An intelligent fusion terminal detection method and system based on artificial intelligence

Abstract

The application relates to the technical field of intelligent fusion terminal detection, and discloses an intelligent fusion terminal detection method and system based on artificial intelligence. The method obtains original running data flow and preset task parameters of a terminal to be detected, and extracts a multi-source heterogeneous feature vector representing terminal running time sequence dynamics, working state spectrum distribution and internal physical state in parallel. Then, a terminal anomaly detection model with an encoder-decoder structure is used to fuse the features, and a multi-head cross attention mechanism is used to capture the correlation between cross-modal features, and output a terminal state anomaly probability. Finally, a terminal health state judgment result is generated based on a dynamic decision threshold and a preset decision logic, so that the detection accuracy and efficiency are improved, and the false positive rate and the false negative rate are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent fusion terminal detection technology, and specifically to an intelligent fusion terminal detection method and system based on artificial intelligence. Background Technology

[0002] In current applications of intelligent converged terminals, as terminal functions become increasingly complex and diversified, the amount of data generated during their operation increases dramatically, and the types of data also tend to become more diversified.

[0003] Traditional terminal detection methods often rely on a single type of data or simple rule matching, making it difficult to comprehensively and accurately capture the complex state changes during terminal operation.

[0004] Especially when dealing with multi-source heterogeneous data, traditional methods cannot effectively integrate feature information from different modalities, resulting in high false alarm and false negative rates in the detection of abnormal terminal states. Furthermore, traditional methods are inefficient when processing large-scale data, making it difficult to meet the needs of real-time detection.

[0005] Therefore, developing a method that can efficiently and accurately detect abnormal states of intelligent fusion terminals is of great practical significance. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this invention is to provide an artificial intelligence-based intelligent fusion terminal detection method and system, solving the problems of low accuracy and poor efficiency in anomaly detection of intelligent fusion terminals.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides an intelligent fusion terminal detection method based on artificial intelligence, comprising: Step S1: Obtain the raw operating data stream and preset task parameters of the intelligent fusion terminal to be tested; Step S2: Perform parallel feature extraction on the original running data stream to obtain the characteristic vector representing the dynamic sequence of the terminal's runtime, the frequency domain vector representing the spectrum distribution of the working state, and the physical parameter vector representing the internal physical state of the terminal. Step S3: Input the feature vector, frequency domain vector, physical parameter vector and preset task parameters into the pre-trained terminal anomaly detection model. The terminal anomaly detection model adopts an encoder-decoder structure to fuse multi-source features and captures the correlation between cross-modal features through a multi-head cross-attention mechanism to output the terminal state anomaly probability. Step S4: Based on the probability of abnormal terminal status and the preset decision logic, generate the terminal health status judgment result.

[0008] Preferably, in one possible implementation of the first aspect, the feature vector extraction process characterizing the terminal runtime dynamics in step S2 includes: The time-series data in the original running data stream is divided into continuous time segments using a sliding window. Each time segment is input into the Long Short-Term Memory network; The hidden states of each temporal segment at the last time step of the Long Short-Term Memory network are aggregated, and then dimensionality reduction and nonlinear transformation are performed through a fully connected layer to output a feature vector of fixed dimension.

[0009] Preferably, in one possible implementation of the first aspect, the frequency domain vector extraction process characterizing the spectral distribution of the operating state in step S2 includes: Perform a short-time Fourier transform on the time segment to convert the time-domain signal into a frequency-domain signal; Calculate the energy distribution of the frequency domain signal in different frequency bands to generate initial spectral characteristics; The initial spectral features are decomposed into multiple sub-band components by applying wavelet transform; Statistical features of each sub-band component are extracted, including energy entropy, spectral centroid, and spectral roll-off point. These statistical features are then concatenated to form a frequency domain vector.

