A fire-fighting robot online state monitoring and fault intelligent diagnosis system

Abstract

The application belongs to the technical field of fire-fighting equipment, and particularly relates to an online state monitoring and intelligent fault diagnosis system for a fire-fighting robot. The system comprises an edge computing terminal, a communication signal collector, a vibration signal collector, a vibration signal sensor, a cloud server and a user terminal. The edge computing terminal is used for receiving data uploaded by the communication signal collector and the vibration signal collector, and is provided with a fault diagnosis algorithm, an EMD decomposer and a support vector machine (SVM) fault diagnosis model optimized by Bayes. The fault diagnosis algorithm is used for pre-processing vibration signals, then the EMD decomposer is used for extracting internal modes of the signals, and multi-dimensional features such as statistics, envelope spectrum and instantaneous frequency are fused. Finally, the SVM fault diagnosis is used for intelligently judging the state of a transmission system of the robot, generating a state monitoring and fault analysis report, and uploading the report to the cloud server. The system has high diagnosis precision, high real-time performance and high operation and maintenance efficiency.

Description

Technical Field

[0001] This invention belongs to the field of fire-fighting equipment technology, specifically relating to an online status monitoring and intelligent fault diagnosis system for fire-fighting robots. Background Technology

[0002] When firefighting robots perform tasks in hazardous environments, their own health status is crucial. Traditional operation and maintenance methods rely on on-site personnel for regular inspections, which cannot achieve real-time, remote monitoring of the robot's operational status, resulting in problems such as delayed response and high operation and maintenance costs. Existing remote monitoring systems mostly focus on simple data collection and display, lacking the ability to conduct in-depth, intelligent fault analysis of core components (such as transmission systems), making it difficult to issue effective early warnings of faults.

[0003] Therefore, there is an urgent need for a condition monitoring and intelligent fault diagnosis system. This system can use advanced signal processing and machine learning algorithms to perform real-time and intelligent fault diagnosis on the vibration signals of key components (especially the transmission system) of fire-fighting robots at the edge, and upload the results to the cloud for remote users to monitor and manage in real time, thereby realizing intelligent and preventive maintenance of fire-fighting robots. Summary of the Invention

[0004] The purpose of this invention is to provide an online status monitoring and intelligent fault diagnosis system for fire-fighting robots with high diagnostic accuracy, low latency, low cost, and high reliability.

[0005] The online status monitoring and intelligent fault diagnosis system for firefighting robots provided by this invention includes an edge computing terminal, a communication signal acquisition unit, a vibration signal acquisition unit, a vibration signal sensor, a cloud server, and a user terminal; wherein:

[0006] The vibration signal sensor is installed on the chassis mobile drive system of the fire-fighting robot to collect the vibration status signal of the mobile drive system in real time.

[0007] The vibration signal acquisition device is connected to the vibration signal sensor and is used to receive and upload the vibration status signal to the edge computing terminal;

[0008] The communication signal acquisition device is connected to the communication bus of the fire-fighting robot and is used to collect the robot's overall communication signals and upload them to the edge computing terminal.

[0009] The edge computing terminal is used to receive data uploaded by the communication signal acquisition device and the vibration signal acquisition device; it is equipped with fault diagnosis algorithms, EMD decomposition (Huang, NE, Shen, Z., Long, SR, Wu, MC, Shih, HH, Zheng, Q., ... & Liu, HH (1998). The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis. Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 454 (1971), 903-995.), after Bayesian optimization (Snoek, J., Larochelle, H., & Adams, RP (2012). Practical Bayesian optimization of machine learning algorithms. Advances in neural information processing systems, 25.) and cross-validation (Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements The high-precision support vector machine (SVM) trained by Springer (Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine learning, 20(3), 273-297.) is used to obtain a fault diagnosis model. The fault diagnosis algorithm is used to preprocess the vibration signal (including resampling, segmentation, and bandpass filtering), and then use the EMD decomposer to extract the intrinsic mode of the signal and fuse multi-dimensional features such as statistics, envelope spectrum and instantaneous frequency. Finally, the state of the (robot) transmission system is intelligently judged through SVM fault diagnosis.

[0010] The cloud server is used to receive and store the fault analysis report, perform data storage, management and visualization, and generate accessible web data pages;

[0011] The user terminal is used to access the cloud server via a webpage or app to remotely view the operating status and fault analysis results of the fire-fighting robot; enabling maintenance personnel or managers to remotely access the robot's health status anytime, anywhere, and respond to early warning information in a timely manner.

