A lithology intelligent identification method and system based on an edge-cloud collaborative architecture

Abstract

The application provides a lithology intelligent identification method and system based on an edge-cloud collaborative architecture, and belongs to the technical field of lithology identification, and comprises the following steps: collecting rock spectrum data and performing pretreatment; a light spectrum analysis model is used in an edge layer to perform preliminary analysis on the pretreated spectrum data, a preliminary lithology identification result is obtained, and it is judged whether the cloud layer needs to be uploaded according to the confidence degree corresponding to the preliminary identification result; if the cloud layer does not need to be uploaded, the preliminary lithology identification result is directly output; if the cloud layer needs to be uploaded, a communication layer judges network classification according to round trip time, bandwidth and packet loss rate, and uploads the spectrum data or the pretreated spectrum data to the cloud layer according to the network classification; a cloud lithology precision analysis model is used in the cloud layer to perform reanalysis on the spectrum data, and accurate element quantitative analysis results, mineral composition and a lithology identification report are obtained. The application realizes rapid identification and analysis of on-site rock chemical components through spectrum data analysis.

Description

Technical Field

[0001] This invention belongs to the field of lithology identification technology, and in particular relates to a lithology intelligent identification method and system based on an edge-cloud collaborative architecture. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, rapid and accurate on-site analysis of lithology is crucial in deep-earth engineering projects such as high-depth tunnel construction and mineral exploration. Traditional lithology identification methods heavily rely on visual observation, microscopic analysis, or laboratory geochemical testing by geological experts. While these methods are accurate, they have significant drawbacks: experts cannot be stationed permanently in the complex tunnel environment with high temperatures and pressures; and laboratory testing is time-consuming, failing to meet the stringent time requirements of construction sites. Therefore, developing technology capable of rapid and accurate on-site rock composition analysis is key to achieving intelligent and unmanned construction, improving efficiency and safety.

[0004] With the development of sensor technology and artificial intelligence, spectral analysis technology has provided a new data acquisition method for on-site lithological analysis. However, traditional spectral analysis methods are often relatively simple, and their accuracy and reliability decrease significantly when faced with complex mineral assemblages or rare lithologies. Furthermore, transmitting large amounts of collected spectral data to the cloud for in-depth analysis in real time is limited by practical conditions such as poor network signals and unstable bandwidth at construction sites, resulting in high analysis latency and a poor user experience. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention proposes a lithological intelligent identification method and system based on an edge-cloud collaborative architecture. It proposes a collaborative paradigm of "rapid preliminary analysis at the edge + precise analysis in the cloud" to achieve both low latency and high accuracy. It also designs an intelligent upload decision mechanism that adaptively selects the upload strategy based on confidence level, element anomaly signals, and network status.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, this invention discloses a lithological intelligent identification method based on an edge-cloud collaborative architecture, comprising: Collect rock spectral data and preprocess it; A lightweight spectral analysis model is used in the edge layer to perform preliminary analysis on the preprocessed spectral data to obtain preliminary lithology identification results. The confidence level of the preliminary identification results is used to determine whether it is necessary to upload to the cloud layer. If it is not necessary to upload, the preliminary lithology identification results are directly output. If uploading is required, the communication layer determines the network classification based on round-trip time, bandwidth, and packet loss rate, and uploads the spectral data or preprocessed spectral data to the cloud according to the network classification. The cloud-based lithology precision analysis model is used to re-analyze the spectral data or the preprocessed spectral data in the cloud layer to obtain accurate elemental quantitative analysis results, mineral composition and lithology identification report; wherein, the cloud-based lithology precision analysis model adopts a deep neural network structure and integrates an attention mechanism to enhance the model's ability to interpret complex spectral features.

[0007] Secondly, this invention discloses a lithological intelligent identification system based on an edge-cloud collaborative architecture, comprising: The data acquisition module is used to collect rock spectral data and perform preprocessing. The edge recognition module is used to perform preliminary analysis on the preprocessed spectral data at the edge layer using a lightweight spectral analysis model to obtain preliminary lithology identification results. Based on the confidence level corresponding to the preliminary identification results, it determines whether to upload to the cloud layer. If no upload is required, the preliminary lithology identification results are directly output. The communication transmission module is used to determine the network classification based on round-trip time, bandwidth and packet loss rate if uploading is required, and upload the spectral data or preprocessed spectral data to the cloud according to the network classification. The cloud-based identification module is used to re-analyze the spectral data or pre-processed spectral data in the cloud using a cloud-based lithology precision analysis model to obtain accurate elemental quantitative analysis results, mineral composition, and lithology identification reports. The cloud-based lithology precision analysis model adopts a deep neural network structure and integrates an attention mechanism to enhance the model's ability to interpret complex spectral features.