[0010] Preferably, in one possible implementation of the first aspect, the physical parameter vector extraction process characterizing the internal physical state of the terminal in step S2 includes: Sensor data, including voltage, current, temperature, and cooling fan speed, are extracted from the raw operating data stream. Perform data processing on sensor data, including outlier removal and smoothing filtering; Physical parameters, including power consumption, energy efficiency ratio, and heat accumulation rate, are calculated based on the processed sensor data. The processed sensor data and physical parameters are normalized and then concatenated to form a physical parameter vector.

[0011] Preferably, in one possible implementation of the first aspect, the terminal anomaly detection model structure in step S3 adopts a multi-source heterogeneous data fusion encoder and a time-series state decoder. The input to the multi-source heterogeneous data fusion encoder is a parallel-stitched feature vector, frequency domain vector, physical parameter vector, and preset task parameters mapped by the embedding layer; The multi-source heterogeneous data fusion encoder consists of a cross-modal feature interaction layer and a temporal context coding layer connected in sequence; The cross-modal feature interaction layer uses a multi-head cross-attention mechanism to calculate the correlation between each modal feature and all other modal features in the feature vector, frequency domain vector, physical parameter vector, and the mapped preset task parameters, and outputs a modal interaction feature sequence. The temporal context coding layer receives the modal interaction feature sequence, uses a dilated convolutional network to capture multi-scale temporal patterns in the modal interaction feature sequence, and outputs a comprehensive state representation sequence. The temporal state decoder receives the comprehensive state representation sequence and maps it to the scalar form of terminal state anomaly probability through a feedforward neural network layer.

[0012] Preferably, in one possible implementation of the first aspect, the cross-modal feature interaction layer assigns learnable modal weight parameters to the feature vector, frequency domain vector, and physical parameter vector respectively before performing multi-head cross-attention calculation.

[0013] Preferably, in one possible implementation of the first aspect, the temporal context coding layer inputs the modal interaction feature sequence into a dilated convolutional network; By using convolutional layers with different dilation rates in a dilated convolutional network, temporal dependencies at different time scales in feature sequences can be captured in parallel. By fusing the multi-scale features output from each convolutional layer, a comprehensive state representation sequence is generated that represents the comprehensive temporal pattern of the terminal state.

[0014] Preferably, in one possible implementation of the first aspect, the preset decision logic in step S4 includes: Set a dynamic decision threshold. The dynamic decision threshold is queried and assigned a value based on the terminal working mode and security level specified in the preset task parameters. Define the terminal health status categories, including normal status, warning status, and abnormal status; Set the judgment rules: when the probability of the terminal state being abnormal is less than the first threshold, it is judged as a normal state; when the probability of the terminal state being abnormal is greater than or equal to the first threshold and less than the second threshold, it is judged as a warning state; when the probability of the terminal state being abnormal is greater than or equal to the second threshold, it is judged as an abnormal state.

[0015] Preferably, in one possible implementation of the first aspect, the process of generating the terminal health status determination result in step S4 includes: The probability of abnormal terminal status is compared with the first threshold and the second threshold in the dynamic decision threshold. According to the judgment rules, the corresponding terminal health status classification is obtained through mapping. The terminal health status classification, terminal status abnormality probability value and trigger judgment rules are formatted and encapsulated to generate a structured data packet containing the terminal health status judgment result.

[0016] Secondly, the present invention provides an artificial intelligence-based intelligent fusion terminal detection system, the system being used to implement the artificial intelligence-based intelligent fusion terminal detection method as described in the first aspect, comprising: The data acquisition module acquires the raw operating data stream and preset task parameters of the intelligent fusion terminal to be tested; The parallel feature extraction module performs parallel feature extraction on the original running data stream to obtain characteristic vectors that characterize the dynamic sequence of the terminal during operation, frequency domain vectors that characterize the spectral distribution of the working state, and physical parameter vectors that characterize the internal physical state of the terminal. The anomaly detection module inputs feature vectors, frequency domain vectors, physical parameter vectors, and preset task parameters into a pre-trained terminal anomaly detection model. The terminal anomaly detection model uses an encoder-decoder structure to fuse multi-source features and captures the correlation between cross-modal features through a multi-head cross-attention mechanism, outputting the probability of terminal state anomalies. The result generation module generates a terminal health status determination result based on the terminal status anomaly probability and preset decision logic.

