A method for early identification of abnormal evolution of a new energy mine car battery system for deep metal mines

Abstract

The application discloses a new energy mine car battery system early abnormal evolution identification method for deep metal mines, and belongs to the technical field of mine equipment safety monitoring and intelligent operation and maintenance. The method comprises the following steps: collecting temperature signals and displacement signals of the battery system, and constructing displacement input sequences and temperature input sequences; adopting a time sequence convolution network to extract features of the displacement input sequences; adopting a long short-term memory network to extract features of the temperature input sequences; mapping the extracted displacement modal features and temperature modal features to a unified shared latent space, then performing feature splicing, and adopting a multilayer perceptron to jointly predict displacement states and temperature states of the battery at the next moment; constructing a joint loss function to optimize and train the network model; in the online inference stage, an abnormal score index is constructed according to the deviation between the prediction result and the actual observation value, the abnormal score index is compared with a threshold value, and when the abnormal score exceeds the threshold value, it is determined that the battery system has an abnormal evolution trend.

Description

Technical Field

[0001] This invention belongs to the field of mine equipment safety monitoring and intelligent operation and maintenance technology, and relates to an early identification method for abnormal evolution of battery systems for new energy mining trucks used in deep metal mines. Background Technology

[0002] With the advancement of green and intelligent mine construction, the application of new energy mining trucks in deep metal mines is becoming increasingly widespread. Compared with traditional fuel-powered equipment, new energy mining trucks have advantages such as low emissions and low noise, but their core power system—the power battery—faces more severe safety challenges in the complex environment of deep mines.

[0003] Deep metal mines typically exhibit characteristics such as high geothermal temperature, high humidity, strong vibration, and high load conditions. Under these conditions, batteries are prone to multiple physical coupling phenomena during operation, including heat accumulation, structural stress changes, and internal material evolution, manifesting as latent features such as abnormal temperature increases and micro-structural deformation. These early anomalies usually evolve slowly and are not easily observed directly, but once they develop into thermal runaway or structural failure, they will seriously threaten the safety of mine car operations and underground workers. Especially in the enclosed environment of deep mines, if a battery experiences thermal runaway or even leads to combustion or explosion, it can easily cause the spread of toxic and harmful gases, damage to interlocking equipment, and casualties. The scope of the hazard is wide, the consequences are severe, and the safety risks are far greater than in surface applications. Therefore, early identification of subtle structural changes and temperature anomalies in the early stages of battery anomaly is of great significance for timely maintenance and risk intervention, and is a key technical requirement for ensuring safe production in mines.

[0004] Existing battery monitoring methods mostly rely on single signals (such as temperature and voltage) for threshold judgment, which makes it difficult to capture the coupling relationship between multiple physical quantities under complex operating conditions, resulting in delayed anomaly detection or a high false alarm rate. In addition, supervised learning methods based on fault samples are difficult to apply in mining scenarios because anomaly samples are difficult to obtain and unevenly distributed.

[0005] Therefore, there is an urgent need for a method that can integrate multi-source information, model the coupled evolution of battery states, and achieve early identification of anomalies in the absence of abnormal samples, so as to improve the safety and reliability of new energy mining truck battery systems in deep metal mines. Summary of the Invention

[0006] The purpose of this invention is to address the problem that the internal mechanical and thermal responses of the battery system of new energy mining trucks used in deep metal mines exhibit strong coupling characteristics and that abnormal evolution is highly concealed and difficult to identify in a timely manner when operating under complex conditions such as high temperature, high humidity, strong vibration and high load. This invention proposes an early identification method for abnormal evolution of the battery system of new energy mining trucks used in deep metal mines, thereby improving the safety and reliability of the battery system operation.

[0007] This invention provides a method for early identification of abnormal evolution in battery systems for new energy mining trucks used in deep metal mines, comprising:

[0008] Step 1: Arrange temperature sensors and displacement sensors in the battery system to collect temperature signals and displacement signals respectively, and construct displacement input sequences and temperature input sequences with a preset time window length;

[0009] Step 2: Extract temporal features from the displacement input sequence and model the displacement sequence using a temporal convolutional network to obtain displacement modal features;

[0010] Step 3: Perform time-series modeling on the temperature input sequence. Use a long short-term memory network to model the temperature sequence to obtain temperature modal features;

[0011] Step 4: Map the displacement mode features and temperature mode features to a unified shared latent space to achieve feature alignment and then perform feature concatenation to construct a joint state representation. Use a multilayer perceptron to jointly predict the displacement state and temperature state of the battery at the next moment.

