Engineering machine cooler temperature curve abnormality detection method and system

Abstract

The present application relates to the technical field of engineering machinery cooler temperature monitoring, and particularly relates to an engineering machinery cooler temperature curve anomaly detection method, which obtains the temperature time series data of the source domain and the target domain cooler, encodes the source domain data in multiple layers of time series features, minimizes the distribution difference with the target domain features to obtain the domain invariant feature representation, and then uses the representation to encode the feature of a small amount of labeled data in the target domain and generates a temperature mode prediction result based on the time series dependency relationship and the running state parameters, and jointly optimizes the encoding weight and the time series weight according to the prediction error. The method can realize the cooler temperature anomaly detection across working conditions and improve the detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of temperature monitoring technology for engineering machinery coolers, and in particular to a method and system for detecting abnormal temperature curves in engineering machinery coolers. Background Technology

[0002] During the operation of cooling systems in construction machinery, the cooler temperature curve is a key indicator reflecting heat dissipation efficiency and equipment health. Currently, the mainstream methods for detecting cooler temperature anomalies in the industry typically rely on statistical models or shallow machine learning classifiers built from historical operating data. For example, anomaly detection is achieved by extracting manual features such as the mean, variance, and rate of change of the temperature curve, combined with preset thresholds or Support Vector Machines (SVM); or by using Long Short-Term Memory (LSTM) networks to predict temperature time series and then judging anomalies based on the magnitude of the prediction residuals. These methods generally rely on a large amount of historical temperature data with precise operating condition labels in practical applications, and require that the training samples and the target equipment's operating environment maintain a high degree of consistency in distribution.

[0003] However, the aforementioned conventional approaches have significant limitations. First, the actual operating conditions of coolers in construction machinery are diverse, including combinations of different loads, ambient temperatures, and cooling medium flow rates. Obtaining fully labeled data covering all operating conditions is extremely costly. When data is sparse or labels are incomplete, the anomaly detection accuracy of the built model drops significantly, easily leading to missed or false alarms. Second, different models of coolers, or even the same cooler, experience slow drift in heat dissipation characteristics due to long-term wear. This makes it difficult to directly apply models trained based on data from a specific time period or device to another stage or device. Traditional methods lack effective compensation mechanisms for differences in temperature distribution across operating conditions and devices. When there is a domain shift between the deployment environment and the training environment, model performance degrades sharply, failing to meet the robustness requirements for anomaly detection in the long-term reliable operation of construction machinery. Summary of the Invention

[0004] This invention provides a method and system for detecting abnormal temperature curves in engineering machinery coolers, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for detecting abnormal temperature curves in engineering machinery coolers, comprising: The historical temperature time series data and corresponding operating condition labels of the source domain cooler are obtained, as well as a small amount of labeled temperature time series data of the target domain cooler. Multi-layer time series feature encoding is performed on the historical temperature time series data of the source domain cooler to obtain the source domain temperature pattern feature representation. By minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, the feature encoding weights are migrated and updated to obtain a domain-invariant feature representation. The domain-invariant feature representation is used to encode a small amount of labeled temperature time series data of the target domain cooler to obtain a target domain feature vector. The time series weight distribution is calculated based on the time series dependency relationship in the target domain feature vector, and the temperature pattern prediction result is generated by combining the cooler operating status parameters at the current moment. Based on the error between the temperature pattern prediction result and the actual temperature pattern label, the feature encoding weight and temporal weight distribution are jointly optimized in reverse.

[0006] Multi-layer time-series feature encoding is performed on the historical temperature time-series data of the source domain cooler to obtain the source domain temperature pattern feature representation, including: The historical temperature time series data of the source domain cooler is divided into multiple time series segments according to the time window. The first time series convolution operation is performed on each time series segment. The local change pattern between temperature values ​​at adjacent time moments is captured by the sliding convolution kernel to obtain the first layer feature mapping. Based on the first layer feature mapping, a multi-scale temporal receptive field is constructed. By parallel convolution operations with different convolution strides, short-term temperature fluctuation features and long-term temperature trend features are captured respectively, resulting in a multi-scale feature representation. The multi-scale feature representation is recalibrated at the channel level, and the channel attention coefficient is calculated based on the contribution of each scale feature to temperature pattern recognition. The channel features are then adaptively weighted. The weighted multi-scale features are spliced ​​and fused, and the fused features are projected onto a feature space of a unified dimension through a fully connected mapping to obtain the source domain temperature pattern feature representation.

