconstructed as: Differentiating the above equation and setting the gradient to zero, we obtain the closed-form analytical solution of the normal equation: For each frequency point ω, the observed data from all hidden state segments are aggregated, and the global sample variance vector is calculated along the time axis to construct the covariance matrix ∑. ω Since the system must be in a healthy, intact state at the start of the test, the initial state probability vector is defined as the Dirac distribution π=[1,0,…,0]. T Meanwhile, the fatigue crushing of materials is an irreversible cumulative process. Therefore, the initial state probability matrix C is constructed as an upper triangular matrix, while the interlayer transition coefficient matrix A is randomly assigned ω.

[0067] In this embodiment, the Baum-Welch iterative framework is then entered. E-step updates the posterior probability: ; ; The M-step updates the model parameters using posterior probabilities as weights: ; ; ; ; ; Where T is the total length of the observation sequence, representing the entire data period used for offline training; T s Let be the length of the s-th subsequence; This represents the original observation submatrix, which is divided into the original frequency response data block within state s; To remove the transmitted frequency response and eliminate the influence of interlayer crosstalk, the pure structural response is obtained; λ is the ridge regression regularization coefficient to prevent the occurrence of singular small penalty constants when inverting the matrix; α is the regularization loss function used to solve the optimization objective function for the feature weights; i,t (s) represents the forward probability / forward variable. β represents the joint probability that, in the i-th sample, given the first t observations, the system is in state s and the observation sequence is as follows: i,t (s) represents the backward probability / backward variable. Let represent the probability of generating the remaining observation sequence in the i-th sample, given the current state s; Let be the posterior probability of state occupancy, representing the probability that, in the i-th sample, at time t, it belongs to state s under the current model; M represents the posterior probability of state transition, specifically the probability that state j jumps to state k at time t in the i-th sample. mask (j, k) is the upper triangular mask matrix, where each element is 0 or 1, used to force the prohibition of reverse transfer (the value is 0 when j > k); A0 is the initial estimate of the interlayer transfer coefficient matrix; Y is the basis function weight matrix in the s-th state; stack A vertically stacked response matrix represents the observation vectors of all frequency points and time steps stacked row-wise; X stack Design a matrix for vertical stacking, by repeating the basis function vector T. s E is obtained by stacking all frequency points. stack The stacked noise matrix; I m K is an m-order identity matrix; i Let a be the total number of time windows for the i-th sample; l Let be the interlayer transfer coefficient vector, representing the parameters of the l-th row of matrix A; Let be the basis function coefficient vector, representing the matrix. The parameters in the l-th row; Let f be the noise variance of the l-th layer at frequency f, and let f be the diagonal elements of the noise covariance matrix. This is the index of the layer above the l-th level; It is the square of the Frobenius norm.

[0068] Finally, the E-step (inferring the posterior probability of the hidden state using the forward-backward algorithm) and the M-step (updating the basis function weight matrix and noise covariance through weighted least squares regression) are executed alternately to achieve autonomous convergence learning of the model parameters, resulting in the energy mode prediction model.

[0069] This embodiment also provides an online fatigue crush detection device for stacked cartons. This device is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0070] Figure 3 This is a structural block diagram of the online fatigue crush detection device for stacked cartons according to an embodiment of this application, as shown below. Figure 3 As shown, the device includes an acquisition module 31, a processing module 32, a prediction module 33, and a judgment module 34, wherein, The acquisition module 31 is used to acquire vibration data corresponding to each target object. The target objects include the support platform for carrying stacked cartons and each layer of cartons. The vibration data includes the acceleration time series corresponding to each target object. The acceleration time series is used to characterize the comprehensive vibration response of each target object at the current moment. The processing module 32, coupled to the acquisition module 31, is used to preprocess the acceleration time series to generate a first acceleration time series corresponding to each target object, and to extract frequency response features and perform differential processing on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked carton. The frequency response features are determined based on the frequency domain self-power spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the carton. The prediction module 33, coupled to the processing module 32, is used to input real-time difference frequency response data into the trained energy mode prediction model to obtain expected frequency response data corresponding to the real-time difference frequency response data. Based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data, the total prediction residual is determined. The expected frequency response data is used to characterize the vibration energy corresponding to each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning and training on the constructed initial hidden Markov state space (HMMSSM) model. The judgment module 34, coupled to the prediction module 33, is used to determine whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and to determine the crushing result of the stacked cartons based on the judgment result.