[0012] Step 5: Construct a joint loss function that includes alignment loss, displacement prediction loss, and temperature prediction loss, and optimize and train the network model;

[0013] Step 6: In the online inference stage, an anomaly scoring index is constructed based on the deviation between the prediction results and the actual observations. The anomaly scoring index is then compared with a threshold. When the anomaly score exceeds the threshold, it is determined that the battery system has an abnormal evolution trend.

[0014] The present invention provides an early identification method for abnormal evolution of battery systems in new energy mining trucks used in deep metal mines, which has the following beneficial effects:

[0015] This invention effectively characterizes the multi-physical coupling evolution of batteries under complex working conditions in deep mines by fusing battery structural response and temperature response information, thereby improving the accuracy of anomaly identification. It enhances semantic consistency between different modes through a cross-modal feature alignment mechanism, thus improving the sensitivity of anomaly detection. Furthermore, this method does not rely on abnormal samples and can achieve anomaly identification based solely on normal operating data, making it suitable for scenarios where abnormal data is difficult to obtain in actual mining engineering. In addition, the joint prediction and continuous early warning mechanism effectively improves the stability and reliability of the system's early warning system. Attached Figure Description

[0016] Figure 1 This is a flowchart of an early identification method for abnormal evolution of battery systems for new energy mining trucks used in deep metal mines, according to the present invention. Detailed Implementation

[0017] like Figure 1 As shown, the present invention provides a method for early identification of abnormal evolution in battery systems for new energy mining trucks used in deep metal mines, comprising:

[0018] Step 1: Arrange temperature sensors and displacement sensors in the battery system to collect temperature signals and displacement signals respectively, and construct displacement input sequences and temperature input sequences with a preset time window length.

[0019] In specific implementations, the displacement sensor includes at least one of a strain sensor, a fiber Bragg grating sensor, or other sensors used to detect structural deformation. The temperature sensor includes at least one of a thermocouple, a thermistor, a digital temperature sensor, or other sensors used to detect temperature changes.

[0020] The temperature sensors are located inside the battery module, on the battery pack cover, and at key locations in the cooling system. They are used to collect temperature change information during battery operation. Specifically, the measuring points inside the battery module reflect the heat accumulation in the cell area, the measuring points on the cover characterize the overall temperature rise trend, and the measuring points in the cooling system monitor changes in heat dissipation performance.

[0021] Displacement sensors are installed on the battery module end plate, module constraint structure, support components, and battery pack shell to detect structural deformation. They are used to detect minute structural deformations of the battery under complex operating conditions. Specifically, the measuring points on the fixed structure of the module reflect changes in the internal stress state of the battery, while the measuring points on the shell side plate characterize the overall structural response of the battery pack.

[0022] In specific implementation, the construction of the displacement input sequence and temperature input sequence with a preset time window length in step 1 is as follows:

[0023] Let the displacement signal and temperature signal collected at time t be denoted as follows: and For a length of L and a step size of L, Constructing displacement input sequences within historical time windows With temperature input sequence They are respectively:

[0024]

[0025]

[0026] in, This indicates the number of displacement signals collected at different measuring point locations. This indicates the number of temperature signals collected at different measuring points.

[0027] Considering the strong vibrations and impact loads experienced by new energy mining trucks during operation, the displacement signals collected from the outer shell measuring points may contain elastic deformation responses under normal operating conditions. To improve the accuracy of anomaly identification, the displacement signals need to be preprocessed before step 2. This preprocessing includes filtering, sliding statistics, or trend decomposition to suppress high-frequency vibration interference and highlight the low-frequency variation characteristics that reflect the evolution of anomalies.

