Fire-fighting robot multi-sensor fusion fire source positioning method

Abstract

The invention relates to the technical field of fire-fighting robots, and discloses a fire-fighting robot multi-sensor fusion fire source positioning method, which comprises the following steps: loading a three-dimensional map of a working area by a robot, and extracting geometric parameters of a strong reflection object in the three-dimensional map; the robot synchronously collects original data of the radar sensor, the hyperspectral infrared sensor, the hyperspectral ultraviolet sensor and the hyperspectral visible light sensor; calculating and screening radar data by using the position information of the known reflector, and eliminating false target points caused by reflection; interference spectrum characteristics of specific chemical substances are matched and stripped from the collected infrared spectrums; carrying out association and fusion calculation on the screened radar target and the purified infrared region, and finally determining a three-dimensional space coordinate of the fire source; and converting the fire source coordinate into a control instruction to drive the robot. According to the method, the priori environment geometric information is used for reflection path pre-screening, and the online reactant spectral feature purification technology is combined, so that three-dimensional positioning of the root of the fire source can be realized in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of firefighting robot technology, specifically to a fire source localization method using multi-sensor fusion for firefighting robots. Background Technology

[0002] In existing commonly used technologies, fire source localization methods for firefighting robots based on multi-sensor fusion typically rely on the direct measurement and fusion of sensor signals. For example, spatial ranging is performed by processing the time difference of arrival or signal strength information of millimeter-wave radar, while fire source identification and verification are performed by combining the radiation spectral characteristics detected by infrared and ultraviolet sensors. Finally, the fire source location is calculated by weighting or geometric intersection of multi-source information.

[0003] However, in specific scenarios such as the reaction vessel areas of fine chemical plants, existing methods face an inherent and coupled technical problem. This environment is not only densely populated with large metal reactors, pipes, and guardrails—strong reflectors—resulting in numerous difficult-to-distinguish reflection path interferences of electromagnetic waves and infrared signals, severely contaminating direct wave signals; simultaneously, in accident conditions, unburned vapors of certain chemicals can form interfering gas clouds with characteristic spectra, whose radiation or absorption lines highly overlap with the flame spectrum. These two types of interference—physical multipath reflection and chemical spectral confusion—often coexist and are coupled together.

[0004] Existing technologies have limitations in addressing this coupling problem. Most methods focus on suppressing single types of interference, such as reducing multipath effects through algorithmic averaging or distinguishing flames from the background using fixed band ratios. They fail to effectively utilize prior static geometric information of the environment for proactive prediction and elimination of reflection paths, and lack the ability to perform real-time spectral feature stripping for dynamically occurring specific chemical interferences. Therefore, coupling interference mixed in the raw signals acquired by the sensors leads to distortion of the positioning basis, causing deviations in subsequent fusion calculations and affecting the targeting of the initial firefighting actions of the firefighting robot. Therefore, the following solution is proposed to address these issues. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a multi-sensor fusion method for fire source localization using a firefighting robot. The firefighting robot is deployed in an environment with a known geometric layout of strong reflectors and specific interfering substances. The robot is equipped with a central processing unit, millimeter-wave radar, a multi-channel hyperspectral infrared imager, and an environmental database. The method includes the following steps: The central processing unit loads the pre-stored geometric information of the strong reflector from the environment database; Simultaneously acquire raw detection data from millimeter-wave radar and multi-channel hyperspectral infrared imager; Based on the geometric information of strong reflectors, the reflection path conformity of each target point detected by millimeter-wave radar is verified, and a set of direct wave candidate target points is selected based on the verification results. The system calls upon the spectral characteristics of specific interfering substances pre-stored in the environmental database to decompose the mixed spectral data acquired by the multi-channel hyperspectral infrared imager, thereby removing interfering spectral components, obtaining purified spectral data, and calculating spectral confidence. The selected direct wave candidate target points are spatially correlated and fused with the purified spectral data and their spectral confidence, and the fusion confidence of each candidate point is calculated. The candidate main fire source is determined based on the fusion confidence level, and the final three-dimensional coordinates of the fire source are calculated based on the geometric relationship of multi-source sensor information.