[0007] By minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, the feature encoding weights are updated through migration, resulting in a domain-invariant feature representation including: A small amount of labeled temperature time-series data of the target domain cooler is encoded using the same multi-layer time-series feature encoding process as the source domain to obtain the temperature pattern feature representation of the target domain. A domain discriminator network is constructed, and the source domain temperature pattern feature representation and the target domain temperature pattern feature representation are respectively input into the domain discriminator network. The domain belonging label of the predicted feature sample is output through binary classification. The domain classification loss of the domain discriminator network is calculated. The gradient of the domain classification loss is inverted by a gradient inversion layer, and the inverted gradient is backpropagated to the weight parameters of the feature encoder, so that the optimization direction of the feature encoder is against the optimization direction of the domain discriminator. During the adversarial optimization process, the source domain temperature pattern classification loss is minimized and the domain classification confusion is maximized simultaneously. When the domain discriminator network cannot effectively distinguish between the features of the source domain and the target domain, the iterative optimization is terminated. Extract the feature encoder weight parameters after the adversarial optimization convergence, and use the feature encoder weight parameters to perform feature encoding on the temperature time series data of the source domain and the target domain to obtain the neighborhood-invariant feature representation.

[0008] A domain discriminator network is constructed, and the source domain temperature pattern feature representation and the target domain temperature pattern feature representation are respectively input into the domain discriminator network. The predicted domain attribution labels of the feature samples are output through binary classification, including: A domain discriminator network containing multiple fully connected layers is constructed. The source domain temperature pattern feature representation and the target domain temperature pattern feature representation are merged to form a hybrid feature sample set. Each feature sample is labeled with a domain source identifier. The source domain feature sample is labeled as the first domain identifier, and the target domain feature sample is labeled as the second domain identifier. The feature samples in the hybrid feature sample set are sequentially input into the first fully connected layer of the domain discriminator network, and the feature vectors are linearly mapped and nonlinearly transformed by nonlinear activation units to obtain the first layer of hidden feature representation. The first hidden feature representation is input into the second fully connected layer for further feature space mapping. By compressing the feature dimension layer by layer, the discriminative feature representation is obtained. The discriminative feature is represented as the input-output fully connected layer. The value of the output neuron is converted into the domain belonging probability through the activation function. The domain identifier corresponding to the output neuron with the highest domain belonging probability is selected as the predicted domain belonging label.

[0009] The target domain feature vector is obtained by using the domain-invariant feature representation to encode a small amount of labeled temperature time-series data of the cooler in the target domain. The temporal weight distribution is calculated based on the temporal dependencies in the target domain feature vector, and a temperature pattern prediction result is generated by combining the cooler's operating state parameters at the current moment. Input a small amount of labeled temperature time-series data of the target domain cooler into the feature encoder that has completed the migration update to obtain the target domain feature vector sequence; A temporal self-attention computation unit is constructed to calculate the similarity score between each feature vector in the target domain feature vector sequence and all feature vectors in the sequence, thereby obtaining the temporal attention matrix; The temporal attention matrix is ​​normalized row by row to obtain a temporal weight distribution. The target domain feature vector sequence is weighted and aggregated according to the temporal weight distribution. The aggregated feature vector is then concatenated with the cooler operating state parameters at the current moment to obtain a fused feature vector. The fused feature vector is input into the temperature pattern classifier, and mapped to the temperature pattern category probability distribution through a fully connected layer and activation function. The category with the highest probability is selected as the temperature pattern prediction result.