[0017] The beneficial effects of this invention are as follows: This invention improves the richness and accuracy of feature information by extracting features in parallel and fusing multi-source heterogeneous data such as terminal runtime dynamics, working state spectrum distribution and internal physical state.

[0018] The terminal anomaly detection model, which adopts an encoder-decoder structure and combines a multi-head cross-attention mechanism, can automatically capture the complex correlations between cross-modal features and enhance the model's ability to identify terminal anomalies.

[0019] Meanwhile, the adoption of dynamic decision thresholds and decision rules makes the judgment results more accurate and can adapt to the detection needs under different working modes and security levels.

[0020] Overall, this invention improves the accuracy and efficiency of anomaly detection in intelligent fusion terminals, reduces false alarm and false negative rates, and provides strong support for the stable operation of terminals. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This application provides a flowchart of an intelligent fusion terminal detection method based on artificial intelligence. Detailed Implementation

[0023] 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.

[0024] Example 1: As Figure 1 As shown, this invention provides an intelligent fusion terminal detection method based on artificial intelligence, comprising: Step S1: Obtain the raw operating data stream and preset task parameters of the intelligent fusion terminal to be tested.

[0025] In this embodiment, the raw operational data stream of the intelligent fusion terminal under test is acquired. The raw operational data stream is a collection of multi-source heterogeneous data generated in real time by the terminal's internal sensors and system logs during operation. The raw operational data stream is continuously output through the data acquisition interface built into the terminal device and includes time-series data, low-level signal data required for spectrum analysis, and sensor data that directly characterizes the internal physical state of the terminal.

[0026] Time-series data covers key performance indicators such as CPU utilization, memory usage, network throughput, and disk read / write latency; raw signal data consists of time-domain electrical signals collected from the terminal hardware monitoring unit; sensor data includes measured physical parameters such as the terminal's internal power supply voltage, operating current, core chip temperature, and cooling fan speed. All of the above data is collected through the terminal device's data bus and aligned and packaged according to a unified time stamp to form a continuous, synchronous raw operating data stream.

[0027] Simultaneously, preset task parameters are obtained. This set of parameters is read from the detection task file issued by the upper-level management platform. The preset task parameters are the metadata and control variables of this detection task. Their content is fixed and includes the terminal's unique identifier, the terminal's currently specified working mode, the security level required for this detection, the time window for triggering the detection task, and the expected duration of the detection.

[0028] The terminal operates in three modes: high-performance, balanced, and energy-saving. Security levels are categorized as low, medium, and high. Simultaneously, preset task parameters are parsed and their format verified and content mapped before being input into the detection system.

[0029] Step S2: Perform parallel feature extraction on the original running data stream to obtain the characteristic vector representing the dynamic sequence of the terminal during operation, the frequency domain vector representing the spectrum distribution of the working state, and the physical parameter vector representing the internal physical state of the terminal.

[0030] In this embodiment, parallel feature extraction of the raw runtime data stream is performed synchronously through three independent processing channels. The first channel is used to extract feature vectors characterizing the dynamic timing of the terminal runtime. Time-series data, including CPU utilization, memory usage, network throughput, and disk read / write latency, is separated from the raw runtime data stream. These time-series data are segmented using a fixed-length sliding window to generate a series of continuous, equal-length time-series segments. Each time-series segment is then input into a Long Short-Term Memory (LSTM) network in time-step order.