[0006] Preferably, the step of performing reflection path conformity verification for each target point includes: The robot's position is taken as the signal transmission point, and the radar detection target point is taken as the assumed direct wave source point; Traverse each reflection plane in the geometric information of the strong reflector and calculate the mirror image of the assumed direct wave source point with respect to each reflection plane; Calculate the reflection path length from the launch point to each mirror point and compare it with the target point distance measured by the radar. When a reflecting plane exists such that the difference between the reflection path length and the measured distance is less than the preset tolerance, and the relationship between the incident angle and the reflection angle satisfies the law of reflection, the signal at the target point is determined to have a suspected risk of reflection.

[0007] Preferably, the step of decomposing the mixed spectral data includes: Retrieve the standard spectral feature vector of at least one known interfering substance related to the current environment from the environmental database; Establish a hybrid spectral model that includes a real flame base spectral template, a linear combination of interfering substance spectra, and background radiation; By solving the optimization problem, the global weight coefficients of the spectra of each interfering substance in the mixed spectrum under the current scenario are estimated. By using global weighting coefficients, the pixel-level spectral data acquired by the multi-channel hyperspectral infrared imager is corrected, and the estimated interference spectral components are subtracted to obtain the purified pixel spectrum.

[0008] Preferably, the step of calculating the spectral confidence is as follows: calculate the similarity measure between the purified spectrum of each pixel and the pre-stored standard flame basic spectral template, and use the similarity measure value as the spectral confidence of the pixel.

[0009] Preferably, the steps for spatial association and fusion include: Project the direct wave candidate target points onto the image plane of the multi-channel hyperspectral infrared imager; For each projected candidate target point, determine its corresponding pixel region on the image plane; Calculate the statistical characteristics of the spectral confidence of all pixels within the pixel region.

[0010] Preferably, in the step of calculating the fusion confidence of each candidate point, the fusion confidence is a function of at least two of the following factors: the radar echo intensity of the candidate target point, the reflection risk level represented by its reflection path verification result, and the statistical characteristic value of the spectral confidence of its corresponding pixel region.

[0011] Preferably, the steps for calculating the final three-dimensional coordinates of the fire source include: Select the candidate target point with the highest fusion confidence as the main fire source candidate point and obtain its three-dimensional coordinates; In the purified spectral data, independent high-temperature regions with high spectral confidence that are not associated with high fusion confidence radar points are identified and converted into spatial direction lines originating from the robot. Based on the coordinates of the candidate main fire source points, and combined with the geometric constraints formed by the spatial direction lines, the coordinates of three-dimensional spatial points that satisfy the constraints are solved through an optimization algorithm, which are then used as the final three-dimensional coordinates of the fire source.

[0012] Preferably, the firefighting robot is also equipped with an ultraviolet flame detector; the method further includes: Acquire signals from the ultraviolet flame detector; The signal strength or its changing trend of the ultraviolet flame detector is used as an auxiliary criterion to verify or alarm the reliability of the final three-dimensional coordinates of the fire source.

[0013] Preferably, after calculating the final three-dimensional coordinates of the fire source, the method further includes: The final three-dimensional coordinates of the fire source are transformed into the body coordinate system of the fire-fighting robot; Based on the transformed coordinates, control commands are generated to drive the fire-fighting robot's fire extinguishing device to align with the three-dimensional coordinates of the fire source.