[0008] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the above-mentioned rock type intelligent identification method based on edge-cloud collaborative architecture.

[0009] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-mentioned intelligent rock identification method based on an edge-cloud collaborative architecture.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires rock spectral data on-site using spectral data acquisition equipment; deploys a lightweight analysis model on edge devices for rapid preliminary identification; and performs precise analysis and big data comparison on a cloud platform, ensuring the system meets the requirements of high precision and high reliability while adapting to harsh on-site environments. The collaborative paradigm of "rapid preliminary analysis at the edge + precise analysis in the cloud" proposed in this invention achieves a balance between low latency and high precision.

[0011] This invention introduces an intelligent decision-making module that monitors and analyzes confidence level, element anomaly indicators, and network bandwidth in real time, driving data to be uploaded on demand, reducing the amount of data uploaded, and improving processing speed.

[0012] This invention proposes a model feedback and evolution mechanism, enabling continuous learning in the cloud and online distillation updates at the edge, thus giving the system self-evolution capabilities.

[0013] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0015] Figure 1 This is a flowchart of the lithology intelligent identification method based on an edge-cloud collaborative architecture as described in Embodiment 1 of the present invention. Detailed Implementation

[0016] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0017] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0018] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0019] Example 1 In one or more embodiments, a lithology intelligent identification method based on an edge-cloud collaborative architecture is disclosed, such as... Figure 1 As shown, it includes the following steps: Step S1: Collect rock spectral data and perform preprocessing.

[0020] Step S1-1: Collect spectral data of the rock using a portable visible-near infrared spectrometer to obtain raw spectral data; Step S1-2: Preprocess the raw spectral data, including dark current correction, whiteboard normalization, noise smoothing, and identification of characteristic spectral peaks. Dark current correction involves measuring the sensor response against a blank background and then applying the correction.

[0021] In the formula, Sensor response against a blank background; This is the raw spectral data; This is the spectral signal after dark current correction; λ represents the wavelength of the spectrum.

[0022] Whiteboard normalization: Normalization is performed based on a standard reflective whiteboard.

[0023] In the formula, A standard reflective whiteboard; This is the normalized reflectance spectrum.

[0024] Noise smoothing: Savitzky-Golay smoothing filter (window length 7, third-order polynomial) is used on the normalized data to remove high-frequency noise.

[0025] Feature spectral peak identification: Peak values ​​are detected by the zero intersection of the first derivative of the smoothed data. The peak value must be higher than three times the standard deviation of the neighborhood average.

[0026] Step S2: Use a lightweight spectral analysis model to perform preliminary analysis on the preprocessed spectral data in the edge layer to obtain preliminary lithology identification results. Determine whether to upload to the cloud layer based on the confidence level corresponding to the preliminary identification results. If no upload is required, output the preliminary lithology identification results directly.

[0027] The edge layer uses portable spectral acquisition all-in-one machines, industrial tablet terminals, or embedded edge computing boxes to achieve local acquisition, preprocessing, and rapid identification of spectral data.

[0028] Step S2-1: Construct a lightweight spectral analysis model. The model adopts a hybrid structure of CNN feature encoder, LightGBM discriminator and fusion layer to simultaneously combine the feature representation ability of CNN with the high robustness and high speed of LightGBM. The number of parameters in this hybrid model is less than 2MB.

[0029] The model first uses a lightweight 1D-CNN to extract a 128-dimensional spectral depth feature vector, and then inputs the feature into the LightGBM multi-task discriminator to achieve element identification, element content estimation and preliminary lithology classification. Subsequently, the output vectors of CNN and LightGBM are weighted and fused through a fusion layer to improve the overall recognition accuracy and model stability.

[0030] The input to the lightweight spectral analysis model is a preprocessed spectral feature vector (512 dimensions); the output is a preliminary lithology identification result, including lithology classification results and an element identification probability list. The lithology classification result includes a preliminary lithology classification and the confidence level of the classification, and the element identification probability list includes element types and their probabilities, as well as element content estimates.