[0031] This Long Short-Term Memory (LSTM) network consists of an input layer, hidden layers, and an output layer. The hidden layers contain memory cells and gating mechanisms to model long-term and short-term dependencies in the sequence. During forward computation, the network receives temporal data points at each time step and updates its hidden state. After processing all time steps of a temporal segment, the hidden state generated by the LSM network at the last time step is taken as the initial representation of that segment. The final hidden states corresponding to all temporal segments are aggregated to form a hidden state sequence. This hidden state sequence is input into a fully connected layer, which performs a linear transformation and applies an activation function to achieve dimensionality reduction and nonlinear mapping of features. The final output is a dense vector of fixed dimensions, which is the feature vector representing the temporal dynamics of the terminal runtime.

[0032] The second channel is used to extract a frequency domain vector representing the spectral distribution of the operating state. Original signal data aligned with the aforementioned time segments is obtained from the raw running data stream. A windowed short-time Fourier transform is applied to the time-domain signal within each time segment to convert it into a time-frequency domain representation, yielding the spectrum. The energy integral of this spectrum over multiple preset frequency bands is calculated to generate initial spectral features describing the distribution of signal energy at different frequencies. Subsequently, a discrete wavelet transform is applied to the initial spectral features, decomposing them into approximation coefficients and detail coefficients at different scales through convolution and downsampling operations. These coefficients correspond to different sub-band components. For each sub-band component, a set of statistical features is extracted: the extracted energy entropy measures the degree of disorder in the energy distribution within that sub-band; the extracted spectral centroid describes the concentrated frequency of the sub-band's spectral energy; and the extracted spectral roll-off point represents the frequency position corresponding to when the accumulated energy within that sub-band reaches a specific percentage of the total energy. Finally, the energy entropy, spectral centroid, and spectral roll-off point of all sub-band components are concatenated sequentially to form a comprehensive frequency domain vector.

[0033] The third channel is used to extract physical parameter vectors characterizing the internal physical state of the terminal. Measurement data from hardware sensors, including supply voltage, operating current, core chip temperature, and cooling fan speed, is separated from the raw operational data stream. First, the sensor data is preprocessed, using statistical quantile-based rules to identify and remove outliers that significantly deviate from the normal range. Then, a moving average filter is applied to smooth the data and suppress random noise. Based on the preprocessed sensor data, physical parameters are calculated: instantaneous power consumption is calculated based on the instantaneous values ​​of voltage and current; the energy efficiency ratio is calculated based on the average power consumption over a period of time and the useful workload completed by the terminal; and the heat accumulation rate is calculated based on the rate of temperature change over time and the cooling fan speed. Finally, the preprocessed sensor data and the calculated power consumption, energy efficiency ratio, and heat accumulation rate are normalized, and all normalized values ​​are concatenated in a predetermined order to form a physical parameter vector.

[0034] Step S3: Input the feature vector, frequency domain vector, physical parameter vector and preset task parameters into the pre-trained terminal anomaly detection model. The terminal anomaly detection model adopts an encoder-decoder structure to fuse multi-source features and captures the correlation between cross-modal features through a multi-head cross-attention mechanism to output the probability of terminal state anomaly.

[0035] In this embodiment, the terminal anomaly detection model adopts an encoder-decoder structure. First, preset task parameters are input into an embedding layer, which is a fully connected network that maps discrete or structured task parameters into a continuous and dense vector representation. Next, the preset task parameter vector mapped by the embedding layer is concatenated in parallel with feature vectors, frequency domain vectors, and physical parameter vectors to form a comprehensive feature sequence that integrates multi-source heterogeneous data, which serves as the input to the multi-source heterogeneous data fusion encoder.

[0036] The multi-source heterogeneous data fusion encoder consists of a cross-modal feature interaction layer and a temporal context coding layer connected sequentially. In the cross-modal feature interaction layer, a learnable modality weight parameter is first assigned to each of the three different feature modalities: feature vector, frequency domain vector, and physical parameter vector. These weight parameters are scalar values ​​learned during model training via backpropagation, used for preliminary automatic calibration of the importance of different modalities before feature fusion. Then, the weighted feature vectors and the mapped pre-defined task parameter vectors are input together into a multi-head cross-attention mechanism for computation.