[0071] In some embodiments, the processing module 32 further includes: The generation unit is used to perform detrending processing on the acceleration time series corresponding to each target object using the least squares method to generate a time-domain series; The filtering unit, coupled to the generation unit, is used to perform zero-phase forward and reverse digital filtering on the time-domain sequence using a preset target digital low-pass filter to obtain the first acceleration time sequence. The target digital low-pass filter includes one of the following: Butterworth digital low-pass filter and Gaussian low-pass filter.

[0072] In some embodiments, the processing module 32 further includes: The conversion unit is used to perform time-frequency conversion on the first acceleration time series to obtain frequency domain power spectrum data corresponding to each target object. The frequency domain power spectrum data is used to characterize the power spectrum generated in each time window associated with the frequency domain power spectrum data after the external excitation of the input preset frequency. The frequency domain power spectrum data includes the external excitation power spectrum and the response power spectrum. The determining unit, coupled to the conversion unit, is used to determine the frequency domain self-power spectrum corresponding to the carrier platform and the frequency domain cross-power spectrum generated by the carrier platform for each layer of carton based on the external excitation power spectrum and the response power spectrum. Based on the ratio of the frequency domain cross-power spectrum to the frequency domain self-power spectrum, the frequency response function corresponding to each carton is generated. The frequency response function is used to characterize the frequency response features corresponding to the first acceleration time series. The differential unit, coupled to the deterministic unit, is used to linearize the frequency response function corresponding to each carton, and then perform differential processing on the full life cycle frequency response data obtained by linearization and the preset reference frequency response data to generate the corresponding real-time differential frequency response data. The linearization process includes logarithmic processing.

[0073] In some embodiments, the prediction module 33 further includes: The decoupling unit is used to perform spatial heteroscedasticity decoupling on real-time difference frequency response data and expected frequency response data using a preset covariance inverse matrix, and generate standard difference frequency response data and standard expected frequency response data respectively. The calculation unit, coupled to the decoupling unit, is used to calculate the Mahalanobis distance between the standard difference frequency response data and the standard expected frequency response data, and to average the generated Mahalanobis distance to generate the one-step prediction residual (OSPE). The total prediction residual includes the one-step prediction residual (OSPE).

[0074] In some embodiments, the stacked carton fatigue crushing online detection device is further configured to, before determining whether the EWMA value corresponding to the total predicted residual is greater than the preset anti-shake control feature energy threshold, perform low-pass smoothing on the total predicted residual in chronological order using the exponentially weighted moving average EWMA algorithm to obtain the EWMA statistic, wherein the EWMA value includes the EWMA statistic; obtain the historical total predicted residual corresponding to multiple detection times in a non-crushing state, and determine the EWMA mean and EWMA standard deviation of the EWMA statistic corresponding to the multiple historical total predicted residuals; and use the sum of the EWMA mean and the anti-shake control feature dynamic value as the anti-shake control feature energy threshold, wherein the anti-shake control feature dynamic value is determined based on the product of the preset control coefficient and the EWMA standard deviation.

[0075] In some embodiments, the determination module 34 further includes: The first judgment unit is used to determine that the cardboard boxes in the stacked boxes are crushed when the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and to trigger a cardboard box crushing alarm.

[0076] The second judgment unit is used to determine that there is no crushing of the stacked cartons when the EWMA value corresponding to the total prediction residual is not greater than the preset anti-shake control characteristic energy threshold.

[0077] In some embodiments, the online fatigue crush detection device for stacked cardboard boxes is further used to generate an energy mode prediction model through the following steps: constructing a frequency response difference matrix based on real-time difference frequency response data corresponding to preset sample vibration data; performing centerless truncated singular value decomposition on the frequency response difference matrix to extract a preset number of orthogonal basis functions; and performing sign alignment constraint processing on the orthogonal basis functions to obtain the corresponding orthogonal basis; constructing an initial hidden Markov state space model (HMMSSM) using the orthogonal basis as the observation basis, wherein the initial HMMSSM is associated with preset damage irreversibility constraints, interlayer unidirectional transmission decoupling, and diagonal noise covariance; the state transition matrix of the initial HMMSSM includes an upper triangular matrix, and the initial state probability of the initial HMMSSM includes a Dirac distribution; using piecewise least squares and ridge regression methods to calibrate the initial parameters of the initial HMMSSM to obtain the corresponding state basis function weight matrix and noise covariance matrix; based on the EM algorithm, inferring the posterior probability of the hidden state through the forward-backward algorithm corresponding to the E step, and then... In the step, the hidden state posterior probability is used as the weight, and the corresponding state transition matrix, basis function weights, inter-layer transfer coefficients and noise variance are updated using weighted least squares. The E-step and M-step are iterated alternately until the parameters converge, generating an energy mode prediction model.