[0031] At the edge layer, this embodiment designs a lightweight one-dimensional convolutional neural network (1D-CNN) model based on a multi-task learning (MTL) architecture. This model shares the underlying feature extractor and has multiple specific output branches to simultaneously complete element identification, content estimation, and lithological classification tasks. After pruning and 8-bit quantization compression, the model is less than 2MB in size, making it suitable for deployment on edge devices.

[0032] A lightweight one-dimensional convolutional neural network consists of an input layer, a shared feature extractor, and a multi-task output layer connected in sequence, specifically: Input Layer: Receives a 512-dimensional spectral feature vector after preprocessing and feature extraction.

[0033] The Shared Feature Extractor is the core computational part of the model. All tasks share the common spectral features extracted by this part, which greatly reduces the total number of parameters and computational cost of the model. It consists of a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a flattening layer connected in sequence. The first convolutional layer (Conv1D) has 16 filters, a kernel length of 5, and uses the ReLU activation function. The output size is (508, 16).

[0034] First pooling layer (MaxPool1D): Pooling size = 2, stride = 2. Output size is approximately (254, 16).

[0035] The second convolutional layer (Conv1D) has 32 filters, a kernel length of 5, and uses the ReLU activation function. The output size is (250, 32).

[0036] Second pooling layer (MaxPool1D): Pooling size = 2, stride = 2. Output size is approximately (125, 32).

[0037] Flatten layer: Flattens the two-dimensional feature map into a one-dimensional vector with a size of (125*32=4000).

[0038] Multi-task output heads are a key special design for achieving multiple outputs. After sharing features, the network branches into three independent fully connected branches, each responsible for learning task-specific features.

[0039] The first branch at the edge is used for element type identification (classification task), which includes the first fully connected layer (Dense, 256 units + ReLU), the first Dropout layer (dropout rate 0.2), and the first fully connected output layer connected in sequence.

[0040] In this design, the number of neurons in the first fully connected output layer is equal to the number of element types N to be identified (e.g., N = 20 common elements), and the Softmax activation function is used. Each neuron represents the probability of the corresponding element's existence. The top three probabilities are selected as the top three element types and their probabilities, which can be used for result display.

[0041] The second branch at the edge is used for element content estimation (regression task) and consists of a second fully connected layer (Dense, 128 units + ReLU), a second Dropout layer (dropout rate 0.2), and a second fully connected output layer connected in sequence.

[0042] The second fully connected output layer also has N neurons (corresponding to the number of element types) and uses a linear activation function. Each neuron outputs an estimated content (e.g., mass percentage) of the corresponding element. This output value is synchronized with the first branch, meaning that only elements with higher probabilities in the first branch are output with meaningful content values.

[0043] The third branch at the edge is used for lithological classification (classification task), which includes a third fully connected layer (Dense, 256 units + ReLU), a third Dropout layer (dropout rate 0.2), and a third fully connected output layer connected in sequence.

[0044] In this system, the number of neurons in the third fully connected output layer is equal to the number of rock types to be identified, M (e.g., M = 10 common rock types), and the Softmax activation function is used. Each neuron represents the probability of the corresponding rock type, and the one with the highest probability is taken as the preliminary rock type classification result. Its probability value is the confidence level of the result.

[0045] In this embodiment, element identification is not achieved through a traditional spectral "peak-finding algorithm." Instead, it is achieved by querying the output probability of the "element type identification" branch of the edge model itself. The output vector of this branch contains the probability of occurrence of all elements seen during model training. During a single forward propagation, the edge model simultaneously outputs the lithological classification result and the element identification probability list. The system only needs to check whether the probability value corresponding to a preset rare element in this readily available probability list exceeds a certain threshold.

[0046] Preferably, model compression and deployment are performed: After training, the model undergoes pruning (removing unimportant neuron connections) and 8-bit integer quantization (converting 32-bit floating-point weights into 8-bit integer representations). These two techniques significantly reduce model size and inference time, compressing it from a potentially original ~10MB to less than 2MB while maintaining almost the same accuracy, thus meeting the stringent resource constraints of edge devices.

[0047] As an alternative implementation, the LightGBM ensemble decision tree model can also be used (which achieves similar functionality through multi-objective regression and classification settings).