[0010] A temporal self-attention computation unit is constructed to calculate the similarity score between each feature vector in the target domain feature vector sequence and all feature vectors in the sequence, resulting in a temporal attention matrix including: The target domain feature vector sequence is linearly transformed by the query mapping matrix, key mapping matrix and value mapping matrix respectively to obtain the query vector sequence, key vector sequence and value vector sequence; For the first query vector in the query vector sequence, calculate the inner product of the first query vector and all key vectors in the key vector sequence to obtain the similarity score vector corresponding to the first query vector; The similarity score vector is scaled by dividing each score value by the square root of the key vector dimension to avoid the inner product value being too large and causing the gradient to vanish. The scaled similarity score vector is subjected to exponential normalization to obtain the attention weight vector corresponding to the first query vector. A temporal position encoding bias term is introduced, and a position encoding vector is generated based on the temporal position index of the feature vector in the sequence. The position encoding vector is then added to the calculation process of the similarity score vector, so that feature vectors that are temporally adjacent obtain higher similarity scores. Repeat the above calculation process to traverse all query vectors in the query vector sequence, and stack the attention weight vectors corresponding to all query vectors row by row to obtain the temporal attention matrix.

[0011] A second aspect of the present invention provides a system for detecting abnormal temperature curves in engineering machinery coolers, comprising: The data encoding unit is used to acquire historical temperature time-series data of the source domain cooler and the corresponding operating condition label, as well as a small amount of labeled temperature time-series data of the target domain cooler, and to perform multi-layer time-series feature encoding on the historical temperature time-series data of the source domain cooler to obtain the source domain temperature pattern feature representation. The migration update unit is used to migrate and update the feature encoding weights by minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, so as to obtain a domain-invariant feature representation. The prediction generation unit is used to encode a small amount of labeled temperature time series data of the target domain cooler using the domain-invariant feature representation to obtain a target domain feature vector, calculate the time series weight distribution based on the time series dependency relationship in the target domain feature vector, and generate a temperature pattern prediction result by combining the cooler operating status parameters at the current moment. The reverse optimization unit is used to perform joint reverse optimization of the feature encoding weights and temporal weight distributions based on the error between the temperature pattern prediction results and the actual temperature pattern labels.

[0012] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] This method significantly reduces the reliance on labeled temperature data of coolers in the target domain, achieving high-precision anomaly detection of temperature curves with only a small number of labeled samples. Through cross-domain feature transfer and alignment, it effectively solves the problem of feature distribution shift caused by differences in operating conditions, models, or usage environments of different coolers, allowing for full utilization of rich historical data in the source domain and greatly improving the model's generalization ability and detection stability in small sample scenarios.

[0015] The extraction of domain-invariant feature representations enables the detection model to accurately identify essential anomalies in temperature patterns, avoiding false alarms and missed detections caused by changes in operating conditions. By combining temporal weight distribution to dynamically strengthen key dependencies in temperature patterns, this method can effectively capture abnormal patterns such as local abrupt changes and trend deviations in temperature curves over time, and its sensitivity to early fault signals is significantly better than traditional fixed threshold methods.

[0016] The temperature pattern prediction results generated by fusing temporal dependencies and current operating state parameters exhibit good real-time performance and adaptability. The joint inverse optimization mechanism can continuously iterate the feature encoding weights and temporal weight distributions, enabling the model to maintain stable detection accuracy even when dealing with dynamic conditions such as cooler performance degradation and seasonal environmental changes, thereby extending equipment maintenance cycles and reducing the risk of unplanned downtime. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for detecting abnormal temperature curves in engineering machinery coolers according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the method for detecting abnormal temperature curves in engineering machinery coolers according to an embodiment of the present invention.