[0012] In this invention, the specific operation steps of the edge computing terminal are as follows:

[0013] (1) The vibration state signal is resampled and segmented;

[0014] (2) Apply a bandpass filter to filter the segmented vibration signal;

[0015] (3) Perform Empirical Mode Decomposition (EMD) on each filtered signal segment to obtain N Intrinsic Mode Function (IMF) components;

[0016] (4) For each IMF component, extract a multi-dimensional feature vector including statistical features, envelope spectrum features and instantaneous frequency features, and concatenate the feature vectors of N IMF components to form the final feature vector.

[0017] (5) Input the final feature vector into the pre-trained SVM fault diagnosis model to obtain the fault diagnosis result of the transmission system;

[0018] (6) Based on the communication signals, comprehensively judge the overall operating status of the fire-fighting robot's communication system, control system, power system, drive system and transmission system, and generate a status monitoring and fault analysis report;

[0019] (7) Upload the status monitoring and fault analysis report to the cloud server via TCP / IP protocol.

[0020] In this invention, the fire-fighting robot chassis moving drive system includes components such as an engine, a drive motor, transmission gears, transmission belts, and bearings.

[0021] In this invention, the vibration signal acquisition device is an acceleration signal acquisition card; the communication signal acquisition device is a bus signal acquisition card; and the vibration signal sensor is an acceleration sensor.

[0022] In step (2) of this invention, the bandpass filter is a 4th-order Butterworth bandpass filter, and its transfer function is:

[0023]

[0024] Where s is the complex frequency variable, and B is the bandwidth, calculated using the following formula:

[0025] B=ω1-ω2 (2)

[0026] ω0 is the center angular frequency of the bandpass filter, measured in radians per second (rad / s). It is determined by the lower cutoff angular frequency ω1 and the upper cutoff angular frequency ω2 of the bandpass filter, and is calculated using the following formula:

[0027]

[0028] In step (3) of this invention, the EMD decomposer is an adaptive signal decomposer used to decompose any complex signal x(t) into several intrinsic mode functions (IMFs) and a residual term;

[0029] The original signal x(t) is represented as:

[0030]

[0031] Among them, IMF i (t) is the i-th eigenmode function, r N (t) is the final residual trend term, and N is the total number of IMFs;

[0032] The definition of the IMF must meet two conditions:

[0033] (I) Throughout the entire data segment, the number of extreme points is equal to or differs from the number of zero-crossing points by at most one;

[0034] (II) At any time, the mean of the upper envelope defined by the local maxima and the lower envelope defined by the local minima is zero.

[0035] The EMD decomposer decomposition algorithm flow is as follows:

[0036] Let the original signal be x(t), and let h(t) = x(t);

[0037] Step a: Find all local maxima and minima of h(t);

[0038] Step b: Using cubic spline interpolation, connect all the maxima and minima to obtain the upper envelope e. max (t) and lower envelope e min (t);

[0039] Step c: Calculate the envelope mean:

[0040]

[0041] Step d: Subtract the mean from the original signal to obtain the candidate IMF:

[0042] IMF1(t)=h new (t);

[0043] Step e: Calculate the residuals:

[0044] r1(t)=x(t)-IMF1(t) (5)

[0045] Step f: Using r1(t) as the new original signal, repeat the above process until the residual r N (t) becomes a monotonic function or contains fewer than two extreme points.

[0046] In step (4) of this invention, for each IMF component, i (t), extract the following three categories of features totaling 17 dimensions; namely:

[0047] (I) Time-domain statistical characteristics (5 dimensions)

[0048] Let the IMF signal be s = [s1, s2, ..., s L The length is L. The specific time-domain statistical characteristics are as follows:

[0049] Mean:

[0050]

[0051] Standard deviation:

[0052]

[0053] Root mean square:

[0054]

[0055] Maximum value:

[0056] s max =max(s) (9)

[0057] Kuroshi:

[0058]

[0059] (II) Envelope Spectrum Features (10 Dimensions)

[0060] The envelope spectrum feature extraction process is as follows:

[0061] Step a: Perform a Hilbert transform on the IMF to obtain the analytic signal:

[0062]

[0063] in, This represents the Hilbert transform.

[0064] Step b: Calculate the amplitude envelope:

[0065]

[0066] Step c: Perform a Fast Fourier Transform (FFT) on the envelope A(t) to obtain the envelope spectrum E(f):

[0067] E(f)=∣FFT{A(t)}∣ (13)

[0068] Step d: In the envelope spectrum E(f), find the top 5 frequency points {f1,f2,...,f5} with the largest amplitude and their corresponding amplitudes {E1,E2,...,E5}.