[0014] The present invention has the following beneficial effects: (1) By using known environmental geometric information to pre-calculate and screen potential reflection path signals, this invention can effectively identify and suppress false target points generated by multiple reflections from the surface of metal equipment during the front-end data processing stage. This process dynamically integrates static prior knowledge into real-time signal analysis, reduces the dependence on posterior identification of reflection interference signals, and makes the radar data source used for positioning itself have higher spatial credibility, which helps to improve the stability of subsequent fusion positioning. (2) By establishing an online spectral decomposition model, this invention can dynamically separate and deduct the characteristic radiation or absorption spectral components of specific reactant vapors from the mixed infrared spectrum. This step reveals the masked background spectral characteristics of the flame, making the flame identification and temperature inversion results based on spectral analysis closer to the real fire source state. This capability improves the sensor's judgment of the nature of the fire source in the scenario of chemical material leakage accompanied by combustion. (3) This invention establishes a fusion credibility model of associated radar target points and purified infrared image features, and deeply couples the evidence from different sensors after preprocessing. The model comprehensively considers multiple dimensions of information such as spatial consistency, spectral matching degree and signal quality, so that the final fire source coordinate calculation is not only based on signal strength, but also on the logical consistency of mutual support between multiple sources of evidence, thereby leading to a more reasonable positioning result. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, please refer to Figure 1 This invention relates to a multi-sensor fusion method for fire source localization in firefighting robots. The firefighting robot is deployed in an environment with strong reflectors and specific interfering substances of known geometry. The robot is equipped with a central processing unit, millimeter-wave radar, a multi-channel hyperspectral infrared imager, and an environmental database. The method includes the following steps: The central processing unit loads the pre-stored geometric information of the strong reflector from the environment database; Simultaneously acquire raw detection data from millimeter-wave radar and multi-channel hyperspectral infrared imager; Based on the geometric information of strong reflectors, the reflection path conformity of each target point detected by millimeter-wave radar is verified, and a set of direct wave candidate target points is selected based on the verification results. The system calls upon the spectral characteristics of specific interfering substances pre-stored in the environmental database to decompose the mixed spectral data acquired by the multi-channel hyperspectral infrared imager, thereby removing interfering spectral components, obtaining purified spectral data, and calculating spectral confidence. The selected direct wave candidate target points are spatially correlated and fused with the purified spectral data and their spectral confidence, and the fusion confidence of each candidate point is calculated. The candidate main fire source is determined based on the fusion confidence level, and the final three-dimensional coordinates of the fire source are calculated based on the geometric relationship of multi-source sensor information.

[0019] The steps for performing reflection path compliance verification for each target point include: The robot's position is taken as the signal transmission point, and the radar detection target point is taken as the assumed direct wave source point; Traverse each reflection plane in the geometric information of the strong reflector and calculate the mirror image of the assumed direct wave source point with respect to each reflection plane; Calculate the reflection path length from the launch point to each mirror point and compare it with the target point distance measured by the radar. When a reflecting plane exists such that the difference between the reflection path length and the measured distance is less than the preset tolerance, and the relationship between the incident angle and the reflection angle satisfies the law of reflection, the signal at the target point is determined to have a suspected risk of reflection.

[0020] The steps for decomposing mixed spectral data include: Retrieve the standard spectral feature vector of at least one known interfering substance related to the current environment from the environmental database; Establish a hybrid spectral model that includes a real flame base spectral template, a linear combination of interfering substance spectra, and background radiation; By solving the optimization problem, the global weight coefficients of the spectra of each interfering substance in the mixed spectrum under the current scenario are estimated. By using global weighting coefficients, the pixel-level spectral data acquired by the multi-channel hyperspectral infrared imager is corrected, and the estimated interference spectral components are subtracted to obtain the purified pixel spectrum.

[0021] The steps for calculating spectral confidence are as follows: calculate the similarity measure between the purified spectrum of each pixel and the pre-stored standard flame base spectrum template, and use the similarity measure value as the spectral confidence of the pixel.

[0022] The steps for spatial association and integration include: Project the direct wave candidate target points onto the image plane of the multi-channel hyperspectral infrared imager; For each projected candidate target point, determine its corresponding pixel region on the image plane; Calculate the statistical characteristics of the spectral confidence of all pixels within the pixel region.