[0048] Step S2-2: Based on the preliminary lithological classification and its confidence level, element types, and probability list, determine whether the data needs to be uploaded to the cloud for cloud-based assessment. If no upload is required, directly output the preliminary lithological identification results. Specifically: From the edge lightweight model, two outputs are obtained: preliminary lithological classification and the confidence level of the classification C-rock (i.e., the maximum classification probability), and element types and their probability list P-elements.

[0049] Determining whether to upload to the cloud based on rare element signals and confidence levels: If the confidence level is greater than or equal to the first confidence threshold and no rare element signal is detected, the edge analysis result is considered reliable enough, and the result can be directly output without uploading. The first confidence threshold can be set to 0.8.

[0050] If the confidence level is less than the second confidence threshold or a rare element signal is detected, the upload process is triggered, and the data is sent to the cloud for precise analysis. If the confidence level is between the two confidence levels and no rare element signal is detected, the previous state is maintained. The second confidence level can be set to 0.85. Depending on the network status, the uploaded data can be spectral data, compressed preprocessed data, or abnormal spectral bands extracted from preprocessed spectral data. This invention uses a network status monitoring module at the edge layer to determine network bandwidth in real time and select the most suitable data upload method to balance data integrity and transmission efficiency.

[0051] The types of data uploaded vary depending on network conditions. Two thresholds are used to form a "hysteresis interval" to prevent the system from fluctuating repeatedly. The upload threshold (second confidence threshold) must be higher than the no-upload threshold (first confidence threshold), otherwise a stable interval cannot be formed.

[0052] The determination of rare element signals is as follows: check whether the probability of any element in the preset rare element list (such as La, Ce, Nd, Pt, Au, etc.) exceeds the rare element probability threshold T-rare (e.g., T-rare=0.3). If the recognition probability of the rare element is not less than the rare element probability threshold, then a rare element signal is detected; if it is less than the threshold, then a rare element signal is not detected.

[0053] Step S3: If uploading is required, the communication layer determines the network classification based on the round-trip time, bandwidth and packet loss rate, and uploads the spectral data or preprocessed spectral data to the cloud according to the network classification; otherwise, the original spectral data or preprocessed spectral data is uploaded to the cloud. If uploading is required, the communication layer will upload the spectral data or preprocessed spectral data to the cloud.

[0054] Specifically, the communication layer uses three metrics—Round-Trip Time (RTT), bandwidth, and packet loss rate—measured in real-time based on heartbeat packets to classify the current network as good, weak, or extremely weak according to preset network classification rules. Under good network conditions (bandwidth > 2 Mbps), the original complete spectral data is uploaded; under weak network conditions (0.5–2 Mbps or RTT ≥ 150 ms), PCA is used to reduce the dimensionality of the preprocessed spectral data to a 50-dimensional feature vector before uploading; under extremely weak network conditions (bandwidth < 0.5 Mbps or RTT ≥ 300 ms), only abnormal spectral band data is uploaded. The abnormal spectral bands are obtained by extracting the bands of the ±10 channels of the characteristic spectral peaks identified in steps S1-2.

[0055] Preferably, the communication uses a combination of CoAP and QUIC protocols, with built-in AES-256 encryption to ensure security.

[0056] Step S4: The cloud-based lithology precision analysis model is used to re-analyze the spectral data, high-dimensional feature vectors, or anomalous spectral bands in the cloud layer to obtain accurate elemental quantitative analysis results, mineral composition, and lithology identification reports. The cloud-based lithology precision analysis model adopts a deep neural network structure and integrates an attention mechanism to enhance the model's ability to interpret complex spectral features.

[0057] The cloud-based lithology precision analysis model adopts a 1D-CNN+Transformer hybrid architecture (64 / 128 / 256 channels of convolutional layers + 4 Transformer layers, 8 attention heads, and 512 hidden layers); it integrates a self-attention mechanism to automatically focus on rare element bands; and outputs element mass percentages, mineral composition lists, lithology identification reports, and confidence levels accurate to 0.001.

[0058] The 1D-CNN+Transformer hybrid architecture is a deep hybrid neural network based on multi-task learning designed for high-precision spectral analysis. It combines 1D-CNN, which excels at extracting local features, with Transformer, which excels at modeling long-range dependencies.