[0028] Step 2: Extract temporal features from the displacement input sequence. A temporal convolutional network is used to model the displacement sequence to obtain displacement modal features. Specifically:

[0029] In the displacement branch, considering that strain gauge displacement signals typically contain strong local dynamic changes and cross-time scale dependencies, TCN is used as the backbone feature extractor. Let the input features of the l-th layer be... ,in The l-th layer constructs a temporal representation through one-dimensional causal convolution, and its convolution output is:

[0030]

[0031] Where k represents the kernel size, Represents the expansion coefficient of the l-th layer. and These represent the convolution kernel parameters and bias terms of the l-th layer, respectively. This is a time index. Due to the use of a causal convolutional structure, the current position... The output depends only on This includes historical information, ensuring no future information is leaked during the prediction process. Furthermore, after processing with the nonlinear activation function ReLU, we obtain:

[0032]

[0033] To improve the training stability of deep networks and enhance the cumulative expressive power of features at different time scales, residual connections are used to construct the output of the l-th layer:

[0034]

[0035] go through After stacking the layers, the high-level temporal features are obtained. To obtain a compact global representation, a temporal aggregation operation is applied to the last temporal output, i.e., global average pooling is performed on the features at the last time step to obtain the feature representation of the displacement mode:

[0036]

[0037] in, This indicates a global average pooling operation, which will... As a global representation of displacement modes:

[0038] .

[0039] Step 3: Perform time-series modeling on the temperature input sequence. A Long Short-Term Memory (LSTM) network is used to model the temperature sequence to obtain temperature modal features. Specifically:

[0040] In the temperature branch, considering that temperature evolution typically exhibits stronger cumulative effects and state memory characteristics, LSTM is used to model the temperature input sequence, given the entire temperature input sequence. , record The input for each time step is The corresponding hidden states and memory units are respectively and The gating update process of LSTM is as follows:

[0041]

[0042] in, and These represent the input gate, forget gate, and output gate, respectively. , , and These represent the input weight matrices for the input gate, forget gate, output gate, and candidate memory, respectively. , , and These represent the cyclic weight matrices for the input gate, forget gate, output gate, and candidate memory, respectively. , , and These represent the biases of the input gate, forget gate, output gate, and candidate memory, respectively. Indicates candidate memory states; This represents the Hadamard product; using the above formula, the LSTM recursively obtains the hidden state sequence. , Let represent the hidden state at time step t. To maintain consistency with the displacement mode features, a linear operation is added to transform the hidden state at the final time step into a feature quantity with the same dimension as the displacement mode features, expressed as:

[0043]

[0044] Will As a temperature modal characteristic, it is denoted as:

[0045]

[0046] This representation can effectively characterize the cumulative thermal state and time-dependent structure of temperature modes within the current observation window.

[0047] Step 4: Map the displacement mode features and temperature mode features to a unified shared latent space to achieve feature alignment and then concatenate the features to construct a joint state representation. A multilayer perceptron is then used to jointly predict the battery's displacement and temperature states at the next time step. Specifically:

[0048] Step 4.1: Since TCN and LSTM construct feature extraction processes for different modalities, the resulting representations... and The features are inconsistent in distribution and semantic scale. To improve the effectiveness of heterogeneous feature fusion, two learnable mapping functions are introduced to project the two types of high-level features, displacement mode features and temperature mode features, into a unified shared latent space, namely:

[0049]

[0050]

[0051] in, To share potential spatial characteristics of displacement, For potential spatial characteristics of temperature sharing, , For the mapping matrix, , For bias terms, and Let Leaky ReLU be the nonlinear mapping function. After mapping, we have: This achieves a unified representation of displacement modal characteristics and temperature modal characteristics in the same dimensional latent space.

[0052] Step 4.2: Construct a joint state representation using feature splicing:

[0053]

[0054] Step 4.3: Based on the joint state representation, a multilayer perceptron is used to jointly predict the displacement state and temperature state of the battery at the next moment, which can be further written as:

[0055]

[0056] in, This represents the high-dimensional features of bimodal classification. ReLU represents the nonlinear activation function. , These are the weights of the first and second layers of a multilayer perceptron. , The first and second layer biases of the multilayer perceptron; and The predicted displacement and temperature values ​​at time t+1 of the battery are represented. This joint prediction structure does not simply perform independent regressions on displacement and temperature separately, but models the future evolution of both types of variables simultaneously on the basis of a shared fusion representation, thereby explicitly utilizing the coupling information between the two modes to improve the ability to approximate the normal state transition law.