[0023] In the step of calculating the fusion confidence of each candidate point, the fusion confidence is a function of at least two of the following factors: the radar echo intensity of the candidate target point, the reflection risk level represented by its reflection path verification result, and the statistical characteristic value of the spectral confidence of its corresponding pixel region.

[0024] The steps to calculate the final three-dimensional coordinates of the fire source include: Select the candidate target point with the highest fusion confidence as the main fire source candidate point and obtain its three-dimensional coordinates; In the purified spectral data, independent high-temperature regions with high spectral confidence that are not associated with high fusion confidence radar points are identified and converted into spatial direction lines originating from the robot. Based on the coordinates of the candidate main fire source points, and combined with the geometric constraints formed by the spatial direction lines, the coordinates of three-dimensional spatial points that satisfy the constraints are solved through an optimization algorithm, which are then used as the final three-dimensional coordinates of the fire source.

[0025] The firefighting robot is also equipped with an ultraviolet flame detector; the method also includes: Acquire signals from the ultraviolet flame detector; The signal strength or its changing trend of the ultraviolet flame detector is used as an auxiliary criterion to verify or alarm the reliability of the final three-dimensional coordinates of the fire source.

[0026] After calculating the final three-dimensional coordinates of the fire source, the method also includes: The final three-dimensional coordinates of the fire source are transformed into the body coordinate system of the fire-fighting robot; Based on the transformed coordinates, control commands are generated to drive the fire-fighting robot's fire extinguishing device to align with the three-dimensional coordinates of the fire source.

[0027] One specific application of this embodiment is: Step S1: Environment initialization and loading of reflector characteristic parameters After the firefighting robot enters the designated patrol area or receives a fire alarm signal, the central processing unit loads a 3D digital map of the current area from the local database; it extracts the surface geometric information of all objects marked as "strong reflectors" on the map, including their spatial coordinates, surface normal vectors, dimensions, and preset reflection coefficients; these objects mainly include stainless steel reactor walls, metal material pipelines, steel platforms, and guardrails; based on the robot's real-time positioning information (from the SLAM system), the processor converts the global coordinates of all reflectors into relative coordinates in the observation coordinate system with the robot itself as the origin, providing a basis for subsequent reflection path calculation.

[0028] Step S2: Synchronous Acquisition and Preprocessing of Raw Data from Multiple Sensors The central processing unit simultaneously triggers the millimeter-wave radar, multi-channel hyperspectral infrared imager, ultraviolet detector and visible light camera to acquire data; Millimeter-wave radar: Acquires the range of multiple suspicious targets within the forward sector. Azimuth and pitch angle ,in The target is indexed, and the echo intensity of each target point is recorded. ; Multi-channel hyperspectral infrared imager: acquires a two-dimensional thermal image, where each pixel corresponds to a spectral vector. , Let k be the radiant intensity value of the k-th spectral channel. The coordinates are pixel coordinates; simultaneously, the processor calculates the average spectrum across the entire field of view. ; Ultraviolet detector: acquiring ultraviolet radiation intensity and its rate of change over time ; Visible light camera: Acquires ambient visible light images, mainly used to assist in identifying device outlines and aid in verification.

[0029] Step S3: Calculation of potential reflection paths and screening of direct wave signals based on prior geometry This step aims to identify and eliminate false target points in radar data that may originate from strong reflectors; Step S31: Direct wave target generation; for each target point detected by the radar... Assuming it is a real fire source (directly reaching the wave source); Step S32: Single reflection path verification; using the robot's position as the launch point. Target point Assuming the source point, traverse all loaded strong reflective planes; for the ... Given a reflection plane, calculate the assumed source point. Regarding the mirror point of this plane According to the principle of the mirror image method, if the signal received by the radar actually comes from the mirror image point... via plane Reflection, then the signal path length ,in, For calculation points (Launch point, i.e., robot position) and point A function of the Euclidean distance between (mirror points); calculate the reflection path length and the radar measured distance. absolute error Set a distance tolerance threshold. (For example, 0.15 meters, considering measurement error); if a reflective plane exists. Make And according to Reflection point The calculated angle of incidence and angle of reflection are approximately equal (within the angular tolerance). If the target point is inside, then determine the target point. The signal height is likely due to reflection; therefore, the risk level of the reflection is marked. This level is related to It is inversely proportional to the reflection coefficient of the reflecting plane; Step S33: Direct wave target screening; integrating all targets and its echo intensity Construct a filtering function; retain those Below the threshold and Higher target points form a preliminary set of radar targets. The core of this step is to use prior environmental geometry to infer the reflection path, thereby directly identifying and reducing the credibility of false targets in radar detection results, rather than simply weighting them in subsequent fusion.