[0059] First, the model's input consists of spectral data or feature vectors uploaded from edge devices. The cloud-based model is designed with a multi-modal input structure. Upon receiving the raw spectrum, the cloud first performs standardization preprocessing. The preprocessed raw spectrum is then input into the 1D-CNN+Transformer hybrid architecture as a one-dimensional sequence with a uniform scale through the first input branch. When PCA compressed features are received, they are processed through the second input branch. Abnormal spectral segments received in cases of extremely weak networks are processed by the third input branch.

[0060] Specifically, the first input branch receives the raw spectral data and maps it to a fixed-dimensional sequence feature space of length S × channel dimension C, matching the hybrid structure, through a first feature mapping layer. The second input branch receives PCA compressed features, activates the PCA feature mapping branch model, and projects the low-dimensional features to a feature space consistent with the backbone network through a second feature mapping layer. These features are then fed into a shared 1D-CNN+Transformer hybrid architecture for recognition. The third input branch receives anomalous spectral segments and uses a local peak enhancement branch model to strengthen key spectral peak features. The enhanced features are then aligned and fed into the shared 1D-CNN+Transformer hybrid architecture for recognition. The local peak enhancement branch model includes a spectral segment alignment layer, a local one-dimensional convolutional feature extraction layer, an attention enhancement layer, and a third feature mapping layer. The spectral segment alignment layer performs preprocessing on the spectral data (interpolation, truncation, or padding) to unify the spectral data scale. It extracts local features of anomalous spectral segments through small-scale one-dimensional convolution, then uses an attention mechanism to enhance the key spectral peak response. The enhanced local features are dimensionally aligned through the feature mapping layer and finally fed into the shared 1D-CNN+Transformer architecture as sequence input. The hybrid architecture completes the subsequent identification.

[0061] Secondly, the 1D-CNN+Transformer hybrid architecture includes three main stages: local feature extraction, global context modeling, and multi-task output.

[0062] Phase 1 (Local Feature Extraction Phase): The 1D-CNN local feature extractor acts as a high-level feature extractor, converting the original spectral signal into a feature map rich in semantic information and gradually reducing the sequence length to alleviate the burden on the subsequent Transformer module, which has a larger computational load.

[0063] Specifically, it includes a third convolutional layer, a first max-pooling layer, a fourth convolutional layer, a second max-pooling layer, and a fifth convolutional layer connected in sequence.

[0064] The third convolutional layer (Conv1D): 64 filters, kernel size = 5, stride = 1, using the ReLU activation function. Output size: (L-4, 64).

[0065] First max pooling layer (MaxPooling1D): Pooling window size = 2, stride = 2.

[0066] Fourth convolutional layer (Conv1D): 128 filters, kernel size = 5, stride = 1, ReLU activation.

[0067] Second Max Pooling (MaxPooling1D): Pooling window size = 2, stride = 2.

[0068] Fifth convolutional layer (Conv1D): 256 filters, kernel size = 5, stride = 1, ReLU activation.

[0069] After these three convolutional and pooling layers, the original long sequence L is significantly shortened to a new sequence length S, but the feature dimension of each point is increased to 256. At this point, the data structure changes from (L,1) to a feature matrix Z of (S,256).

[0070] Phase Two (Global Context Modeling Phase): Transformer Global Context Modeling (Integrating Self-Attention Mechanism) Specifically, the feature matrix Z∈R^(S×256) output by the 1D-CNN local feature extractor is considered as a sequence of length S. To compensate for the positional information lost by the CNN, positional encoding needs to be added to this sequence to obtain the encoded features. This is the key step connecting the CNN and the Transformer, expressed as: H0 = Z + PositionalEncoding In the formula, H0 is the initial Transformer input sequence feature matrix after incorporating positional information; PositionalEncoding is the positional information encoding of the feature sequence, with the same dimension as the feature matrix Z, used to compensate for the sequence positional information lost by the convolutional network during feature extraction. Positional encoding can be a fixed-positional encoding method constructed using sine and cosine functions, i.e., generating a set of sine and cosine values ​​for each sequence position based on the channel dimension of Z; alternatively, a trainable positional embedding method can be used, predefining a learnable vector for each position and updating its parameters during model training. H0, obtained by element-wise addition of the positional encoding and the feature matrix Z, is used as the input to the Transformer encoder.