[0021] Methods for detecting abnormal temperature profiles in construction machinery coolers include: The historical temperature time series data and corresponding operating condition labels of the source domain cooler are obtained, as well as a small amount of labeled temperature time series data of the target domain cooler. Multi-layer time series feature encoding is performed on the historical temperature time series data of the source domain cooler to obtain the source domain temperature pattern feature representation. By minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, the feature encoding weights are migrated and updated to obtain a domain-invariant feature representation. The domain-invariant feature representation is used to encode a small amount of labeled temperature time series data of the target domain cooler to obtain a target domain feature vector. The time series weight distribution is calculated based on the time series dependency relationship in the target domain feature vector, and the temperature pattern prediction result is generated by combining the cooler operating status parameters at the current moment. Based on the error between the temperature pattern prediction result and the actual temperature pattern label, the feature encoding weight and temporal weight distribution are jointly optimized in reverse.

[0022] In one optional implementation, multi-layer time-series feature encoding is performed on the historical temperature time-series data of the source domain cooler to obtain a source domain temperature pattern feature representation, including: The historical temperature time series data of the source domain cooler is divided into multiple time series segments according to the time window. The first time series convolution operation is performed on each time series segment. The local change pattern between temperature values ​​at adjacent time moments is captured by the sliding convolution kernel to obtain the first layer feature mapping. Based on the first layer feature mapping, a multi-scale temporal receptive field is constructed. By parallel convolution operations with different convolution strides, short-term temperature fluctuation features and long-term temperature trend features are captured respectively, resulting in a multi-scale feature representation. The multi-scale feature representation is recalibrated at the channel level, and the channel attention coefficient is calculated based on the contribution of each scale feature to temperature pattern recognition. The channel features are then adaptively weighted. The weighted multi-scale features are spliced ​​and fused, and the fused features are projected onto a feature space of a unified dimension through a fully connected mapping to obtain the source domain temperature pattern feature representation.

[0023] After acquiring the historical temperature time-series data of the source domain cooler, systematic feature extraction processing is required. Historical temperature time-series data is typically recorded at a fixed sampling frequency, such as collecting temperature values ​​from multiple measuring points (cooler outlet temperature, inlet temperature, and ambient temperature) once per second. This raw data constitutes a multi-dimensional time series, which contains the temperature variation patterns of the cooler under different operating conditions.

[0024] To effectively capture temporal patterns in temperature data, continuous historical temperature time-series data is first segmented according to a preset time window. The length of the time window is determined based on the thermal response characteristics of the cooler, typically set to 30 to 120 seconds. Assuming a sampling frequency of 1 Hz and a time window length of 60 seconds, each time segment contains 60 consecutive temperature sampling points. The complete historical data sequence was divided using a sliding window method, with the sliding step size set to 50% of the time window length to ensure appropriate overlap between adjacent segments and thus avoid the loss of boundary information.

[0025] When performing the first layer of temporal convolution for each time segment, a one-dimensional convolution kernel is used to slide along the time dimension. The kernel size is typically set to 3 to 7 time steps to capture local variation patterns between adjacent temperature values. Specifically, for the input time segment... ,in Indicates the number of temperature measuring points. This represents the number of sampling points within the time window. The first convolution operation can be represented as... ,in This is the weight matrix of the first layer convolution kernel. For bias vectors, This represents a convolution operation, with ReLU as the activation function. This convolution operation extracts the basic change patterns of the temperature curve over a short timescale, such as rising, falling, and plateauing. The number of output channels in the first feature map layer is typically set to 32 to 64, with each output channel corresponding to a specific local temperature change pattern.

[0026] After obtaining the first layer of feature maps, a multi-scale temporal receptive field structure is constructed to simultaneously capture temperature change patterns at different time scales. This structure employs parallel convolutional branches, each using different convolutional parameter configurations. Specifically, three parallel branches are set up: the first branch uses a convolutional layer with a kernel size of 3 and a stride of 1 to capture short-term temperature fluctuation features, which reflect the cooler's rapid response to sudden load changes; the second branch uses a convolutional layer with a kernel size of 5 and a stride of 2 to capture medium-term temperature change trends; and the third branch uses a convolutional layer with a kernel size of 7 and a stride of 3 to capture long-term temperature evolution patterns, which reflect the cooler's thermal inertia characteristics. The output feature maps of each branch have different temporal resolutions; the short-term branch retains more temporal details, while the long-term branch performs greater temporal compression on the temperature sequence.

