Mobile XRF spectrum intelligent analysis system based on deep learning

Abstract

The invention discloses a mobile XRF spectrum intelligent analysis system based on deep learning, and relates to the technical field of analysis and detection. The task scheduling management module is used for comparing a standardized task with each distributed detection module to determine an optimal distributed detection module; the distributed detection module is used for executing the standardized task through autonomous navigation and a mechanical arm of a robot, and acquiring a corresponding X-ray fluorescence spectrum signal; and the intelligent analysis module is used for preprocessing the X-ray fluorescence spectrum signal to obtain a spectrum data packet, and analyzing the spectrum data packet to obtain a corresponding signal analysis result. According to the invention, by dynamically selecting the optimal execution module, efficient distribution of tasks can be ensured, sample positioning and signal acquisition without manual intervention can be realized, dependence on artificial experience is eliminated, and the spatial range and application scene of detection operation are expanded.

Description

Technical Field

[0001] This invention relates to the field of analytical testing technology, specifically to a mobile XRF spectral intelligent analysis system based on deep learning. Background Technology

[0002] X-ray fluorescence (XRF) spectroscopy is a non-destructive elemental detection technique widely used in geological exploration, environmental monitoring, archaeological research, and industrial quality inspection. Its basic principle is to use high-energy X-rays to excite atoms in a sample, causing their inner-shell electrons to transition and release secondary X-rays (i.e., fluorescence) with characteristic energies. By detecting the energy and intensity of these characteristic X-rays, qualitative and quantitative analysis of the types and amounts of elements in the sample can be achieved.

[0003] Chinese invention patent CN114037894A discloses a deep learning-based target detection system, including an X-ray hazardous materials intelligent detection system. This system comprises a hazardous materials detection module and a system management module. By analyzing the working process of commonly used X-ray security inspection equipment, the system assesses the requirements of security equipment for the positioning accuracy, classification accuracy, and real-time performance of hazardous materials detection. Simultaneously, it utilizes X-ray image data collected in actual security scenarios to create and preprocess a dataset. Addressing the needs of existing security inspection systems for hazardous materials detection, the system employs deep learning to aid in hazardous materials identification, researches convolutional neural network technologies, and compares and analyzes existing hazardous materials detection technologies to determine a detection algorithm suitable for X-ray hazardous materials. This security inspection system features an automatic alarm function, reducing manual labor and mitigating the impact of human uncertainty, effectively saving time and human resources while lowering costs.

[0004] However, the above-mentioned and similar technical solutions still have the following shortcomings: Current XRF equipment is mainly handheld or desktop in structure. Handheld equipment relies on manual operation and is difficult to automate detection, while desktop equipment cannot move autonomously. Both types of equipment are difficult to move autonomously and provide on-demand service in complex experimental environments or cross-regional scenarios, which can easily lead to low efficiency in instrument sharing and interruption of cross-space detection processes. At the same time, since current XRF data analysis mostly relies on human experience for judgment, it is difficult to achieve rapid and accurate solutions for multi-material and multi-element spectra. Summary of the Invention

[0005] The purpose of this invention is to provide a mobile XRF spectral intelligent analysis system based on deep learning to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a mobile XRF spectral intelligent analysis system based on deep learning, comprising a task scheduling and management module, a distributed detection module, and an intelligent analysis module, wherein multiple distributed detection modules are configured, and all multiple distributed detection modules interact with the task scheduling and management module and the intelligent analysis module through a communication module; Task scheduling and management module: Based on the user's detection request, it parses and obtains the corresponding standardized task, compares the standardized task with each distributed detection module, and determines the optimal distributed detection module to execute the standardized task; Distributed detection module: It performs autonomous path planning through a set sensor network, and executes the standardized tasks through the robot's autonomous navigation and robotic arm, and collects the corresponding X-ray fluorescence spectral signals; Intelligent analysis module: preprocesses the X-ray fluorescence spectrum signal to obtain the corresponding spectral data package, and analyzes the spectral data package to obtain the corresponding signal analysis results, including element type, content and prediction confidence.