[0069] (III) Instantaneous frequency characteristics, which have two dimensions:

[0070] a: Calculate the instantaneous phase from the analytic signal z(t):

[0071] φ(t)=arg(z(t)) (14)

[0072] b: Unwind the phase to obtain a continuous phase.

[0073] c: Calculate the instantaneous frequency:

[0074]

[0075] In the discrete domain, the difference approximation is used;

[0076] d: Calculate the mean μ of the instantaneous frequency sequence finst and standard deviation σ finst ;

[0077] In step (4) of this invention, N is 5 for the N intrinsic mode functions (IMF) components; the multi-dimensional feature vector includes 17 features: mean, standard deviation, root mean square value, maximum value, kurtosis, the top 5 frequency points with the largest amplitude in the envelope spectrum, the top 5 amplitudes with the largest amplitude in the envelope spectrum, the mean of the instantaneous frequency, and the standard deviation of the instantaneous frequency; the final feature vector dimension is 5×17=85.

[0078] In step (5) of this invention, the SVM fault diagnosis model uses an SVM classifier with a radial basis function (RBF) kernel, as follows:

[0079] Decision function:

[0080]

[0081] Where: x i ,y i These are the training samples and their labels, α i It is a Lagrange multiplier (0≤α) i≤C, for all i), b is the bias term, and K(·,·) is the kernel function.

[0082] RBF kernel function:

[0083] K(x i ,x j )=exp(-γ||x i -x j || 2 (17)

[0084] C is the regularization parameter, which controls the degree of penalty for misclassification. γ is the coefficient of the RBF kernel, which controls the influence range of a single sample.

[0085] By employing a Bayesian optimization algorithm combined with a cross-validation strategy, the SVM hyperparameters in this pipeline are automatically optimized to obtain the optimal model.

[0086] The Bayesian optimization method aims to find the optimal hyperparameter θ. * =(C * ,γ * To maximize the accuracy of cross-validation.

[0087] Optimization issues:

[0088]

[0089] Where, Θ=[C min C max ]×[γ min ,γ max ] is the hyperparameter search space.

[0090] The specific Bayesian optimization process is as follows:

[0091] (1) Proxy model: The objective function CV-Accuracy(θ) is modeled using Gaussian process (GP) to obtain the posterior mean μ(θ) and variance σ(θ);

[0092] (2) Acquisition Function: Based on the surrogate model, selects the next most promising sampling point. Common desired improvements:

[0093] EI(θ)=E[max(f best -f(θ),0)] (19)

[0094] Among them, f best It is the currently known best target value;

[0095] (3) Iterative update: at θ nest =argmax θThe objective function is evaluated at EI(θ), and the GP model is updated until the maximum number of iterations is reached;

[0096] The trained optimal pipeline is deployed as a whole on the edge computing terminal.

[0097] In this invention, the entire system adopts a B / S architecture. The user terminal accesses the Web page provided by the cloud server through a browser to realize remote operation and maintenance management, remote service, and fault early warning of the fire-fighting robot.

[0098] The system architecture adopts a B / S model, mainly including:

[0099] Sensing layer: Composed of vibration signal sensors (installed in the transmission system), vibration signal acquisition unit and communication signal acquisition unit, responsible for collecting raw data;

[0100] Edge computing layer: The edge computing terminal is the core intelligent unit of this system;

[0101] Cloud service layer: The cloud server is responsible for receiving structured analysis reports uploaded by edge computing terminals, and for storing, managing, and visualizing them;

[0102] Application layer: The user terminal (Web) enables maintenance personnel or managers to remotely access the robot's health status anytime, anywhere, and respond to early warning information in a timely manner.

[0103] The main technical features and functional advantages of this invention are:

[0104] (1) Intelligent diagnosis: The advanced EMD-SVM algorithm is applied to the transmission system of the fire robot, realizing the leap from "data acquisition" to "intelligent diagnosis" with high diagnostic accuracy;

[0105] (2) Real-time performance and low latency: Computationally intensive signal processing and diagnostic tasks are completed at the edge, greatly reducing the dependence on network bandwidth and enabling near real-time fault detection.