[0027] To enable subsequent fusion of features extracted from different branches, the outputs of each branch need to be aligned in the temporal dimension. Adaptive pooling is used to unify feature maps with different temporal resolutions to the same time length. Assume the time length of the feature map output by the first branch is... The second branch is The third branch is They are uniformly adjusted to the target length through adaptive average pooling. The adjusted multi-scale feature representation is denoted as... These correspond to short-term, medium-term, and long-term characteristics, respectively.

[0028] When performing channel-level feature recalibration on multi-scale feature representations, a channel attention mechanism is employed to evaluate the contribution of features at each scale to temperature pattern recognition. First, global average pooling is performed on the feature maps at each scale, compressing the temporal information into a single channel descriptor. For the feature maps... The global average pooling operation is calculated as follows: , thus obtaining the channel description vector The channel description vectors at each scale are then concatenated and input into a two-layer fully connected network. The first layer performs dimensionality reduction and compression using ReLU activation. The second layer restores the original number of channels and uses the Sigmoid activation function, outputting channel attention coefficients. Each element has a value ranging from 0 to 1, representing the importance weight of the corresponding channel feature.

[0029] The calculation process of the channel attention coefficient can be expressed as follows: ,in This represents a vector concatenation operation. and These are the weight matrices for the two fully connected network layers, This represents the Sigmoid function. After obtaining the channel attention coefficients, element-wise multiplication is performed on each channel of the feature map at each scale to achieve adaptive weighting. ,in This represents a broadcast multiplication along the channel dimension. This operation can enhance feature channels that contribute significantly to temperature pattern recognition while suppressing redundant or noisy channels.

[0030] After completing the channel-level recalibration, the weighted multi-scale features will be... The data is spliced ​​and fused along the channel dimension to obtain a comprehensive feature representation. Because the concatenated features have a high dimensionality, using them directly would lead to excessive computational complexity. Therefore, a fully connected mapping layer is used to project the fused features onto a feature space of uniform dimension. Specifically, the first step is to... Global average pooling is performed over the time dimension to obtain a fixed-length feature vector. Then, the vector is mapped to the target dimension through a fully connected layer. Typically set to 128 or 256 dimensions, the mapping operation is represented as ,in Let be the projection weight matrix. This is the bias vector.

[0031] The obtained feature vector This is the source domain temperature pattern feature representation, which integrates temperature change information across multiple time scales and highlights the most valuable feature components for pattern recognition through an attention mechanism. For each time segment in the source domain, a corresponding feature representation vector is generated. These feature vectors not only preserve the key variation patterns of the original temperature data but also achieve a mapping from the original signal space to the abstract feature space through a multi-layer coding process, providing a high-quality input representation for subsequent domain adaptation and anomaly detection.

[0032] In practical applications, the weights of the first-layer convolutional kernel are learned using the backpropagation algorithm, and the Xavier initialization method is used during initialization to ensure training stability. The kernel parameters of the multi-scale branches are trained independently, allowing each branch to focus on feature extraction at a specific time scale. The parameters of the fully connected layers of the channel attention network are also obtained through end-to-end training, with the training objective being to enable the recalibrated feature representation to better distinguish different temperature pattern categories. The entire feature encoding process employs batch processing, processing 32 to 64 time segments each time to improve computational efficiency and utilize batch normalization techniques to stabilize the training process.