[0030] Step S4: Matching and Dynamic Purification of Interference Spectral Feature Library of Reactants This step targets hyperspectral data to remove spectral interference from reactant vapors; Step S41: Local spectral feature library query; based on the information in the database about chemical substances (such as acrylonitrile, styrene, toluene) that may be stored in this region, retrieve their standard infrared absorption / radiation spectral feature vectors. ,in For material indexing; these feature vectors are measured or simulated calibration data that match the hyperspectral imager channel response of this robot; Step S42: Mixed Spectral Decomposition Model; assuming that the mixed spectrum in a certain region within the field of view is derived from the true flame spectrum. Interference spectrum of reactants and background radiation Linearly mixed; average spectrum for the current frame Solve for the optimal weight coefficients. To minimize the residual of the following expression: In the formula, To use all weight coefficients To optimize the variables, we seek a solution that minimizes the value of the subsequent expression; To calculate the square of the L2 norm of a vector, which is the sum of the squares of all elements of the vector, this is used here to calculate the square of the Euclidean distance of the residual vector, as the fitting error; This is the measured global average spectral vector; A standard clean hydrocarbon flame-based spectral template (pre-calibrated); This represents the total number of possible reactants involved in the calculation. Represents all weight coefficients The sum of absolute values ​​(L1 norm); For the first The standard characteristic spectral vector of a chemical substance; Background radiation estimated based on ambient temperature; The regularization coefficient is used to prevent overfitting; this is a constrained linear optimization problem, and the weight estimates of the interference from each reactant are obtained by solving it. ; Step S43: Pixel-level spectral cleansing; utilizing the estimated global interference weights Spectrum of each pixel The purified suspected flame spectrum was obtained. : In the formula, This is an adjustment factor related to pixel temperature (interference from reactants may be more significant in high-temperature regions, so a value of 1-1.5 may be used; in low-temperature regions, a value of 0.5-1 may be used); after purification, the purification spectrum of each pixel is calculated. Compared with the standard flame basic spectrum cosine similarity This serves as the spectral confidence level that the pixel belongs to a real flame.