[0106] (3) Remote operation and maintenance: Through the B / S architecture, maintenance personnel can fully grasp the robot status without being on-site, which significantly improves operation and maintenance efficiency and reduces costs;

[0107] (4) Preventive maintenance: The system can provide early warning of potential faults, transforming passive maintenance into proactive maintenance, thus ensuring the reliability and availability of the fire-fighting robot at critical moments. Attached Figure Description

[0108] Figure 1 This is a block diagram of the overall architecture of the system of the present invention.

[0109] Figure 2This is a flowchart of a fault diagnosis algorithm for a transmission system embedded in an edge computing terminal.

[0110] Figure 3 This is a schematic diagram of the overall assembly of the system of the present invention.

[0111] Figure 4 This is a schematic diagram of the installation of the system of the present invention.

[0112] Figure 5 The transmission pulley in the embodiment is in a toothed state. (a) Pulley with missing teeth, (b) Missing tooth section.

[0113] Figure 6 The vibration state time-domain / frequency-domain signals are shown in the four states described in the embodiment.

[0114] Figure 7 This is the confusion matrix for the test set in the example.

[0115] Figure 8 The example shows the multi-class ROC curve. Detailed Implementation

[0116] In the specific implementation of this system, when the firefighting robot is performing its tasks, vibration signal sensors on its transmission system continuously collect vibration data. This data is then transmitted to the edge computing terminal via a vibration signal acquisition unit. Simultaneously, a communication signal acquisition unit also transmits data from the robot's CAN bus or other communication buses to this terminal.

[0117] Upon receiving the data, the edge computing terminal immediately initiates its built-in fault diagnosis program. This program first resamples and segments the vibration signal, then filters out noise using a 4th-order Butterworth bandpass filter. Next, it performs EMD decomposition on each filtered signal segment (up to 5 IMFs), calculating 17-dimensional features for each IMF, ultimately forming an 85-dimensional feature vector. This feature vector is input into a pre-trained SVM model, which is optimized offline using Pipeline and Bayesian optimization algorithms. This model accurately outputs the health status of the transmission system (e.g., "standby," "normal," "bearing inner race fault," "gear wear," "missing tooth on the drive pulley," "drive pulley slippage," etc.).

[0118] The edge computing terminal combines vibration diagnostic results with communication signal analysis results to generate a comprehensive "Firefighting Robot Operation Status Report," which is then uploaded to a cloud server via 4G / 5G or a private network. The cloud server parses the report and updates the web page. Remote users can then open a browser or app to clearly see the robot's real-time status, historical trends, and fault warning information, enabling them to make appropriate maintenance decisions.

[0119] 1. Experiment and Results Analysis

[0120] 1.1 Experimental Setup

[0121] The experimental data were derived from common pulley failures in the transmission system of a simulated fire-fighting robot, including four states: "standby," "normal operation," "missing tooth on the transmission pulley," and "slipping transmission pulley." Multiple sets of signals were collected for each state. The data were divided into training / validation and test sets in an 8:2 ratio.

[0122] The experimental equipment mainly consists of a fire-fighting robot and vibration data acquisition equipment. The data acquisition point is located on the motor casing, which can directly collect the vibration characteristic signals transmitted from the pulley to the motor casing through the motor rotor.

[0123] For the firefighting robot in "standby" and "normal operation" states, vibration characteristic signals of the firefighting robot when stationary and during normal operation need to be collected respectively.

[0124] To address the issue of a missing tooth on the transmission pulley of a firefighting robot, an experiment was conducted to simulate a transmission pulley with a missing tooth, such as... Figure 5 As shown, it is installed on a fire-fighting robot and the robot operates continuously. Vibration characteristic signals of the transmission pulley in the missing tooth state are collected using vibration data acquisition equipment.

[0125] To address the slippage of the drive belt pulley in a firefighting robot, the position of the robot's drive motor was adjusted to loosen the drive belt pulley. When the robot operates, slippage occurs between the loose pulley and the motor rotor, generating corresponding vibration signals on the motor housing. These data were then collected using vibration data acquisition equipment.

[0126] Observing the vibration characteristic signals collected under four conditions, it can be found that in the states of "missing teeth on the drive pulley" and "slipping drive pulley," the low-frequency signals are higher compared to the "standby" and "normal operation" states. Especially around 100Hz, the frequency signal fluctuation in the "missing teeth on the drive pulley" state is more pronounced, and the relevant characteristic signals can characterize the corresponding pulley fault state. For example... Figure 6 As shown.

[0127] 1.2 Results Analysis

[0128] After Bayesian optimization, the model achieved a best accuracy of 93.69% in cross-validation. On the independent test set, the final model achieved a diagnostic accuracy of 94.01%.