[0033] In one optional implementation, the feature encoding weights are updated by minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, resulting in a domain-invariant feature representation including: A small amount of labeled temperature time-series data of the target domain cooler is encoded using the same multi-layer time-series feature encoding process as the source domain to obtain the temperature pattern feature representation of the target domain. A domain discriminator network is constructed, and the source domain temperature pattern feature representation and the target domain temperature pattern feature representation are respectively input into the domain discriminator network. The domain belonging label of the predicted feature sample is output through binary classification. The domain classification loss of the domain discriminator network is calculated. The gradient of the domain classification loss is inverted by a gradient inversion layer, and the inverted gradient is backpropagated to the weight parameters of the feature encoder, so that the optimization direction of the feature encoder is against the optimization direction of the domain discriminator. During the adversarial optimization process, the source domain temperature pattern classification loss is minimized and the domain classification confusion is maximized simultaneously. When the domain discriminator network cannot effectively distinguish between the features of the source domain and the target domain, the iterative optimization is terminated. Extract the feature encoder weight parameters after the adversarial optimization convergence, and use the feature encoder weight parameters to perform feature encoding on the temperature time series data of the source domain and the target domain to obtain the neighborhood-invariant feature representation.

[0034] After obtaining the multi-scale temperature pattern feature representation of the source domain cooler, this feature representation needs to be transferred to the target domain cooler to solve the problem of inconsistent temperature data distribution caused by differences in operating conditions and sensor characteristics among different engineering machinery equipment. The target domain cooler usually only has a small amount of labeled data and cannot independently train an effective temperature anomaly detection model. Therefore, domain adaptation technology is used to align the feature space, so that the temperature pattern knowledge learned in the source domain can be effectively transferred to the target domain.

[0035] For the limited amount of labeled temperature time-series data collected from the target domain cooler, a multi-layer temporal feature encoding process identical to that used in the source domain is employed. Specifically, the target domain temperature data is segmented according to the same time window length, and short-term, medium-term, and long-term temperature change features are extracted using the same three-branch multi-scale convolutional structure. These features are then subjected to the same feature fusion and dimensionality reduction operations to obtain the target domain temperature pattern feature representation. At this point, the source domain feature representation and the target domain feature representation reside in the same feature space dimension. However, due to the difference in data distribution between the two domains, the distribution position of the feature representation in this space is significantly offset, causing the classifier trained in the source domain to be unsuitable for direct application to the target domain data.

[0036] To align feature distributions across domains, a domain discriminator network is constructed as a core component of adversarial learning. The domain discriminator employs a multi-layer fully connected neural network structure. Its input is the temperature pattern feature representation output from the feature encoder, and its output is a binary classification probability used to determine whether the input feature originates from the source or target domain. The source domain temperature pattern feature representation is labeled as domain label 0, and the target domain temperature pattern feature representation is labeled as domain label 1. The domain discriminator is trained through supervised learning to accurately distinguish features between the two domains. The loss function of the domain discriminator uses a binary cross-entropy form to calculate the difference between the predicted and true domain labels.

[0037] While training the domain discriminator, an adversarial optimization mechanism for the feature encoder is implemented through a gradient inversion layer. The gradient inversion layer is a special network layer that directly transmits features as an identity mapping during forward propagation, but inverts the gradient during backpropagation. Specifically, after the domain discriminator calculates the domain classification loss, the gradient of this loss with respect to the feature encoder weights is multiplied by a negative number when passing through the gradient inversion layer, causing the feature encoder's parameter update direction to be opposite to the domain discriminator's optimization direction. This adversarial mechanism prompts the feature encoder to learn a domain-invariant feature representation, meaning the extracted features retain both the discriminative information of the temperature pattern and eliminate domain-specific distribution differences.

[0038] During adversarial optimization, the feature encoder simultaneously optimizes two objectives. The first objective is to minimize the source domain temperature pattern classification loss, ensuring that the extracted features can accurately distinguish between normal and abnormal temperature patterns. The source domain classification loss is obtained by inputting the source domain features into the temperature pattern classifier and calculating the cross-entropy loss between the predicted and true labels. The second objective is to maximize the domain classification confusion, i.e., by using a gradient inversion mechanism to prevent the domain discriminator from accurately determining the domain from which the features originate. When the feature encoder successfully extracts domain-invariant features, the distribution of source and target domain features in the common feature space tends to be consistent, and the classification accuracy of the domain discriminator drops to approximately 50% of that of random guessing.