[0007] Furthermore, the optimal distributed detection module for performing the standardized task is determined, including: SA1: Node Filtering: Based on the target location and required detection requirements corresponding to the standardized task, the attribute data and status data of each distributed detection module are filtered to obtain the distributed detection modules that meet the target location and required detection requirements, and a candidate detection set is constructed. SA2: Node Evaluation: Using the set scoring model, obtain the utility score corresponding to each distributed detection module in the candidate detection set, including efficiency score, load score and resource score; SA3: Node Determination: By setting the scoring weights, the efficiency score, load score, and resource score corresponding to each of the distributed detection modules are combined to obtain the corresponding comprehensive utility score. All comprehensive utility scores are compared to determine the maximum comprehensive utility score. The distributed detection module corresponding to the maximum comprehensive utility score is the optimal distributed detection module.

[0008] Furthermore, the attribute data includes the identifier corresponding to the distributed detection module, the model of the XRF spectrometer, the list of functions to be detected, the working radius of the robotic arm, and the list of sensors. The attribute data is compared with the required detection requirements corresponding to the standardized task, and the distributed detection modules corresponding to the attribute data that meet the required detection requirements are selected to construct an initial candidate detection set. The status data includes the current operating status, current battery level, current location coordinates, current task queue, and health status of the distributed detection module. The current operating status includes offline, maintenance, fault, idle, running, and moving. The health status includes normal, sensor fault, and robotic arm fault. Distributed detection modules in the initial candidate detection set whose current operating status is offline, maintenance, or fault, or whose health status is sensor fault or robotic arm fault, are deleted. At the same time, according to the set battery safety threshold, distributed detection modules in the initial candidate detection set whose current battery level is lower than the battery safety threshold are deleted, thus constructing the final candidate detection set.

[0009] Furthermore, the utility score corresponding to each distributed detection module in the candidate detection set is obtained, including: SA2.1: Spatiotemporal Efficiency Assessment: Based on the current task queue of each distributed detection module, obtain the task target position corresponding to each task in the current task queue, and determine the task planning path length between two adjacent tasks in the set global digital map according to the task order corresponding to each task in the current task queue. Based on the movement speed set by the robot in the distributed detection module, determine the movement time corresponding to each task. At the same time, combine the movement time and working time corresponding to each task to determine the total time spent for each task, and determine the corresponding total required time based on the total time spent for each task in the current task queue. SA2.2: Load balancing assessment: The total required time for each of the distributed detection modules is averaged to obtain the corresponding average total time. The load rate is determined based on the ratio between the total required time and the average total time for each of the distributed detection modules. SA2.3: Resource Availability Assessment: Based on the average power consumption set by the robot in the distributed detection module and the task planning path length corresponding to each task, the mobile energy consumption corresponding to each task is determined. At the same time, the mobile energy consumption and working energy consumption corresponding to each task are combined to determine the expected energy consumption corresponding to each task. The expected energy consumption corresponding to each task in the distributed detection module is combined with the total expected energy consumption corresponding to the standardized task to determine the total expected energy consumption corresponding to each distributed detection module. The remaining power consumption is determined by the ratio between the remaining power consumption and the total power consumption.

[0010] Furthermore, based on the total estimated energy consumption and current power consumption of each of the distributed detection modules, the corresponding energy difference is obtained to determine the remaining power consumption of the distributed detection module. At the same time, the remaining power consumption is compared with the power safety threshold, and the distributed detection modules with remaining power consumption below the power safety threshold are deleted, while the distributed detection modules with remaining power consumption not below the power safety threshold are retained.

[0011] Furthermore, the corresponding X-ray fluorescence spectral signals are acquired, including: SB1: Path planning: Using the lidar and depth camera installed on the robot, a 3D point cloud map of the robot's surrounding environment is constructed. Using the inertial measurement unit and SLAM algorithm installed on the robot, the corresponding planned path is determined in the 3D point cloud map, and autonomous navigation is performed. SB2: Robotic arm control: The depth camera at the end of the robotic arm acquires images of the sample area, processes and identifies the images to determine the corresponding preset sample marker or test area, and controls the operation of the robotic arm through a PID controller based on the spatial vector between the current base coordinates of the robotic arm and the preset sample marker or test area. SB3: Signal Acquisition and Detection: Through the robot's autonomous navigation and the operation of the robotic arm, X-ray fluorescence spectral signals of the preset sample marker or the area to be tested are acquired.