[0129] Figure 7The confusion matrix of the test set is shown. Also known as the error matrix, it's a standard format for evaluating accuracy. Each column represents the predicted class, and the total in each column indicates the number of data points predicted as belonging to that class. Each row represents the true class of the data, and the total number of data points in each row indicates the number of data instances belonging to that class. As can be seen from the graph, the recognition rate for all types of faults is high, demonstrating strong discriminative ability for major fault modes.

[0130] Figure 8 This is a multi-class ROC (Receiver Operating Characteristic) curve. The ROC curve is a commonly used tool for evaluating the performance of binary classification models. It displays the model's performance under different classification conditions through different threshold settings. The ROC curve plots the false positive rate (FPR) on the horizontal axis and the true positive rate (TPR) on the vertical axis, comprehensively evaluating the model's classification ability. The AUC values ​​for all classes are greater than 0.98, further demonstrating the model's excellent classification performance and generalization ability.

Claims

1. A system for online status monitoring and intelligent fault diagnosis of a firefighting robot, characterized in that, This includes edge computing terminals, communication signal acquisition devices, vibration signal acquisition devices, vibration signal sensors, cloud servers, and user terminals; among which: The vibration signal sensor is installed on the chassis mobile drive system of the fire-fighting robot to collect the vibration status signal of the mobile drive system in real time. The vibration signal acquisition device is connected to the vibration signal sensor and is used to receive and upload the vibration status signal to the edge computing terminal; The communication signal acquisition device is connected to the communication bus of the fire-fighting robot and is used to collect the robot's overall communication signals and upload them to the edge computing terminal. The edge computing terminal is used to receive data uploaded by the communication signal collector and the vibration signal collector; it is equipped with a fault diagnosis algorithm, an EMD decomposer, and a high-precision support vector machine (SVM) fault diagnosis model trained through Bayesian optimization and cross-validation; the fault diagnosis algorithm is used to preprocess the vibration signal, then use the EMD decomposer to extract the intrinsic modes of the signal, and fuse multi-dimensional features such as statistics, envelope spectrum, and instantaneous frequency; finally, through SVM fault diagnosis, the state of the robot transmission system is intelligently judged; a state monitoring and fault analysis report is generated; and the state monitoring and fault analysis report is uploaded to the cloud server via TCP / IP protocol; The cloud server is used to receive and store the fault analysis report, perform data storage, management and visualization, and generate accessible web data pages; The user terminal is used to access the cloud server via a webpage or app to remotely view the operating status and fault analysis results of the fire-fighting robot; enabling maintenance personnel or managers to remotely access the robot's health status anytime and anywhere and respond to early warning information in a timely manner. The specific operation steps of the edge computing terminal are as follows: (1) The vibration state signal is resampled and segmented; (2) Apply a bandpass filter to filter the segmented vibration signal; (3) Perform Empirical Mode Decomposition (EMD) on each filtered signal segment to obtain N Intrinsic Mode Function (IMF) components; (4) For each IMF component, extract a multi-dimensional feature vector including statistical features, envelope spectrum features and instantaneous frequency features, and concatenate the feature vectors of N IMF components to form the final feature vector; (5) Input the final feature vector into the pre-trained SVM fault diagnosis model to obtain the fault diagnosis result of the transmission system; (6) Based on the communication signals, comprehensively judge the overall operating status of the fire-fighting robot's communication system, control system, power system, drive system and transmission system, and generate a status monitoring and fault analysis report; (7) Upload the status monitoring and fault analysis report to the cloud server via TCP / IP protocol; In step (2), the bandpass filter is a 4th-order Butterworth bandpass filter, and its transfer function is: Where s is the complex frequency variable, and B is the bandwidth, calculated using the following formula: It is the center angular frequency of the bandpass filter, measured in radians per second (rad / s). It is determined by the lower cutoff angular frequency of the bandpass filter. and upper cutoff angular frequency The decision was made jointly, and the calculation formula is as follows: ； In step (3), the EMD decomposer is an adaptive signal decomposer used to decompose any complex signal x(t) into several intrinsic mode functions (IMFs) and a residual term. The original signal x(t) is represented as: Among them, IMF i (t) is the i-th eigenmode function, r N (t) is the final residual trend term. for