[0039] The iterative process of adversarial optimization employs an alternating training strategy. In each training batch, the weight parameters of the feature encoder are first fixed, and only the parameters of the neighborhood discriminator are updated to improve its ability to distinguish the neighborhood. Subsequently, the parameters of the neighborhood discriminator are fixed, and the parameters of the feature encoder are updated through a gradient inversion layer to generate more confusing feature representations. This alternating optimization process is similar to the training mechanism of generative adversarial networks, where the feature encoder and the neighborhood discriminator form a dynamic game relationship, eventually reaching a Nash equilibrium.

[0040] To monitor the convergence of adversarial optimization, domain discrimination accuracy is set as a termination condition. In the early stages of training, the domain discriminator can easily distinguish features between the source and target domains, with a discrimination accuracy approaching 100%. As adversarial training progresses, the feature encoder gradually learns domain-invariant feature representations, and the accuracy of the domain discriminator continuously decreases. When the domain discrimination accuracy stabilizes around 50% over several consecutive training epochs, and the fluctuation range is less than a preset threshold, adversarial optimization is considered to have converged. At this point, the features extracted by the feature encoder have achieved effective alignment between domains.

[0041] After the adversarial optimization converges, all weight parameters of the feature encoder are extracted, including the kernel weights of the multi-scale convolutional layers, the scaling and offset parameters of the batch normalization layers, and the weight matrix of the fully connected layers. These weight parameters are used to construct a domain-invariant feature extractor to uniformly encode the temperature time-series data from both the source and target domains. The domain-adaptive feature encoder automatically filters out domain-specific information related to specific equipment and operating conditions, retaining only the essential features related to temperature anomaly patterns. This domain-invariant feature representation has good generalization ability, allowing the temperature anomaly detection model trained on source domain data to be directly applied to the target domain cooler without requiring large-scale data annotation and model retraining in the target domain.

[0042] The hyperparameter settings in the adversarial learning process have a significant impact on the final domain adaptation effect. The inversion coefficient of the gradient inversion layer controls the adversarial strength. This coefficient is usually dynamically adjusted during training. A smaller value is used in the early stages of training to ensure the stability of the feature encoder, and then gradually increased to enhance the domain alignment effect. The network depth and hidden layer dimension of the domain discriminator need to be adjusted according to the complexity of the feature space. An overly deep discriminator leads to overfitting, while an overly shallow discriminator cannot effectively capture domain differences. The weight balance coefficient between the source domain classification loss and the domain adversarial loss determines the trade-off between maintaining discriminative ability and achieving domain alignment. This coefficient is usually fine-tuned by evaluating performance on the validation set.

[0043] Through the aforementioned adversarial learning mechanism, feature transfer from the source domain to the target domain was achieved. The resulting domain-invariant feature representation not only retains the discriminative information of temperature anomaly patterns but also eliminates the distribution differences between devices, laying the foundation for subsequent fine-tuning based on a small amount of target domain labeled data.

[0044] In one optional implementation, a domain discriminator network is constructed, and the source domain temperature pattern feature representation and the target domain temperature pattern feature representation are respectively input into the domain discriminator network. The domain attribution label of the predicted feature sample is output through binary classification, including: A domain discriminator network containing multiple fully connected layers is constructed. The source domain temperature pattern feature representation and the target domain temperature pattern feature representation are merged to form a hybrid feature sample set. Each feature sample is labeled with a domain source identifier. The source domain feature sample is labeled as the first domain identifier, and the target domain feature sample is labeled as the second domain identifier. The feature samples in the hybrid feature sample set are sequentially input into the first fully connected layer of the domain discriminator network, and the feature vectors are linearly mapped and nonlinearly transformed by nonlinear activation units to obtain the first layer of hidden feature representation. The first hidden feature representation is input into the second fully connected layer for further feature space mapping. By compressing the feature dimension layer by layer, the discriminative feature representation is obtained. The discriminative feature is represented as the input-output fully connected layer. The value of the output neuron is converted into the domain belonging probability through the activation function. The domain identifier corresponding to the output neuron with the highest domain belonging probability is selected as the predicted domain belonging label.