[0037] The specific calculation process of the multi-head cross-attention mechanism is as follows: For each set of features in the concatenated comprehensive feature sequence—namely, the feature vector, frequency domain vector, physical parameter vector, and mapped task parameter vector—corresponding query vector, key vector, and value vector are generated through different linear transformation layers. In each attention head, for a target modality's feature vector, the query vector is multiplied by the key vectors of all modal features to obtain the attention score matrix between the target modality and all other modalities. This score matrix is ​​scaled and normalized to form attention weights. Then, the attention weights are used to perform a weighted summation of the value vectors of all modal features to obtain a feature representation that integrates cross-modal information. The outputs of multiple attention heads are concatenated and integrated through a linear projection layer, ultimately outputting a modal interaction feature sequence containing rich cross-modal correlation information.

[0038] The temporal context encoding layer receives a sequence of modal interaction features and employs a dilated convolutional network to capture the temporal dependencies at different time scales within the feature sequence. The dilated convolutional network contains multiple parallel dilated convolutional layers, each with a different dilation rate. The dilation rate determines the sampling interval of the input data when the convolutional kernel processes the sequence. A larger dilation rate allows the convolutional kernel to perceive features over a wider time interval, capturing long-range temporal patterns; a smaller dilation rate focuses on feature changes at local, nearby time points, capturing short-range temporal patterns. Each dilated convolutional layer independently performs a convolution operation on the input modal interaction feature sequence, extracting temporal features at the corresponding time scale. Subsequently, the multi-scale temporal features from the outputs of all convolutional layers with different dilation rates are fused. The fusion operation uses channel concatenation followed by convolution. Finally, the temporal context encoding layer outputs a comprehensive state representation sequence that represents the overall temporal pattern of the terminal state.

[0039] The temporal state decoder receives the comprehensive state representation sequence. The decoder first performs global average pooling on the sequence, compressing the variable-length sequence features into a fixed-dimensional global feature vector. This global feature vector is then input into a feedforward neural network layer. This feedforward neural network layer consists of multiple fully connected layers, with non-linear activation functions between layers. The feedforward neural network performs non-linear transformations and dimensional mappings on the input features, ultimately outputting a scalar value. This scalar value represents the probability of the terminal being in an abnormal state, ranging from 0 to 1, indicating the likelihood of the terminal being in an abnormal state.

[0040] The training process of the terminal anomaly detection model is based on supervised learning using labeled historical operational data. First, a training dataset is constructed by collecting raw operational data streams and corresponding task parameters from intelligent fusion terminals during their normal operation and known anomaly periods. Each historical raw operational data stream is processed, simultaneously generating corresponding feature vectors, frequency domain vectors, and physical parameter vectors. Based on the actual state of the terminal corresponding to that historical data, a true label is assigned: 0 for a normal terminal and 1 for a terminal in a definitive anomaly state. This forms a sample set consisting of input feature groups (feature vector, frequency domain vector, physical parameter vector, and task parameters) and output labels, which is then divided into a training set, a validation set, and a test set in a 7:2:1 ratio.

[0041] The terminal anomaly detection model is initialized with randomly assigned parameters and uses a binary cross-entropy loss function. For a single training sample, let the probability of the model predicting an abnormal terminal state be... Its corresponding real label is ( If the value is 0 or 1, then the loss L is calculated as follows: The total loss of the model on the entire training set is the average of the losses of all samples.