[0078] This embodiment also provides an online fatigue crush detection system for stacked cartons, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0079] Optionally, the above-mentioned online fatigue crush detection system for stacked cartons may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0080] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0081] S1, acquire vibration data corresponding to each target object, wherein the target object includes the support platform for carrying stacked cardboard boxes and each layer of cardboard boxes, and the vibration data includes the acceleration time series corresponding to each target object, which is used to characterize the comprehensive vibration response of each target object at the current moment.

[0082] S2, preprocess the acceleration time series to generate the first acceleration time series corresponding to each target object, and extract and differentially process the frequency response features of the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked carton. The frequency response features are determined based on the frequency domain self-power spectrum corresponding to the first acceleration time series corresponding to the carrier platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the carton.

[0083] S3. Input the real-time difference frequency response data into the trained energy mode prediction model to obtain the expected frequency response data corresponding to the real-time difference frequency response data. Based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data, determine the total prediction residual. The expected frequency response data is used to characterize the vibration energy corresponding to each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning and training on the constructed initial hidden Markov state space (HMMSSM) model.

[0084] S4. Determine whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and determine the crushing result of the stacked carton based on the judgment result.

[0085] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0086] Furthermore, in conjunction with the online fatigue crush detection method for stacked cartons in the above embodiments, this application embodiment can provide a storage medium for implementation. This storage medium stores a computer program; when executed by a processor, the computer program implements any of the online fatigue crush detection methods for stacked cartons in the above embodiments.

[0087] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for online detection of fatigue crushing of stacked cardboard boxes, characterized in that, include: The vibration data corresponding to each target object is obtained. The target object includes a support platform for stacking cardboard boxes and each layer of cardboard boxes. The vibration data includes an acceleration time series corresponding to each target object. The acceleration time series is used to characterize the comprehensive vibration response of each target object at the current moment. The acceleration time series is preprocessed to generate a first acceleration time series corresponding to each target object, and frequency response features are extracted and differentially processed on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked cardboard box. The frequency response features are determined based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the cardboard box. The real-time differential frequency response data is input into the trained energy mode prediction model to obtain the expected frequency response data corresponding to the real-time differential frequency response data. Based on the similarity between the real-time differential frequency response data and the corresponding expected frequency response data, the total prediction residual is determined. The expected frequency response data is used to characterize the vibration energy of each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time differential frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning and training on the constructed initial hidden Markov state space (HMMSSM) model. Determine whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and determine the crushing result of the stacked cartons based on the determination result.

2. The method according to claim 1, characterized in that, The acceleration time series is preprocessed to generate a first acceleration time series corresponding to each target object, including: Using the least squares method, the acceleration time series corresponding to each target object is detrended to generate a time-domain series. The time-domain sequence is subjected to zero-phase forward and reverse digital filtering using a preset target digital low-pass filter to obtain the first acceleration time sequence. The target digital low-pass filter includes one of the following: a Butterworth digital low-pass filter and a Gaussian low-pass filter.

3. The method according to claim 2, characterized in that, Frequency response feature extraction and differential processing are performed on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked carton, including: The first acceleration time series is converted to a time-frequency signal to obtain frequency domain power spectrum data corresponding to each target object. The frequency domain power spectrum data is used to characterize the power spectrum generated in each time window associated with the frequency domain power spectrum data after an external excitation at a preset frequency is input. The frequency domain power spectrum data includes the external excitation power spectrum and the response power spectrum. Based on the external excitation power spectrum and the response power spectrum, the frequency domain self-power spectrum corresponding to the carrier platform and the frequency domain cross-power spectrum generated by the carrier platform for each layer of cardboard are determined. Based on the ratio of the frequency domain cross-power spectrum to the frequency domain self-power spectrum, a frequency response function corresponding to each cardboard is generated, wherein the frequency response function is used to characterize the frequency response characteristics corresponding to the first acceleration time series. The frequency response function corresponding to each carton is linearized, and the full life cycle frequency response data obtained by linearization is differentially processed with the preset reference frequency response data to generate the corresponding real-time difference frequency response data. The linearization process includes logarithmic processing.

4. The method according to claim 1, characterized in that, Based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data, the total prediction residual is determined, including: Using a preset covariance inverse matrix, spatial heteroscedasticity decoupling is performed on the real-time difference frequency response data and the expected frequency response data to generate standard difference frequency response data and standard expected frequency response data, respectively. Calculate the Mahalanobis distance between the standard difference frequency response data and the standard expected frequency response data, and average the generated Mahalanobis distance to generate a one-step prediction residual (OSPE), wherein the total prediction residual includes the one-step prediction residual (OSPE).