[0012] Furthermore, the corresponding signal analysis results are obtained, including: SC1: Edge Pre-processing: The X-ray fluorescence spectrum signal is filtered and denoised to obtain pre-processed X-ray fluorescence energy spectrum data. The pre-processed X-ray fluorescence energy spectrum data and environmental parameters are spliced ​​together to form a spectral data package. At the same time, the spectral data package is used as the input of the lightweight AI analysis model set in the distributed detection module to output the corresponding preliminary analysis results, including element type, content and initial prediction confidence. SC2: Cloud-based reprocessing: Combining convolutional neural network models and attention mechanisms to construct a deep learning spectral analysis model. Simultaneously, the preprocessed X-ray fluorescence energy spectrum data in the spectral data package is used as the input to the deep learning spectral analysis model, and the output is the corresponding signal analysis results, including element types and contents.

[0013] Furthermore, the initial prediction confidence level is compared with a preset confidence threshold, and based on the comparison result, the spectral data packet is divided, including: When the initial prediction confidence level is greater than the preset confidence threshold, the corresponding spectral data packet is transmitted to the user initiator or local display through the communication module and archived on the cloud server; otherwise, the corresponding spectral data packet is transmitted to the cloud server.

[0014] Furthermore, the preprocessed X-ray fluorescence energy spectrum data is used as the input to the first convolutional layer in the convolutional neural network model, and the corresponding feature vector is obtained as the output. At the same time, the feature vector is used as the input to the attention mechanism, and the importance weights of different energy regions in the energy spectrum are obtained as the output. Through a multi-task learning framework, the feature vector and the importance weights are combined to obtain the predicted content corresponding to each element type.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Firstly, this invention intelligently analyzes user needs and dynamically selects the optimal execution module from multiple distributed detection modules, thereby ensuring efficient task allocation. At the same time, through the autonomous navigation of the robot and the precise control of the robotic arm, sample positioning and signal acquisition can be achieved without human intervention, thus eliminating the reliance on human experience and expanding the spatial range and applicable scenarios of detection operations. Secondly, this invention quantifies the total time required, load rate, and remaining power percentage to perform a comprehensive utility score and select the optimal execution module. This not only avoids overload of a single node and achieves load balancing among multiple detection modules, but also improves the utilization rate of the entire system resources and the efficiency of task completion. Thirdly, this invention uses a lightweight AI model to quickly perform preliminary analysis of spectral data and provide prediction confidence levels. Based on the prediction confidence levels, high-confidence results are directly fed back to the user, while low-confidence data is uploaded to the cloud for in-depth analysis. This enables the accurate extraction of spectral features and the allocation of importance weights to different energy regions, achieving high-precision and high-reliability analysis of the types and contents of various elements, effectively overcoming the limitations of traditional methods that rely on manual interpretation. Attached Figure Description

[0016] Figure 1 This is a system block diagram of the mobile XRF spectral intelligent analysis system of the present invention; Figure 2 This is a flowchart illustrating the task scheduling and management method of the present invention; Figure 3 This is a schematic diagram of the process for obtaining a comprehensive utility score in this invention; Figure 4 This is a flowchart illustrating the distributed detection method of the present invention; Figure 5 This is a flowchart illustrating the intelligent analysis method in this invention; Figure 6 This is a schematic diagram of the process for obtaining signal analysis results in this 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] refer to Figure 1 This embodiment provides a mobile XRF spectral intelligent analysis system based on deep learning. The system includes a task scheduling and management module, a distributed detection module, and an intelligent analysis module. Multiple distributed detection modules are configured, and each module connects to the task scheduling and management modules via a communication module (e.g., MQTT, TCP / IP, 4G / 5G) for information exchange. Specifically, the task scheduling and management module parses user detection requests to obtain corresponding standardized tasks. Based on the status, location, and task queue of each distributed detection module, it determines the optimal distributed detection module for each standardized task. Simultaneously, the distributed detection modules execute the standardized tasks determined by the task scheduling and management module. Through a set sensor network, they perform autonomous path planning and, via autonomous navigation and robotic arm control, reach the target location to collect X-ray fluorescence spectral signals. Simultaneously, the intelligent analysis module receives the X-ray fluorescence spectral signals collected by the distributed detection module, and analyzes the received X-ray fluorescence spectral signals through a set deep learning spectral analysis model to obtain the corresponding signal analysis results, including element types, contents, and corresponding prediction confidence levels.