[0031] Step S5: Multi-dimensional evidence fusion and calculation of three-dimensional coordinates of the fire source; This step involves spatial correlation and fusion positioning of radar data that has undergone reflection filtering and infrared data that has undergone spectral purification. Step S51: Evidence Space Projection and Correlation; Projecting the initially screened radar target points... Projected onto the hyperspectral infrared image; for each radar target point This corresponds to a projection region on the image; calculate the spectral confidence of all pixels within this projection region. average Spatial variance; Step S52: Fusion credibility calculation; for each radar target point Calculate a fusion credibility score : In the formula, The fusion confidence score represents the overall confidence that the m-th radar candidate target point is determined to be a real fire source, and is a scalar value between 0 and 1. Let represent the current target echo intensity, indicating the echo signal intensity of the m-th target point detected by the millimeter-wave radar; Maximum echo intensity represents the maximum value among all target point echo intensities detected by the radar in the current frame, used for... Perform normalization; The reflection risk level is a non-negative scalar calculated for the m-th target point; Here, it is the natural exponential function. Used to map the reflection risk level to a credibility decay factor between 0 and 1; The average spectral confidence score represents the flame spectral similarity among all pixels within the corresponding projection area of ​​the m-th radar target point in the infrared image. The average value; The slope parameter of the Sigmoid function is a positive scalar used to control the steepness of the spectral confidence transition function; The sigmoid function center threshold is a scalar representing the decision center point of the spectral confidence. Greater than At that time, the output value quickly approaches 1; Let be the spatial variance of the spectral confidence, representing the spectral confidence of all pixels within the projected region corresponding to the m-th target point. The statistical variance is used to measure the consistency of spectral characteristics in a region. The larger the variance, the worse the consistency. (Penalty factor) The smaller; m is the radar target index after filtering, representing a point in the set of radar target points retained after reflection filtering in step S33, where m is an integer; The first part is based on radar evidence (intensity and reflection risk), and the second part is based on the Sigmoid function of infrared spectral confidence. and For calibration parameters), the third part penalizes the spatial inhomogeneity of spectral confidence ( (This refers to the confidence variance of the region). It combines the credibility of radar evidence after reflection screening with the consistency and strength of infrared evidence after spectral purification; Step S53: Determine the main fire source and coordinates; Select highest target point As a candidate for the primary ignition source; simultaneously, examine the hyperspectral images to determine if other spatial regions exist. High but not with any high Radar point association; if present, the center of the high-temperature pixel cluster in that area is inversely calculated as a directional line (ray) in three-dimensional space using the camera calibration model. Step S54: Multi-source geometric intersection localization; the final three-dimensional coordinates of the fire source are determined as follows: based on the main fire source point confirmed by radar. The three-dimensional coordinates are used as the initial solution; if the infrared independent direction line described in S530 exists, then this direction line is used as a constraint, and the least squares method is used to solve for a solution that is simultaneously closest to the given one. And the coordinates of the three-dimensional point located on this direction line If no such independent direction line exists, then directly use... The coordinates are used as the final fire source coordinates. Ultraviolet detector signal strength and its rate of change As a verification of the effectiveness of the overall solution, if If the value is extremely low, an alarm may be triggered indicating a non-open flame leak.

[0032] Step S6: Output the fire source coordinates and align with the fire extinguishing system; The central processing unit will finally calculate the three-dimensional coordinates of the fire source. The coordinates are converted to the robot's base coordinate system, and the robot's gimbal and the fire extinguishing agent (such as water mist or foam) spraying device are driven to precisely align the nozzle axis with the coordinate point; simultaneously, the fire source coordinates and confidence level are converted to the coordinates. and the types of potential reactants identified (based on) Uploaded to the command center.

[0033] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for fire source localization using multi-sensor fusion in a firefighting robot, characterized in that: The firefighting robot is deployed in an environment with strong reflectors and specific interfering substances of known geometry. The robot is equipped with a central processing unit, millimeter-wave radar, a multi-channel hyperspectral infrared imager, and an environmental database. The method includes the following steps: The central processing unit loads the geometric information of strong reflectors pre-stored in the environmental database; Simultaneously acquire raw detection data from the millimeter-wave radar and the multi-channel hyperspectral infrared imager; Based on the geometric information of the strong reflector, the reflection path conformity verification is performed on each target point detected by the millimeter-wave radar, and a set of direct wave candidate target points is selected according to the verification results. The system calls upon the spectral characteristics of specific interfering substances pre-stored in the environmental database to decompose the mixed spectral data acquired by the multi-channel hyperspectral infrared imager, thereby removing interfering spectral components, obtaining purified spectral data, and calculating spectral confidence. The selected direct wave candidate target points are spatially correlated and fused with the purified spectral data and their spectral confidence, and the fusion confidence of each candidate point is calculated. Based on the fusion confidence level, the main fire source candidate is determined, and the final three-dimensional coordinates of the fire source are calculated based on the geometric relationship of the multi-source sensor information.