[0042] The model is trained using a gradient descent-based optimization algorithm, with training iteratively performed in batches. Within each training batch, forward propagation is performed: the input feature set of a batch of samples is fed into the model, processed through the embedding layer, cross-modal feature interaction layer, temporal context encoding layer, and temporal state decoder, ultimately yielding the predicted probabilities for that batch of samples. Next, the binary cross-entropy loss between the batch's predictions and the true labels is calculated. Then, backpropagation is performed: the gradient of the loss value with respect to each trainable parameter of the model (including the weights and biases of each neural network layer, and the learnable modal weights in the cross-modal feature interaction layer, etc.) is calculated. The optimizer updates all parameters of the model based on the calculated gradients, reducing the loss value.

[0043] During training, the model is periodically evaluated using an independent validation set. Performance metrics such as loss and classification accuracy on the validation set are monitored. Training is terminated early if validation set performance no longer improves over several consecutive training epochs to prevent overfitting, and a snapshot of the model parameters with the best validation set performance is saved as the final trained model. Finally, the final trained model is evaluated using a test set to confirm its generalization ability.

[0044] Step S4: Based on the probability of abnormal terminal status and the preset decision logic, generate the terminal health status judgment result.

[0045] In this embodiment, the preset decision logic includes a dynamic decision threshold module, a terminal health status classification definition, and decision rules. The dynamic decision threshold module maintains a lookup table that defines a first threshold and a second threshold corresponding to different combinations of terminal operating modes and security levels. The first threshold and the second threshold are real numbers between 0 and 1, and the first threshold is less than the second threshold. When preset task parameters are obtained, the system retrieves and assigns the corresponding first threshold and second threshold from the lookup table according to the terminal operating mode and security level specified in the parameters.

[0046] The terminal operates in three modes: high-performance, balanced, and energy-saving. Security levels are categorized as low, medium, and high. In high-performance mode, the first and second thresholds are typically set to lower values ​​to monitor performance fluctuations. In energy-saving mode, the thresholds are set to higher values ​​to tolerate a wider range of performance fluctuations. Higher security levels correspond to lower first and second thresholds.

[0047] Terminal health status is categorized into three discrete states: normal, warning, and abnormal. The decision rule compares the probability P of the terminal's abnormal state output by the terminal anomaly detection model with a dynamically assigned threshold. If probability P is less than the first threshold T1, the terminal is considered to be in a normal state; if probability P is greater than or equal to the first threshold T1 and less than the second threshold T2, the terminal is considered to be in a warning state; and if probability P is greater than or equal to the second threshold T2, the terminal is considered to be in an abnormal state.

[0048] The process of generating a terminal health status determination result is as follows: First, the calculated terminal state anomaly probability P is compared with the judgment thresholds T1 and T2, which are dynamically assigned according to preset task parameters. Next, according to the judgment rules, the probability value is mapped to the corresponding terminal health status category. Then, the determination result is formatted and encapsulated. The formatting and encapsulation process encapsulates the terminal health status category, the specific numerical value of the terminal state anomaly probability, the specific rule clauses on which the judgment is triggered, the timestamp of this detection, and the unique identifier of the terminal to be detected into a structured data packet with predetermined fields and format. This structured data packet is the final terminal health status determination result and is reported to the upper-level management system.

[0049] Example 2: This invention provides an intelligent fusion terminal detection method based on artificial intelligence, and also includes a progressive training and optimization strategy for the terminal anomaly detection model.

[0050] First, self-supervised pre-training is performed on a large amount of unlabeled historical operational data from terminals. By designing pre-tasks such as mask reconstruction and contrastive learning, the model learns general multimodal data representations. Subsequently, supervised fine-tuning is performed on a small number of labeled anomaly samples to adapt to specific anomaly detection targets.

[0051] Furthermore, this method includes an automated hyperparameter optimization loop. During the model training phase, a Bayesian optimization method is employed to automatically optimize key hyperparameters of the model within a defined search space. These key hyperparameters include the number of heads in the multi-head cross-attention mechanism, the dilation rate combination of the dilated convolutional network, and the learning rate decay strategy during training, achieving an optimal balance between model performance and computational efficiency.