[0019] In this embodiment, by parsing user detection requests, the corresponding standardized task and the optimal distributed detection module for executing the standardized task are determined. (Reference) Figure 2 and Figure 3 This embodiment provides a task scheduling and management method, which specifically includes the following steps: Step SA1: Node Filtering. This involves setting up user detection requests through the interactive interface, parsing these requests, and determining the corresponding standardized tasks, including the target location and required detection criteria. Based on the determined target location and requirements, multiple distributed detection modules are then filtered to select those capable of executing the standardized tasks, forming the candidate detection set.

[0020] Furthermore, in this embodiment, each distributed detection module is equipped with corresponding attribute data and status data. The attribute data includes the identifier of the distributed detection module, the model of the XRF spectrometer it carries, a list of functions to be detected (including at least the equipped vacuum pump or helium purging device), the robotic arm's working radius, and a list of sensors (including at least a depth camera). Simultaneously, the status data in this embodiment includes the current operating status, current battery level, current location coordinates, current task queue, and health status of the distributed detection module.

[0021] In this embodiment, the required detection requirements corresponding to the standardized task are compared with the attribute data of each distributed detection module to determine the attribute data that can meet the required detection requirements. Based on the distributed detection modules corresponding to all the determined attribute data, the corresponding initial candidate detection set is constructed.

[0022] Furthermore, based on the constructed initial candidate detection set, the distributed detection modules in the initial candidate detection set are filtered through the status data corresponding to each distributed detection module and the target location corresponding to the standardized task. Specifically, based on the current operating status of the distributed detection module (including offline, maintenance, fault, idle, running, and moving), distributed detection modules in the offline, maintenance, and fault states are deleted. Simultaneously, based on the health status of the distributed detection module (including normal, sensor fault, and robotic arm fault), distributed detection modules in the sensor fault and robotic arm fault states are deleted. Additionally, based on the current battery level of the distributed detection module, distributed detection modules with a current battery level below the battery safety threshold (which can be specifically set according to actual needs, and is not specifically described in this embodiment) are deleted. In other words, by using the distributed detection modules that have been deleted from the initial candidate detection set, the remaining distributed detection modules that need to be retained in the initial candidate detection set are determined, thereby constructing the final candidate detection set.

[0023] Step SA2: Node Evaluation. Based on the final candidate detection set constructed in Step SA1, the utility score for each distributed detection module in the final candidate detection set is obtained using the set scoring model. Specifically, based on the state data of each distributed detection module in the final candidate detection set and the target location of the standardized task, the efficiency score, load score, and resource score for each distributed detection module are determined. Details are as follows: Step SA2.1: Spatiotemporal efficiency evaluation. This involves determining the target location of each task in the current task queue of each distributed detection module. Simultaneously, based on the task order in the current task queue, a path planning algorithm is used to determine the length of the planned path between two adjacent tasks on the established global digital map. Furthermore, based on the target location of the last task in the current task queue, a path planning algorithm is used to determine the corresponding planned path length on the established global digital map.

[0024] Furthermore, based on the robot's movement speed set in the distributed detection module and the path length planned for each task, the movement time for each task is determined. Simultaneously, the movement time and working time for each task are combined to determine the total time spent on each task.

[0025] In this embodiment, based on the total time spent on each task in the current task queue of the distributed detection module, the total time required for each distributed detection module to process the standardized task set in step SA1 is determined, specifically as follows: in: The total time required. This is the index of the task in the current task queue. This represents the total number of tasks in the current task queue. Let be the movement time of the i-th task. Let be the working time for the i-th task.