[0071] The Transformer encoder consists of four stacked encoder layers. Each encoder layer includes a multi-head self-attention layer, a residual connection and layer normalization layer, a feedforward neural network, and a residual connection and layer normalization layer connected in sequence. The input sequence H0 is fed into a multi-head self-attention layer. Eight attention heads work in parallel, each calculating the association weights of each feature point in the sequence with all other points from a different perspective. This allows the model to automatically focus on key bands relevant to the task (e.g., a head may have specifically learned to identify weak feature peaks of rare earth elements).

[0072] Residual Connections and Layer Normalization (Add & Norm): The output of the attention layer is added to the original input H0 (residual connection), and then layer normalization is performed to stabilize the training process.

[0073] A feed-forward network is a multilayer perceptron structure containing two fully connected layers. The first fully connected layer maps the input dimension from 256 to 512 and uses ReLU or GELU nonlinear activation. The second fully connected layer then maps the dimension back from 512 to 256 so as to perform residual connections with the input of the previous sublayer and perform nonlinear transformations on the features at each location.

[0074] The residual connection and layer normalization are performed again. The output FFN(H1) of the feedforward sub-layer is residually connected with its input H1 and then normalized to obtain H2, so as to stabilize the gradient and enhance the training capability of deep networks.

[0075] After processing by four or more encoder layers, the model outputs a deeply processed feature sequence H_out ∈ R^(S × 256), which incorporates global spectral context information.

[0076] Phase 3 (Multi-task Output Phase) includes a multi-task output layer. This layer is not generated from a simple output layer, but rather through three independent task-specific heads. This is a significant improvement over a single output layer, allowing the model to specialize in each task.

[0077] Specifically, the output H_out of the Transformer is first subjected to global average pooling to obtain a comprehensive 256-dimensional global feature vector.

[0078] The first branch in the cloud: Quantitative regression of elements, which includes a fully connected layer and a linear activation function layer. The output dimension is the number of element types N. Using a linear activation function, it directly outputs the percentage of element quality accurate to 0.001.

[0079] The second branch in the cloud: mineral composition classification, which includes a fully connected layer and a sigmoid activation function. The output dimension is the number of mineral types M. Using the sigmoid activation function (multi-label classification), it outputs the probability of the existence of each mineral.

[0080] The third branch in the cloud: lithology identification, which includes a fully connected layer and a Softmax activation function. The output dimension is the number of lithology types K. Using the Softmax activation function, the final lithology classification and confidence level are output.

[0081] This embodiment utilizes a model with high-precision analytical capabilities: the 1D-CNN-Transformer hybrid architecture combines the advantages of local feature extraction and global dependency modeling, enabling it to capture complex and weak spectral signals (such as rare elements), thus achieving high-precision (0.001 level) quantitative analysis of elemental content. It exhibits excellent robustness: the self-attention mechanism allows the model to make judgments based on the contextual information of the entire spectrum, rather than looking at certain bands in isolation, thus providing stronger resistance to noise, baseline drift, and other interference. It achieves efficient multi-task learning: the shared feature extraction backbone network combined with discrete output branches allows the model to simultaneously and efficiently complete the three key tasks of elemental, mineral, and lithological analysis, with complementary knowledge between tasks, improving overall performance.

[0082] This embodiment is specifically optimized for the sequential characteristics of spectral data. The introduction of Transformer greatly enhances the ability to model long-range spectral correlations, which is difficult to achieve with traditional CNN models or pure sequence models.

[0083] Preferably, the loss function of the 1D-CNN+Transformer hybrid architecture is jointly optimized using cross-entropy and L1 regression loss, and the expression is:

[0084] In the formula, The loss function; For classification loss; For regression loss; For classification loss weights; The regression loss weights.

[0085] The classification loss expression is as follows:

[0086] In the formula, Loss due to lithological classification; This represents a loss of mineral composition. The expression for lithological classification loss is:

[0087] In the formula, This is the true label for lithology (one-hot encoding; 1 if it belongs to this category, 0 otherwise); This represents the probability of the lithology predicted by the model.

[0088]

[0089]

[0090] In the formula, A true label for a mineral or element (present / absent, or one-hot encoded); The probability of a mineral or element predicted by the model; This represents the number of mineral types.

[0091] The regression loss expression is:

[0092] In the formula, N is the number of element types; This represents the true value of the element content. These are the model predictions for element content.