[0052] After the model is deployed, it supports a model update mechanism based on online learning or periodic incremental learning. The model parameters can be safely updated based on newly generated and validated abnormal data samples, continuously improving the ability to detect new abnormal patterns.

[0053] Example 3: This invention provides an intelligent fusion terminal detection system based on artificial intelligence, comprising: The data acquisition module acquires the raw operating data stream and preset task parameters of the intelligent fusion terminal to be tested; The parallel feature extraction module performs parallel feature extraction on the original running data stream to obtain characteristic vectors that characterize the dynamic sequence of the terminal during operation, frequency domain vectors that characterize the spectral distribution of the working state, and physical parameter vectors that characterize the internal physical state of the terminal. The anomaly detection module inputs feature vectors, frequency domain vectors, physical parameter vectors, and preset task parameters into a pre-trained terminal anomaly detection model. The terminal anomaly detection model uses an encoder-decoder structure to fuse multi-source features and captures the correlation between cross-modal features through a multi-head cross-attention mechanism, outputting the probability of terminal state anomalies. The result generation module generates a terminal health status determination result based on the terminal status anomaly probability and preset decision logic.

[0054] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting intelligent fusion terminals based on artificial intelligence, characterized in that, include: Step S1: Obtain the raw operating data stream and preset task parameters of the intelligent fusion terminal to be tested; Step S2: Perform parallel feature extraction on the original running data stream to obtain the characteristic vector representing the dynamic sequence of the terminal's runtime, the frequency domain vector representing the spectrum distribution of the working state, and the physical parameter vector representing the internal physical state of the terminal. Step S3: Input the feature vector, frequency domain vector, physical parameter vector and preset task parameters into the pre-trained terminal anomaly detection model. The terminal anomaly detection model adopts an encoder-decoder structure to fuse multi-source features and captures the correlation between cross-modal features through a multi-head cross-attention mechanism to output the terminal state anomaly probability. Step S4: Based on the probability of abnormal terminal status and the preset decision logic, generate the terminal health status judgment result.

2. The method for detecting intelligent fusion terminals based on artificial intelligence as described in claim 1, characterized in that, The feature vector extraction process characterizing the terminal runtime dynamics in step S2 includes: The time-series data in the original running data stream is divided into continuous time segments using a sliding window. Each time segment is input into the Long Short-Term Memory network; The hidden states of each temporal segment at the last time step of the Long Short-Term Memory network are aggregated, and then dimensionality reduction and nonlinear transformation are performed through a fully connected layer to output a feature vector of fixed dimension.

3. The method for detecting an intelligent fusion terminal based on artificial intelligence as described in claim 2, characterized in that, The frequency domain vector extraction process characterizing the spectral distribution of the working state in step S2 includes: Perform a short-time Fourier transform on the time segment to convert the time-domain signal into a frequency-domain signal; Calculate the energy distribution of the frequency domain signal in different frequency bands to generate initial spectral characteristics; The initial spectral features are decomposed into multiple sub-band components by applying wavelet transform; Statistical features of each sub-band component are extracted, including energy entropy, spectral centroid, and spectral roll-off point. These statistical features are then concatenated to form a frequency domain vector.

4. The method for detecting an intelligent fusion terminal based on artificial intelligence as described in claim 3, characterized in that, The physical parameter vector extraction process characterizing the internal physical state of the terminal in step S2 includes: Sensor data, including voltage, current, temperature, and cooling fan speed, are extracted from the raw operating data stream. Perform data processing on sensor data, including outlier removal and smoothing filtering; Physical parameters, including power consumption, energy efficiency ratio, and heat accumulation rate, are calculated based on the processed sensor data. The processed sensor data and physical parameters are normalized and then concatenated to form a physical parameter vector.