[0026] Step SA2.2: Load Balancing Assessment. This involves averaging the total required time for each distributed detection module obtained in Step SA2.1 to obtain the average total required time. In this embodiment, the load rate for each distributed detection module is determined based on the ratio between the total required time and the average total required time. Specifically: in: For load rate, The total time required. This represents the average total time taken.

[0027] Step SA2.3: Resource Availability Assessment. This involves determining the mobility energy consumption for each task based on the average power consumption set for the robot in the distributed detection module and the task planning path length obtained in Step SA2.1. Simultaneously, the working time and set unit-time operation energy consumption coefficient for each task obtained in Step SA2.1 are combined to determine the operational energy consumption for each task. In other words, the operational energy consumption and mobility energy consumption for each task are combined to obtain the estimated energy consumption for each task.

[0028] Furthermore, the estimated energy consumption of each task in the current task queue of each distributed detection module is combined with the estimated energy consumption of the standardized task set in step SA1 to obtain the total estimated energy consumption of each distributed detection module for processing the standardized task set in step SA1.

[0029] In this embodiment, the energy difference between the current power consumption and the total estimated energy consumption for each distributed detection module is obtained, thereby determining the remaining power consumption for each distributed detection module. Simultaneously, the remaining power consumption for each distributed detection module is compared with a set power safety threshold. Distributed detection modules with power consumption below the power safety threshold are identified and directly deleted.

[0030] Furthermore, based on the distributed detection modules corresponding to power safety thresholds, the remaining power of each distributed detection module is determined, and the percentage of remaining power for each distributed detection module is determined based on the ratio between the remaining power and the total power.

[0031] Step SA3: Node Determination. Based on the efficiency score (i.e., total required time), load score (i.e., load rate), and resource score (i.e., remaining power percentage) corresponding to each distributed detection module determined in Step SA2, these scores are combined with the corresponding set score weights to obtain the combined score. Simultaneously, the combined scores are further combined to obtain the comprehensive utility score for each distributed detection module. It is worth noting that the comprehensive utility score at this stage is obtained solely through numerical summation, without considering units. In other words, the units of measurement can be ignored during the specific formula addition calculation.

[0032] Furthermore, the comprehensive utility scores corresponding to each distributed detection module are compared collectively to determine the maximum comprehensive utility score. The distributed detection module corresponding to the maximum comprehensive utility score is the optimal distributed detection module for executing the standardized task set in step SA1.

[0033] In this embodiment, a standardized task set in step SA1 is executed using a configured sensor network and a robotic arm to acquire the corresponding X-ray fluorescence spectral signals. (Reference) Figure 4 This embodiment provides a distributed detection method, which specifically includes the following steps: Step SB1: Path Planning. This involves autonomous path planning using the sensor network installed on the robot. Specifically, the sensor network in this embodiment includes a LiDAR, a depth camera, and an inertial measurement unit (IMU). The LiDAR constructs a 3D point cloud map of the robot's surrounding environment by emitting and receiving laser pulse beams, depicting the contours of each object's surface in the robot's environment. Simultaneously, the depth camera acquires corresponding RGB-D images, including color and depth information. In other words, by acquiring the color and distance information for each object, the 3D point cloud map constructed by the LiDAR is supplemented and enriched. Furthermore, the path planning algorithm determines the corresponding planned path based on the robot's current position coordinates and the target position coordinates on the 3D point cloud map, respectively, according to the standardized task set in Step SA1.

[0034] Furthermore, through the inertial measurement unit, the robot's acceleration and angular velocity are acquired in real time. Then, using the SLAM algorithm, the robot's position and orientation are determined within the constructed 3D point cloud map. In other words, the robot's motion controller, based on the issued operating commands, determines the rotational speed of each drive wheel using a kinematic model, and controls the corresponding drive motors to move, thereby driving the drive wheels. It is worth noting that the control processes of the motion controller, kinematic model, and drive motors in this embodiment are all conventional techniques, and therefore are not specifically described in this embodiment.