[0093] Furthermore, to enhance the long-term stability and prediction accuracy of the edge-cloud collaborative architecture, this invention introduces a model feedback and evolution mechanism, enabling the system to automatically accumulate experience in multiple work areas and scenarios, and achieve continuous self-evolution. This mechanism includes five stages: (1) data feedback, (2) quality screening, (3) incremental learning in the cloud, (4) knowledge distillation and distribution, and (5) adaptive optimization of the edge model.

[0094] Example 2 In one or more embodiments, a lithology intelligent identification system based on an edge-cloud collaborative architecture is disclosed, specifically including: The data acquisition module is used to collect rock spectral data and perform preprocessing. The edge recognition module is used to perform preliminary analysis on the preprocessed spectral data at the edge layer using a lightweight spectral analysis model to obtain preliminary lithology identification results. Based on the confidence level corresponding to the preliminary identification results, it determines whether to upload to the cloud layer. If no upload is required, the preliminary lithology identification results are directly output. The communication transmission module is used to determine the network classification based on round-trip time, bandwidth and packet loss rate if uploading is required, and upload the spectral data or preprocessed spectral data to the cloud according to the network classification. The cloud-based identification module is used to re-analyze the spectral data or pre-processed spectral data in the cloud using a cloud-based lithology precision analysis model to obtain accurate elemental quantitative analysis results, mineral composition, and lithology identification reports. The cloud-based lithology precision analysis model adopts a deep neural network structure and integrates an attention mechanism to enhance the model's ability to interpret complex spectral features.

[0095] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-described intelligent rock identification method based on an edge-cloud collaborative architecture.

[0096] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the aforementioned rock type intelligent identification method based on an edge-cloud collaborative architecture.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0101] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A lithology intelligent identification method based on an edge-cloud collaborative architecture, characterized in that, include: Collect rock spectral data and preprocess it; A lightweight spectral analysis model is used in the edge layer to perform preliminary analysis on the preprocessed spectral data to obtain preliminary lithology identification results. Based on the confidence level corresponding to the preliminary identification results, it is determined whether it is necessary to upload to the cloud layer. If no upload is required, the preliminary lithology identification results will be output directly. If uploading is required, the communication layer determines the network classification based on round-trip time, bandwidth, and packet loss rate, and uploads the spectral data or preprocessed spectral data to the cloud according to the network classification. The communication layer uses the three indicators of round-trip time, bandwidth, and packet loss rate measured by heartbeat packets in real time to determine the current network as a good network, a weak network, or an extremely weak network according to preset network classification rules. Under good network conditions, the original complete spectral data is uploaded. Under weak network conditions, PCA is used to reduce the dimensionality of the preprocessed spectral data before uploading. Under extremely weak network conditions, only abnormal spectral band data is uploaded. Abnormal spectral bands are obtained by extracting the bands of the ±10 channels of the identified characteristic spectral peaks. The cloud-based lithology precision analysis model is used to re-analyze the spectral data or the preprocessed spectral data in the cloud layer to obtain accurate elemental quantitative analysis results, mineral composition and lithology identification report; wherein, the cloud-based lithology precision analysis model adopts a deep neural network structure and integrates an attention mechanism to enhance the model's ability to interpret complex spectral features.

2. The lithology intelligent identification method based on the edge-cloud collaborative architecture according to claim 1, characterized in that, The lightweight spectral analysis model employs a lightweight one-dimensional convolutional neural network, comprising an input layer, a shared feature extractor, and a multi-task output layer connected in sequence. The input layer receives a preprocessed spectral feature vector; The shared feature extractor includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a flattening layer connected in sequence. The multi-task output layer enables various outputs. 3.The lithology intelligent identification method based on the edge-cloud collaborative architecture according to claim 2, characterized in that, The multi-task output layer branches into three independent fully connected branches: The first branch at the edge is used for element type identification, including the first fully connected layer, the first Dropout layer, and the first fully connected output layer connected in sequence; The second edge branch is used for element content estimation, and includes a second fully connected layer, a second Dropout layer, and a second fully connected output layer connected in sequence. The third branch at the edge is used for lithological classification and includes the third fully connected layer, the third Dropout layer, and the third fully connected output layer connected in sequence.