[0057] Step 5: Construct a joint loss function that includes alignment loss, displacement prediction loss, and temperature prediction loss, and optimize and train the network model;

[0058] In practical implementation, to simultaneously constrain displacement prediction, temperature prediction, and cross-modal feature alignment, a joint loss function is constructed:

[0059]

[0060] Considering that displacement and temperature sequences within the same time window reflect the same battery state, they should possess consistency in high-level semantics. Therefore, an alignment constraint is introduced. This constraint prompts the model to establish a more stable cross-modal semantic correspondence between displacement and temperature while preserving modal discrimination information, thus providing a consistent representational basis for subsequent state prediction. The alignment loss in the shared space is defined as:

[0061]

[0062] Define displacement prediction loss and temperature prediction loss as follows:

[0063]

[0064] in, The latent space features are shared for the i-th displacement in a training batch. The i-th temperature-shared latent space feature in a training batch; and These are the i-th displacement prediction value and temperature prediction value in a training batch, respectively. and These are the true values ​​of the i-th displacement signal and the true value of the temperature signal in a training batch, respectively. , and These represent the weighting coefficients of the three loss terms, with values ​​of 1, 1, and 0.1 respectively. This represents the number of samples in a training batch.

[0065] Through the joint optimization of the above objective functions, the model learns the future evolution relationship of the bimodal state variables under normal operating conditions on the one hand, and establishes a consistent semantic representation of heterogeneous modes in the shared space on the other hand. Considering that abnormal samples are usually difficult to obtain in actual engineering scenarios, the training phase uses normal operating condition data for end-to-end parameter learning, so that the model can form a stable state prediction capability based on the normal mode.

[0066] Step 6: In the online inference phase, an anomaly scoring index is constructed based on the deviation between the predicted results and the actual observed values. This index is then compared to a threshold. When the anomaly score exceeds the threshold, it is determined that the battery system exhibits an abnormal evolution trend. Specifically:

[0067] Step 6.1: After training is complete, during the online inference phase, given the displacement input sequence at the current time step... With temperature input sequence The model outputs the displacement prediction and temperature prediction for the next time step; when actual observations... Upon arrival, the prediction errors for the two types of variables are calculated separately:

[0068]

[0069] in, and These represent the predicted displacement and temperature at time t+1, respectively. and These represent the true values ​​of the displacement signal and the temperature signal at time t+1, respectively. and These represent the displacement prediction error and temperature prediction error at time t+1, respectively.

[0070] Step 6.2: Under normal conditions, the system evolution follows the state transition relationships learned from the training data, thus the residuals are usually kept within a small range. However, when the battery enters an abnormal evolution phase, its displacement and temperature responses deviate from the normal coupling trajectory, leading to a significant increase in prediction error. Based on this, a comprehensive anomaly score is constructed based on displacement prediction error and temperature prediction error:

[0071]

[0072] in, and The weighting coefficients are set to 0.5, which are used to adjust the contribution of the two modes in anomaly detection.

[0073] Step 6.3: Set the discrimination threshold based on the abnormal score distribution on the normal validation set. When the following expression is satisfied, an abnormal sign is determined to exist at that moment, and an early warning is triggered:

[0074]

[0075] Step 6.4: To reduce false alarms caused by single-point noise disturbances, a continuous triggering mechanism is introduced, that is, the final alarm signal is only output when the abnormal score exceeds the threshold for a certain number of consecutive time periods.

[0076] This invention addresses the problem that when new energy mining trucks used in deep metal mines operate under complex conditions such as high temperature, high humidity, strong vibration, and high load, the internal mechanical and thermal responses of the battery system exhibit strong coupling characteristics, and the abnormal evolution is highly concealed and difficult to identify in a timely manner. It proposes an early identification method for abnormal evolution of the battery system of new energy mining trucks used in deep metal mines, thereby improving the safety and reliability of the battery system operation.

[0077] This method uses structural displacement and temperature signals collected during battery operation as inputs. It employs a temporal convolutional network for feature extraction, addressing the local dynamic changes and cross-timescale dependencies in the displacement signals. For the thermal accumulation effect and state memory characteristics of the temperature signals, a long short-term memory network is used for modeling. By performing deep feature extraction on both modes, displacement modal features and temperature modal features are obtained.