[0035] Step SB2: Robotic Arm Control. Following the planned path obtained in Step SB1, the robot is automatically moved to the target position. A depth camera mounted at the end of the robot arm then captures an image of the sample area in front of the robot. Simultaneously, image recognition algorithms (such as template matching, feature point detection, or deep learning models) are used to process and identify the corresponding preset sample markers or areas to be tested.

[0036] Furthermore, through visual recognition and coordinate transformation (it is worth noting that the visual recognition and coordinate transformation in this embodiment are conventional image processing methods, so they are not specifically described in this embodiment), the spatial vector between the current base coordinates of the robotic arm and the detection point on the sample surface is determined from the identified preset sample mark or test area. Simultaneously, the encoder of each joint motor of the robotic arm obtains the angle corresponding to the position of each joint motor. Based on the determined spatial vector and the angle corresponding to the robotic arm joint, the PID controller determines the current or torque command for each joint motor, thereby controlling the operation of the robotic arm. It is worth noting that the PID controller used to regulate the operation of the joint motors in this embodiment is the same as the conventional PID controller operation control method, so it is not specifically described in this embodiment.

[0037] Step SB3: Signal Acquisition and Detection. After the robotic arm completes its positioning in Step SB2, the X-ray fluorescence spectrum signal of the preset sample marker or the area to be tested is acquired using the XRF component mounted on the robot. Specifically, the X-ray tube in the XRF component emits a high-energy primary X-ray beam towards the preset sample marker or the area to be tested, and the preset sample marker or the area to be tested emits characteristic X-ray fluorescence representing the corresponding elemental composition. Simultaneously, the silicon drift detector in the XRF component receives the corresponding characteristic X-ray fluorescence and obtains the corresponding electrical pulse.

[0038] Furthermore, the acquired electrical pulses are analyzed and processed using a pre-set analyzer to obtain the amplitude of each pulse. Based on the amplitude, all electrical pulses are then classified. In other words, based on the classified electrical pulses within a preset time period, the corresponding raw energy spectrum, i.e., the corresponding raw X-ray fluorescence energy spectrum data, is obtained.

[0039] In this embodiment, the raw X-ray fluorescence energy spectrum data obtained in step SB3 is analyzed using a set deep learning spectral analysis model to obtain the corresponding signal analysis results. (Reference) Figure 5 and Figure 6 This embodiment provides an intelligent analysis method, which specifically includes the following steps: Step SC1: Initial edge processing. This involves using a smoothing algorithm, such as Savitzky-Gore filtering, to filter and reduce noise in the raw X-ray fluorescence spectrometry data obtained in step SB3, resulting in preprocessed X-ray fluorescence spectrometry data. Simultaneously, the preprocessed X-ray fluorescence spectrometry data is combined with synchronously acquired environmental parameters (including acquisition time, tube pressure and current, robotic arm pressure, and ambient temperature and humidity) to construct the corresponding spectral data package.

[0040] Furthermore, each distributed detection module is equipped with a corresponding lightweight AI parsing model (such as a CNN and Attention network model), which is used to parse the spectral data packets acquired by the module. Specifically, the spectral data packets acquired by each distributed detection module are initially parsed using the corresponding lightweight AI parsing model.

[0041] In this embodiment, the acquired spectral data packets are used as input to a lightweight AI analysis model, and the output includes the corresponding element types, abundances, and initial prediction confidence levels. Simultaneously, the acquired spectral data packets are divided based on the initial prediction confidence level of each spectral data packet. Specifically, the initial prediction confidence level of each spectral data packet is compared with a preset confidence threshold (which can be set according to actual needs, and is not specifically described in this embodiment), and the division result of the corresponding spectral data packets is determined based on the comparison result, as follows: When the initial prediction confidence level is greater than the preset confidence threshold, the corresponding spectral data packet is directly transmitted to the user initiator or local display via the configured communication module (e.g., MQTT, TCP / IP, 4G / 5G), and can also be uploaded to the cloud server for archiving. Conversely, when the initial prediction confidence level is not greater than the preset confidence threshold, the corresponding spectral data packet is directly transmitted to the cloud server for further analysis using the deep learning spectral analysis model configured on the cloud server.