5. The method for detecting an intelligent fusion terminal based on artificial intelligence as described in claim 1, characterized in that, In step S3, the terminal anomaly detection model structure adopts a multi-source heterogeneous data fusion encoder and a time-series state decoder. The input to the multi-source heterogeneous data fusion encoder is a parallel-stitched feature vector, frequency domain vector, physical parameter vector, and preset task parameters mapped by the embedding layer; The multi-source heterogeneous data fusion encoder consists of a cross-modal feature interaction layer and a temporal context coding layer connected in sequence; The cross-modal feature interaction layer uses a multi-head cross-attention mechanism to calculate the correlation between each modal feature and all other modal features in the feature vector, frequency domain vector, physical parameter vector, and the mapped preset task parameters, and outputs a modal interaction feature sequence. The temporal context coding layer receives the modal interaction feature sequence, uses a dilated convolutional network to capture multi-scale temporal patterns in the modal interaction feature sequence, and outputs a comprehensive state representation sequence. The temporal state decoder receives the comprehensive state representation sequence and maps it to the scalar form of terminal state anomaly probability through a feedforward neural network layer.

6. The method for detecting an intelligent fusion terminal based on artificial intelligence as described in claim 5, characterized in that, Before performing multi-head cross-attention calculation, the cross-modal feature interaction layer assigns learnable modal weight parameters to the feature vector, frequency domain vector, and physical parameter vector, respectively.

7. The method for detecting an intelligent fusion terminal based on artificial intelligence as described in claim 6, characterized in that, The temporal context coding layer inputs the modal interaction feature sequence into the dilated convolutional network; By using convolutional layers with different dilation rates in a dilated convolutional network, temporal dependencies at different time scales in feature sequences can be captured in parallel. By fusing the multi-scale features output from each convolutional layer, a comprehensive state representation sequence is generated that represents the comprehensive temporal pattern of the terminal state.

8. The method for detecting intelligent fusion terminals based on artificial intelligence as described in claim 1, characterized in that, The preset decision logic in step S4 includes: Set a dynamic decision threshold. The dynamic decision threshold is queried and assigned a value based on the terminal working mode and security level specified in the preset task parameters. Define the terminal health status categories, including normal status, warning status, and abnormal status; Set the judgment rules: when the probability of the terminal state being abnormal is less than the first threshold, it is judged as a normal state; when the probability of the terminal state being abnormal is greater than or equal to the first threshold and less than the second threshold, it is judged as a warning state; when the probability of the terminal state being abnormal is greater than or equal to the second threshold, it is judged as an abnormal state.

9. The method for detecting an intelligent fusion terminal based on artificial intelligence as described in claim 8, characterized in that, The process of generating the terminal health status determination result in step S4 includes: The probability of abnormal terminal status is compared with the first threshold and the second threshold in the dynamic decision threshold. According to the judgment rules, the corresponding terminal health status classification is obtained through mapping. The terminal health status classification, terminal status abnormality probability value and trigger judgment rules are formatted and encapsulated to generate a structured data packet containing the terminal health status judgment result.

10. An intelligent fusion terminal detection system based on artificial intelligence, characterized in that, The system is used to implement an artificial intelligence-based intelligent fusion terminal detection method as described in any one of claims 1 to 9, comprising: The data acquisition module acquires the raw operating data stream and preset task parameters of the intelligent fusion terminal to be tested; The parallel feature extraction module performs parallel feature extraction on the original running data stream to obtain characteristic vectors that characterize the dynamic sequence of the terminal during operation, frequency domain vectors that characterize the spectral distribution of the working state, and physical parameter vectors that characterize the internal physical state of the terminal. The anomaly detection module inputs feature vectors, frequency domain vectors, physical parameter vectors, and preset task parameters into a pre-trained terminal anomaly detection model. The terminal anomaly detection model uses an encoder-decoder structure to fuse multi-source features and captures the correlation between cross-modal features through a multi-head cross-attention mechanism, outputting the probability of terminal state anomalies. The result generation module generates a terminal health status determination result based on the terminal status anomaly probability and preset decision logic.

Images