[0078] Building upon this foundation, a cross-modal feature mapping mechanism is introduced to project the two types of modal features onto a unified shared latent space. Furthermore, an intrinsic correlation between displacement and temperature responses is established through feature alignment constraints, thereby achieving a consistent representation of multimodal features in the semantic space. The aligned modal features are then fused to construct a joint state representation, which is used to jointly predict the battery's displacement and temperature states at the next time step.

[0079] During the model training phase, a joint loss function, incorporating displacement prediction error, temperature prediction error, and cross-modal alignment constraints, is constructed based on normal operating condition data. This function is then used to perform end-to-end optimization of the model, enabling it to accurately characterize the evolution of the battery system under normal conditions. During the online operation phase, an anomaly scoring index is constructed by calculating the deviation between the model's predicted values ​​and the actual observed values. When the deviation exceeds a preset threshold, an anomaly warning is triggered, thereby achieving online monitoring of the abnormal evolution process of the battery.

[0080] Furthermore, by introducing a continuous triggering mechanism, an alarm signal is only output when the abnormal score exceeds the threshold at multiple consecutive times, in order to reduce false alarms caused by environmental noise and occasional disturbances.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early identification of abnormal evolution of a new energy mine car battery system for deep metal mines, characterized in that, include: Step 1: Arrange temperature sensors and displacement sensors in the battery system to collect temperature signals and displacement signals respectively, and construct displacement input sequences and temperature input sequences with a preset time window length; Step 2: Extract temporal features from the displacement input sequence and model the displacement sequence using a temporal convolutional network to obtain displacement modal features; Step 3: Perform time-series modeling on the temperature input sequence. Use a long short-term memory network to model the temperature sequence to obtain temperature modal features; Step 4: Map the displacement mode features and temperature mode features to a unified shared latent space to achieve feature alignment and then perform feature concatenation to construct a joint state representation. Use a multilayer perceptron to jointly predict the displacement state and temperature state of the battery at the next moment. Step 5: Construct a joint loss function that includes alignment loss, displacement prediction loss, and temperature prediction loss, and optimize and train the network model; Step 6: In the online inference stage, an anomaly scoring index is constructed based on the deviation between the prediction results and the actual observations. The anomaly scoring index is then compared with a threshold. When the anomaly score exceeds the threshold, it is determined that the battery system has an abnormal evolution trend.

2. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 1, characterized in that, The temperature sensor is located inside the battery module and on the battery pack cover; the displacement sensor is located on the battery module end plate, module constraint structure, support component and battery pack shell.

3. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 1, characterized in that, The specific steps in step 1 for constructing the displacement input sequence and temperature input sequence with a preset time window length are as follows: Let the displacement signal and temperature signal collected at time t be denoted as follows: and For a length of L and a step size of L, Constructing displacement input sequences within historical time windows With temperature input sequence They are respectively: in, This indicates the number of displacement signals collected at different measuring point locations. This indicates the number of temperature signals collected at different measuring points.

4. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 1, characterized in that... Before step 2, the displacement signal is preprocessed, including filtering, sliding statistics or trend decomposition, to suppress high-frequency vibration interference and highlight the low-frequency variation characteristics that reflect the abnormal evolution.

5. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 3, characterized in that, Step 2 specifically involves: Using TCN as the backbone feature extractor, let the input features of the l-th layer be... ,in The l-th layer constructs a temporal representation through one-dimensional causal convolution, and its convolution output is: Where k represents the kernel size, Represents the expansion coefficient of the l-th layer. and These represent the convolution kernel parameters and bias terms of the l-th layer, respectively. For time indexing; further, after processing with the non-linear activation function ReLU, we obtain: To improve the training stability of deep networks and enhance the cumulative expressive power of features at different time scales, residual connections are used to construct the output of the l-th layer: go through After stacking the layers, the high-level temporal features are obtained. To obtain a compact global representation, a temporal aggregation operation is applied to the last temporal output, i.e., global average pooling is performed on the features at the last time step to obtain the feature representation of the displacement mode: in, This indicates a global average pooling operation, which will... As a global representation of displacement modes: 。 6. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 5, characterized in that, Step 3 specifically involves: LSTM is used to model the temperature input sequence, given the entire temperature input sequence. , record The input for each time step is The corresponding hidden states and memory units are respectively and The gating update process of LSTM is as follows: in, and These represent the input gate, forget gate, and output gate, respectively. , , and These represent the input weight matrices for the input gate, forget gate, output gate, and candidate memory, respectively. , , and These represent the cyclic weight matrices for the input gate, forget gate, output gate, and candidate memory, respectively. , , and These represent the biases of the input gate, forget gate, output gate, and candidate memory, respectively. Indicates candidate memory states; This represents the Hadamard product; using the above formula, the LSTM recursively obtains the hidden state sequence. , Let represent the hidden state at time step t. To maintain consistency with the displacement mode features, a linear operation is added to transform the hidden state at the final time step into a feature quantity with the same dimension as the displacement mode features, expressed as: Will As a temperature modal characteristic, it is denoted as: This representation can effectively characterize the cumulative thermal state and time-dependent structure of temperature modes within the current observation window.

7. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 6, characterized in that, Step 4 specifically involves: Step 4.1: Introduce two learnable mapping functions to project the two types of high-level features, displacement mode features and temperature mode features, into a unified shared latent space, namely: in, To share potential spatial characteristics of displacement, For potential spatial characteristics of temperature sharing, , For the mapping matrix, , For bias terms, and Let Leaky ReLU be the nonlinear mapping function. After mapping, we have: ; This achieves a unified representation of displacement modal characteristics and temperature modal characteristics in the same dimensional latent space; Step 4.2: Construct a joint state representation using feature splicing: Step 4.3: Based on the joint state representation, a multilayer perceptron is used to jointly predict the displacement state and temperature state of the battery at the next moment, which can be further written as: in, This represents the high-dimensional features of bimodal classification. ReLU represents the nonlinear activation function. , These are the weights of the first and second layers of a multilayer perceptron. , The first and second layer biases of the multilayer perceptron; and This represents the predicted displacement and temperature of the battery at time t+1.

8. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 7, characterized in that, In step 5, a joint loss function is constructed to simultaneously constrain displacement prediction, temperature prediction, and cross-modal feature alignment. The alignment loss in the shared space is defined as: Define displacement prediction loss and temperature prediction loss as follows: in, The latent space features are shared for the i-th displacement in a training batch. The i-th temperature-shared latent space feature in a training batch; and These are the i-th displacement prediction value and temperature prediction value in a training batch, respectively. and These are the true values ​​of the i-th displacement signal and the true value of the temperature signal in a training batch, respectively. , and These represent the weighting coefficients of the three loss terms. The number of samples in a training batch; By constructing a joint loss function and performing joint optimization, the model learns the future evolution relationship of bimodal state variables under normal operating conditions on the one hand, and establishes a consistent semantic representation of heterogeneous modes in the shared space on the other hand. Considering that abnormal samples are usually difficult to obtain in actual engineering scenarios, end-to-end parameter learning is carried out using normal operating condition data during the training phase, so that the model can form a stable state prediction capability based on the normal mode.

9. The method for early identification of abnormal evolution of battery systems for new energy mining trucks in deep metal mines according to claim 1, characterized in that, Step 6 specifically involves: Step 6.1: After training is complete, during the online inference phase, given the displacement input sequence at the current time step... With temperature input sequence The model outputs the displacement prediction and temperature prediction for the next time step; when actual observations... Upon arrival, the prediction errors for the two types of variables are calculated separately: in, and These represent the predicted displacement and temperature at time t+1, respectively. and These represent the true values ​​of the displacement signal and the temperature signal at time t+1, respectively. and These represent the displacement prediction error and temperature prediction error at time t+1, respectively. Step 6.2: Construct a comprehensive anomaly score based on displacement prediction error and temperature prediction error: in, and The weighting coefficients are set to 0.5, which are used to adjust the contribution of the two modes in anomaly detection. Step 6.3: Set the discrimination threshold When the following expression is satisfied, an abnormal sign is determined to exist at that moment, and an early warning is triggered: Step 6.4: To reduce false alarms caused by single-point noise disturbances, a continuous triggering mechanism is introduced, that is, the final alarm signal is only output when the abnormal score exceeds the threshold for a certain number of consecutive time periods.

Images