5. The method according to claim 4, characterized in that, Before determining whether the EWMA value corresponding to the total prediction residual is greater than the preset image stabilization control feature energy threshold, the method further includes: Using the exponentially weighted moving average (EWMA) algorithm, the total prediction residuals are low-pass smoothed in time order to obtain the EWMA statistic, wherein the EWMA value includes the EWMA statistic. Obtain the historical total prediction residuals corresponding to multiple detection times in a non-crushable state, and determine the EWMA mean and EWMA standard deviation of the EWMA statistic corresponding to the multiple historical total prediction residuals; The sum of the EWMA mean and the dynamic value of the image stabilization control feature is used as the energy threshold of the image stabilization control feature, wherein the dynamic value of the image stabilization control feature is determined by multiplying the preset control coefficient by the standard deviation of the EWMA.

6. The method according to claim 5, characterized in that, Based on the assessment results, the crushing outcome of the stacked cardboard boxes is determined, including: If the EWMA value corresponding to the total prediction residual is found to be greater than the preset anti-shake control characteristic energy threshold, it is determined that the carton in the stacked carton is crushed, and a carton crushing alarm is triggered. If the EWMA value corresponding to the total prediction residual is not greater than the preset anti-shake control characteristic energy threshold, it is determined that there is no crushing of the stacked cartons.

7. The method according to claim 1, characterized in that, Generating the energy mode prediction model includes: A frequency response difference matrix is ​​constructed based on the real-time difference frequency response data corresponding to the preset sample vibration data. The frequency response difference matrix is ​​subjected to centerless truncated singular value decomposition to extract a preset number of orthogonal basis functions. The orthogonal basis functions are then subjected to sign alignment constraint processing to obtain the corresponding orthogonal basis. Using the orthogonal basis as the observation basis, an initial hidden Markov state-space model (HMMSSM) is constructed. The initial HMMSSM is associated with a preset irreversible damage constraint, inter-layer unidirectional transmission decoupling, and diagonal noise covariance. The state transition matrix of the initial HMMSSM includes an upper triangular matrix, and the initial state probability of the initial HMMSSM includes a Dirac distribution. The initial parameters of the initial HMMSSM are calibrated using piecewise least squares and ridge regression to obtain the corresponding state basis function weight matrix and noise covariance matrix. Based on the EM algorithm, the posterior probability of the hidden state is inferred through the forward-backward algorithm corresponding to the E step, and in the M step, the posterior probability of the hidden state is used as the weight to update the corresponding state transition matrix, basis function weights, inter-layer transfer coefficients and noise variance using weighted least squares. The E step and M step are iterated alternately until the parameters converge to generate the energy mode prediction model.

8. An online fatigue crush detection device for stacked cardboard boxes, characterized in that, include: The acquisition module is used to acquire vibration data corresponding to each target object. The target object includes a support platform for stacking cardboard boxes and each layer of cardboard boxes. The vibration data includes an acceleration time series corresponding to each target object. The acceleration time series is used to characterize the comprehensive vibration response of each target object at the current moment. The processing module is used to preprocess the acceleration time series to generate a first acceleration time series corresponding to each target object, and to extract frequency response features and perform differential processing on the first acceleration time series to generate real-time differential frequency response data corresponding to each stacked carton. The frequency response features are determined based on the frequency domain autopower spectrum corresponding to the first acceleration time series corresponding to the support platform and the frequency domain cross-power spectrum corresponding to the first acceleration time series corresponding to the carton. The prediction module is used to input the real-time difference frequency response data into the trained energy mode prediction model to obtain the expected frequency response data corresponding to the real-time difference frequency response data, and to determine the total prediction residual based on the similarity between the real-time difference frequency response data and the corresponding expected frequency response data. The expected frequency response data is used to characterize the vibration energy of each carton at present. The similarity is determined based on the Mahalanobis distance between the real-time difference frequency response data and the expected frequency response data. The energy mode prediction model is generated by using the expectation-maximization (EM) algorithm to perform unsupervised parameter calibration learning and training on the constructed initial hidden Markov state space (HMMSSM) model. The judgment module is used to determine whether the EWMA value corresponding to the total prediction residual is greater than the preset anti-shake control characteristic energy threshold, and to determine the crushing result of the stacked cartons based on the judgment result.

9. An online fatigue crush detection system for stacked cardboard boxes, comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the steps of the online fatigue crush detection method for stacked cartons as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the online fatigue crush detection method for stacked cartons as described in any one of claims 1 to 7.

Images