2. The fire source localization method for a fire-fighting robot using multi-sensor fusion according to claim 1, characterized in that: The step of performing reflection path conformity verification for each target point includes: The robot's position is taken as the signal transmission point, and the radar detection target point is taken as the assumed direct wave source point; Traverse each reflection plane in the geometric information of the strong reflector and calculate the mirror image point of the assumed direct wave source point with respect to each reflection plane; Calculate the reflection path length from the emission point to each of the mirror points and compare it with the target point distance measured by the radar. When a reflecting plane exists such that the difference between the reflection path length and the measured distance is less than a preset tolerance, and the relationship between the incident angle and the reflection angle satisfies the law of reflection, the target point signal is determined to have a suspected risk of reflection.

3. The fire source localization method for a fire-fighting robot using multi-sensor fusion according to claim 1, characterized in that: The steps for decomposing the mixed spectral data include: Query the environmental database for the standard spectral feature vector of at least one known interfering substance related to the current environment; Establish a hybrid spectral model that includes a real flame base spectral template, a linear combination of interfering substance spectra, and background radiation; By solving the optimization problem, the global weight coefficients of the spectra of each interfering substance in the mixed spectrum under the current scenario are estimated. Using the global weighting coefficients, the pixel-level spectral data acquired by the multi-channel hyperspectral infrared imager is corrected, and the estimated interference spectral components are subtracted to obtain the purified pixel spectrum.

4. The fire source localization method for a fire-fighting robot using multi-sensor fusion according to claim 3, characterized in that: The step of calculating the spectral confidence is as follows: calculate the similarity measure between the purified spectrum of each pixel and the pre-stored standard flame basic spectral template, and use the similarity measure value as the spectral confidence of the pixel.

5. The fire source localization method for a fire-fighting robot using multi-sensor fusion according to claim 1, characterized in that: The steps for spatial association and fusion include: The direct wave candidate target points are projected onto the image plane of the multi-channel hyperspectral infrared imager; For each projected candidate target point, determine its corresponding pixel region on the image plane; Calculate the statistical characteristics of the spectral confidence of all pixels within the pixel region.

6. The fire source localization method for a fire-fighting robot using multi-sensor fusion according to claim 5, characterized in that: In the step of calculating the fusion confidence of each candidate point, the fusion confidence is a function of at least two of the following factors: the radar echo intensity of the candidate target point, the reflection risk level represented by its reflection path verification result, and the statistical characteristic value of the spectral confidence of its corresponding pixel region.

7. The fire source localization method for a fire-fighting robot using multi-sensor fusion according to claim 1, characterized in that: The steps for calculating the final three-dimensional coordinates of the fire source include: Select the candidate target point with the highest fusion confidence as the main fire source candidate point and obtain its three-dimensional coordinates; In the purified spectral data, independent high-temperature regions with high spectral confidence that are not associated with high fusion confidence radar points are identified and converted into spatial direction lines originating from the robot. Based on the coordinates of the candidate main fire source points, and combined with the geometric constraints formed by the spatial direction lines, the coordinates of three-dimensional spatial points that satisfy the constraints are obtained through an optimization algorithm, and these coordinates are used as the final three-dimensional coordinates of the fire source.

8. A fire source localization method for a fire-fighting robot using multi-sensor fusion according to any one of claims 1 to 7, characterized in that: The firefighting robot is also equipped with an ultraviolet flame detector; the method further includes: Collect the signal from the ultraviolet flame detector; The signal strength or its changing trend of the ultraviolet flame detector is used as an auxiliary criterion to verify or issue an alarm for the reliability of the final three-dimensional coordinates of the fire source.

9. A multi-sensor fusion fire source localization method for a fire-fighting robot according to claim 1, characterized in that: After calculating the final three-dimensional coordinates of the fire source, the method further includes: Transform the final three-dimensional coordinates of the fire source into the body coordinate system of the fire-fighting robot; Based on the transformed coordinates, control commands are generated to drive the fire-fighting robot's fire extinguishing device to align with the three-dimensional coordinates of the fire source.

Images