The total number; The definition of the IMF must meet two conditions: (I) The number of extreme points is equal to or differs from the number of zero-crossing points by at most one throughout the entire data segment; (II) At any given time, the mean of the upper envelope defined by the local maxima and the lower envelope defined by the local minima is zero; The EMD decomposer decomposition algorithm flow is as follows: Let the original signal be x(t), and let h(t) = x(t); Step a: Find all local maxima and minima of h(t); Step b: Using cubic spline interpolation, connect all the maxima and minima to obtain the upper envelope e. max (t) and lower envelope e min (t); Step c: Calculate the envelope mean: ； Step d: Subtract the mean from the original signal to obtain the candidate IMF: IMF1(t) = h new (t); Step e: Calculate the residuals: r1(t) = x(t) - IMF1(t) Step f: Using r1(t) as the new original signal, repeat the above process until the residual r N Until (t) becomes a monotonic function or contains fewer than two extreme points; In step (3), for each IMF component, the IMF... i (t), extract the following three categories of features totaling 17 dimensions; namely: (I) Time-domain statistical characteristics, consisting of 5 dimensions; Let the IMF signal be s = [s1, s2, ..., s]. L ], with a length of L; the specific time-domain statistical features are: Mean: ， Standard deviation: Root mean square: Maximum value: Kuroshi: (II) Envelope spectrum features, consisting of 10 dimensions; The envelope spectrum feature extraction process is as follows: Step a: Perform a Hilbert transform on the IMF to obtain the analytic signal: in, Represents the Hilbert transform; Step b: Calculate the amplitude envelope: Step c: For the envelope Perform a Fast Fourier Transform (FFT) to obtain the envelope spectrum E(f): E(f) = |FFT{A(t)}| Step d: In the envelope spectrum E(f), find the top 5 frequency points {f1,f2,...,f5} with the largest amplitude and their corresponding amplitudes {E1,E2,...,E5}; (III) Instantaneous frequency characteristics, which have two dimensions: a: Calculate the instantaneous phase from the analytic signal z(t): b: Unwind the phase to obtain a continuous phase. ; c: Calculate the instantaneous frequency: In the discrete domain, the difference approximation is used; d: Calculate the mean μ of the instantaneous frequency sequence finst and standard deviation σ finst ; In step (4) of this invention, N is 5 for the N intrinsic mode functions (IMF) components; the multi-dimensional feature vector includes 17 features: mean, standard deviation, root mean square value, maximum value, kurtosis, the top 5 frequency points with the largest amplitude in the envelope spectrum, the top 5 amplitudes with the largest amplitude in the envelope spectrum, the mean of the instantaneous frequency, and the standard deviation of the instantaneous frequency; the final feature vector dimension is 5 × 17 = 85; In step (5), the SVM fault diagnosis model uses an SVM classifier with radial basis function (RBF) kernels, as follows: Decision function: Where: x i ,y i These are the training samples and their labels, α i It is a Lagrange multiplier, 0≤α i ≤C, for all i, b is a bias term. It is a kernel function; RBF kernel function: C is the regularization parameter, which controls the degree of penalty for misclassification; The coefficients of the RBF kernel control the influence range of a single sample; The optimal model is obtained by automatically optimizing the SVM hyperparameters in the Pipeline using a Bayesian optimization algorithm combined with a cross-validation strategy. The Bayesian optimization method aims to find the optimal hyperparameters. To maximize cross-validation accuracy; The optimization problem is: in, It is the hyperparameter search space; The Bayesian optimization process is as follows: (1) Proxy model: using Gaussian process (GP) for the objective function Modeling to obtain the posterior mean and variance ; (2) Acquisition function: Based on the surrogate model, select the next most promising sampling point; improve using expectation: in, It is the currently known best target value; (3) Iterative update: In The objective function is evaluated and the GP model is updated until the maximum number of iterations is reached; The trained optimal pipeline is deployed as a whole on the edge computing terminal.

2. The online status monitoring and intelligent fault diagnosis system for firefighting robots according to claim 1, characterized in that, Using a B / S architecture, the user terminal accesses the web page provided by the cloud server through a browser to realize remote operation and maintenance management, remote service and fault early warning of the fire-fighting robot; The B / S model of system architecture mainly includes: Sensing layer: Composed of vibration signal sensors, vibration signal acquisition unit and communication signal acquisition unit, responsible for collecting raw data; Edge computing layer: This refers to edge computing terminals; Cloud service layer: The cloud server is responsible for receiving structured analysis reports uploaded by edge computing terminals, and for storing, managing, and visualizing them; Application layer: The user terminal enables maintenance personnel or managers to remotely access the robot's health status anytime, anywhere, and respond to early warning information in a timely manner.

Images