[0042] Step SC2: Cloud-based reprocessing. This involves combining the convolutional neural network model and the attention mechanism to construct the corresponding deep learning spectral analysis model. Specifically, the convolutional neural network model in this embodiment includes a first convolutional layer, a deep convolutional layer, and a pooling layer. The first convolutional layer has multiple one-dimensional convolutional kernels to receive the preprocessed X-ray fluorescence energy spectrum data. The first convolutional layer and the deep convolutional layer are connected through a pooling layer, and the width of the convolutional kernels in the first convolutional layer is smaller than the width of the convolutional kernels in the deep convolutional layer.

[0043] Furthermore, the feature vectors obtained from the convolutional neural network model are used as input to the attention mechanism model to obtain the importance weights of different energy regions in the energy spectrum.

[0044] In this embodiment, the feature vectors obtained by the convolutional neural network model and the importance weights obtained by the attention mechanism model are combined through a multi-task learning framework (such as a Sigmoid activation function layer) to obtain the predicted content corresponding to each element type.

[0045] In other words, the preprocessed X-ray fluorescence energy spectrum data in the spectral data package is used as the input to the deep learning spectral analysis model, and the output is the corresponding signal analysis results, including element types and contents.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A mobile XRF spectral intelligent analysis system based on deep learning, characterized in that, It includes a task scheduling and management module, a distributed detection module, and an intelligent analysis module. Multiple distributed detection modules are provided, and all of the multiple distributed detection modules interact with the task scheduling and management module and the intelligent analysis module through a communication module. Task scheduling and management module: Based on the user's detection request, it parses and obtains the corresponding standardized task, compares the standardized task with each distributed detection module, and determines the optimal distributed detection module to execute the standardized task; Distributed detection module: It performs autonomous path planning through a set sensor network, and executes the standardized tasks through the robot's autonomous navigation and robotic arm, and collects the corresponding X-ray fluorescence spectral signals; Intelligent analysis module: preprocesses the X-ray fluorescence spectrum signal to obtain the corresponding spectral data package, and analyzes the spectral data package to obtain the corresponding signal analysis results, including element type, content and prediction confidence.

2. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 1, characterized in that, The optimal distributed detection module for performing the standardized task is determined, including: SA1: Node Filtering: Based on the target location and required detection requirements corresponding to the standardized task, the attribute data and status data of each distributed detection module are filtered to obtain the distributed detection modules that meet the target location and required detection requirements, and a candidate detection set is constructed. SA2: Node Evaluation: Using the set scoring model, obtain the utility score corresponding to each distributed detection module in the candidate detection set, including efficiency score, load score and resource score; SA3: Node Determination: By setting the scoring weights, the efficiency score, load score, and resource score corresponding to each of the distributed detection modules are combined to obtain the corresponding comprehensive utility score. All comprehensive utility scores are compared to determine the maximum comprehensive utility score. The distributed detection module corresponding to the maximum comprehensive utility score is the optimal distributed detection module.

3. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 2, characterized in that, The attribute data includes the identifier corresponding to the distributed detection module, the model of the XRF spectrometer, the list of functions to be detected, the working radius of the robotic arm, and the list of sensors. The attribute data is compared with the required detection requirements corresponding to the standardized task, and the distributed detection modules corresponding to the attribute data that meet the required detection requirements are selected to construct an initial candidate detection set. The status data includes the current operating status, current battery level, current location coordinates, current task queue, and health status of the distributed detection module. The current operating status includes offline, maintenance, fault, idle, running, and moving. The health status includes normal, sensor fault, and robotic arm fault. Distributed detection modules in the initial candidate detection set whose current operating status is offline, maintenance, or fault, or whose health status is sensor fault or robotic arm fault, are deleted. At the same time, according to the set battery safety threshold, distributed detection modules in the initial candidate detection set whose current battery level is lower than the battery safety threshold are deleted, thus constructing the final candidate detection set.

4. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 2, characterized in that, Obtain the utility score corresponding to each distributed detection module in the candidate detection set, including: SA2.1: Spatiotemporal Efficiency Assessment: Based on the current task queue of each distributed detection module, obtain the task target position corresponding to each task in the current task queue, and determine the task planning path length between two adjacent tasks in the set global digital map according to the task order corresponding to each task in the current task queue. Based on the movement speed set by the robot in the distributed detection module, determine the movement time corresponding to each task. At the same time, combine the movement time and working time corresponding to each task to determine the total time spent for each task, and determine the corresponding total required time based on the total time spent for each task in the current task queue. SA2.2: Load balancing assessment: The total required time for each of the distributed detection modules is averaged to obtain the corresponding average total time. The load rate is determined based on the ratio between the total required time and the average total time for each of the distributed detection modules. SA2.3: Resource Availability Assessment: Based on the average power consumption set by the robot in the distributed detection module and the task planning path length corresponding to each task, the mobile energy consumption corresponding to each task is determined. At the same time, the mobile energy consumption and working energy consumption corresponding to each task are combined to determine the expected energy consumption corresponding to each task. The expected energy consumption corresponding to each task in the distributed detection module is combined with the total expected energy consumption corresponding to the standardized task to determine the total expected energy consumption corresponding to each distributed detection module. The remaining power consumption is determined by the ratio between the remaining power consumption and the total power consumption.

5. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 4, characterized in that, Based on the total estimated energy consumption and current power consumption of each distributed detection module, the corresponding energy difference is obtained to determine the remaining power consumption of the distributed detection module. At the same time, the remaining power consumption is compared with the power safety threshold, and the distributed detection module with remaining power consumption below the power safety threshold is deleted, while the distributed detection module with remaining power consumption not below the power safety threshold is retained.

6. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 1, characterized in that, Acquire the corresponding X-ray fluorescence spectral signals, including: SB1: Path planning: Using the lidar and depth camera installed on the robot, a 3D point cloud map of the robot's surrounding environment is constructed. Using the inertial measurement unit and SLAM algorithm installed on the robot, the corresponding planned path is determined in the 3D point cloud map, and autonomous navigation is performed. SB2: Robotic arm control: The depth camera at the end of the robotic arm acquires images of the sample area, processes and identifies the images to determine the corresponding preset sample marker or test area, and controls the operation of the robotic arm through a PID controller based on the spatial vector between the current base coordinates of the robotic arm and the preset sample marker or test area. SB3: Signal Acquisition and Detection: Through the robot's autonomous navigation and the operation of the robotic arm, X-ray fluorescence spectral signals of the preset sample marker or the area to be tested are acquired.

7. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 1, characterized in that, The corresponding signal analysis results obtained include: SC1: Edge Pre-processing: The X-ray fluorescence spectrum signal is filtered and denoised to obtain pre-processed X-ray fluorescence energy spectrum data. The pre-processed X-ray fluorescence energy spectrum data and environmental parameters are spliced ​​together to form a spectral data package. At the same time, the spectral data package is used as the input of the lightweight AI analysis model set in the distributed detection module to output the corresponding preliminary analysis results, including element type, content and initial prediction confidence. SC2: Cloud-based reprocessing: Combining convolutional neural network models and attention mechanisms to construct a deep learning spectral analysis model. Simultaneously, the preprocessed X-ray fluorescence energy spectrum data in the spectral data package is used as the input to the deep learning spectral analysis model, and the output is the corresponding signal analysis results, including element types and contents.

8. The mobile XRF spectral intelligent analysis system based on deep learning according to claim 7, characterized in that, The initial prediction confidence level is compared with a preset confidence threshold, and the spectral data packet is divided according to the comparison result, including: When the initial prediction confidence level is greater than the preset confidence threshold, the corresponding spectral data packet is transmitted to the user initiator or local display through the communication module and archived on the cloud server; otherwise, the corresponding spectral data packet is transmitted to the cloud server.

9. A mobile XRF spectral intelligent analysis system based on deep learning according to claim 7, characterized in that, The preprocessed X-ray fluorescence energy spectrum data is used as the input to the first convolutional layer in the convolutional neural network model, and the output obtains the corresponding feature vector. At the same time, the feature vector is used as the input to the attention mechanism, and the output obtains the importance weights of different energy regions in the energy spectrum. Through a multi-task learning framework, the feature vector and the importance weights are combined to obtain the predicted content corresponding to each element type.

Images