4. The lithological intelligent identification method based on an edge-cloud collaborative architecture as described in claim 1, characterized in that, The step of determining whether to upload to the cloud based on the confidence level corresponding to the preliminary identification results is as follows: Two main outputs are obtained from the lightweight spectral analysis model: preliminary lithological classification and its confidence level, and a list of elemental types and their probabilities; The system determines whether to upload to the cloud based on the presence of rare element signals and the confidence level. If the confidence level is greater than or equal to the first confidence threshold and no rare element signal is detected, the edge analysis results are considered reliable enough, and the edge analysis results are directly output without uploading. If the confidence level is less than the second confidence threshold or a rare element signal is detected, the upload process is triggered, and the data is sent to the cloud for detailed analysis.

5. The lithological intelligent identification method based on an edge-cloud collaborative architecture as described in claim 4, characterized in that, The determination of rare element signals is as follows: check whether there is any element in the preset rare element list whose probability exceeds the rare element probability threshold in the element type and its probability list. If the recognition probability of the rare element is not less than the rare element probability threshold, then a rare element signal is detected; if it is less than the threshold, then a rare element signal is not detected.

6. The lithological intelligent identification method based on an edge-cloud collaborative architecture as described in claim 1, characterized in that, The cloud-based lithology precision analysis model adopts a 1D-CNN+Transformer hybrid architecture, which integrates a self-attention mechanism to automatically focus on rare element bands and outputs element mass percentage, mineral composition list, lithology identification report and confidence level. The 1D-CNN+Transformer hybrid architecture comprises three stages: a local feature extraction stage, a global context modeling stage, and a multi-task output stage. The local feature extraction stage employs a 1D-CNN local feature extractor, which acts as a high-level feature extractor, converting the original spectral signal into a feature map rich in semantic information and progressively reducing the sequence length. The global context modeling stage uses Transformer global context modeling. The multi-task output stage includes a multi-task output layer.

7. The lithological intelligent identification method based on an edge-cloud collaborative architecture as described in claim 6, characterized in that, The multi-task output layer includes three independent task-specific branches: The first branch in the cloud is used for quantitative regression of elements, which includes a fully connected layer and a linear activation function layer. The output dimension is the number of element types, and the linear activation function is used to output the percentage of element quality. The second branch in the cloud is used for mineral composition classification, including a fully connected layer and a sigmoid activation function. The output dimension is the number of mineral types, and the sigmoid activation function is used to output the probability of the existence of each mineral. The third branch in the cloud is used for lithology identification. It includes a fully connected layer and a Softmax activation function. The output dimension is the number of lithology types. Using the Softmax activation function, it outputs the final lithology classification and confidence level.

8. A lithological intelligent identification system based on an edge-cloud collaborative architecture, characterized in that, include: The data acquisition module is used to collect rock spectral data and perform preprocessing. The edge recognition module is used to perform preliminary analysis on the preprocessed spectral data at the edge layer using a lightweight spectral analysis model to obtain preliminary lithology identification results, and to determine whether it is necessary to upload to the cloud layer based on the confidence level corresponding to the preliminary identification results. If no upload is required, the preliminary lithology identification results will be output directly. The communication transmission module is used to determine the network classification based on round-trip time, bandwidth, and packet loss rate when uploading is required. Based on the network classification, the communication layer uploads the spectral data or pre-processed spectral data to the cloud. The communication layer uses real-time data from heartbeat packets to measure the round-trip time, bandwidth, and packet loss rate, and then uses preset network classification rules to determine the current network as good, weak, or extremely weak. Under good network conditions, the original complete spectral data is uploaded. Under weak network conditions, PCA is used to reduce the dimensionality of the pre-processed spectral data before uploading. Under extremely weak network conditions, only abnormal spectral band data is uploaded; abnormal spectral bands are obtained by extracting the bands of the ±10 channels of the identified characteristic spectral peaks. The cloud-based identification module is used to re-analyze the spectral data or pre-processed spectral data in the cloud using a cloud-based lithology precision analysis model to obtain accurate elemental quantitative analysis results, mineral composition, and lithology identification reports. The cloud-based lithology precision analysis model adopts a deep neural network structure and integrates an attention mechanism to enhance the model's ability to interpret complex spectral features.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the lithology intelligent identification method based on an edge-cloud collaborative architecture as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the rock type intelligent identification method based on an edge-cloud collaborative architecture as described in any one of claims 1-7.

Images