[0045] In the transfer learning framework for detecting anomalies in the temperature curves of engineering machinery coolers, the construction and training of the domain discriminator network is a crucial step in achieving domain adaptation. The role of the domain discriminator is to promote the feature representation extracted by the feature encoder to have domain invariance between the source and target domains through adversarial training mechanisms, thereby eliminating data distribution differences caused by different working conditions, different equipment models, or different operating environments.

[0046] The domain discriminator network employs a multi-layer fully connected neural network architecture. Specifically, this network includes an input layer, several hidden layers, and an output layer. The input layer receives temperature pattern feature representations from the output of the feature encoder; these feature representations typically have a dimension of [missing information]. This dimension is consistent with the feature dimension after multi-scale fusion in the aforementioned feature encoding process. The number of hidden layers is set according to the feature complexity and the difficulty of the discrimination task, typically between 2 and 4 layers. The number of neurons in each hidden layer decreases progressively; for example, the first hidden layer has 512 neurons, the second hidden layer has 256, and the third hidden layer has 128, achieving feature space compression and abstraction through layer-by-layer dimensionality reduction. The output layer contains two neurons, corresponding to the source domain and the target domain, respectively.

[0047] During training, the first step is to construct a hybrid feature sample set. The feature samples extracted from the historical temperature time-series data of the source domain cooler are denoted as follows: ,in Indicates the source domain sample index. , Let be the total number of samples in the source domain. The feature samples extracted from the limited labeled temperature time-series data of the target domain cooler are denoted as . ,in Indicates the target domain sample index. , This represents the total number of samples in the target domain. These two types of feature samples are then merged to form a mixed feature sample set. .

[0048] Each feature sample is labeled with a domain source identifier; source domain feature samples The first domain identifier is labeled with the value 0; target domain feature samples. The label is designated as the second domain identifier and represented by the value 1. This binary labeling method transforms the domain discrimination task into a standard binary classification problem. During actual training, the mixed feature sample set is randomly shuffled to avoid the network learning spurious patterns related to sample order.

[0049] Feature samples from the mixed feature sample set are sequentially input into the first fully connected layer of the neighborhood discriminator network. The weight matrix of the first fully connected layer is denoted as... The bias vector is denoted as ,in This represents the number of neurons in the first hidden layer. For the input feature vector... The linear mapping operation is calculated as follows: The result after linear mapping undergoes a nonlinear transformation using nonlinear activation units. A commonly used activation function is the modified linear unit, whose mathematical expression is: After processing with the activation function, the first layer of hidden feature representation is obtained. To enhance the generalization ability of the network, dropout regularization is typically introduced after the first layer of hidden feature representations, using probability... The activation values ​​of some neurons are randomly dropped, and the dropout probability is usually set to 0.3 to 0.5.

[0050] First layer hidden feature representation The second fully connected layer is then input for further feature space mapping. The weight matrix of the second fully connected layer is denoted as... The bias vector is denoted as ,in The number of neurons in the second hidden layer, and This achieves feature dimension compression. The linear mapping and activation operations in the second layer are as follows: , Similarly, after the second hidden feature representation, a dropout operation is introduced, with the dropout probability denoted as... .

[0051] If the network contains a third hidden layer, then the second hidden feature representation... Continue with the third fully connected layer. The weight matrix of the third fully connected layer is denoted as... The bias vector is denoted as ,in The number of neurons in the third hidden layer, and The calculation process for the third layer is as follows: , After layer-by-layer dimensionality reduction through multiple fully connected layers, the feature dimension is reduced from the initial... Gradually compress to The discriminant feature representation is obtained. This layer-by-layer dimensionality reduction design allows the network to retain key discriminative information while filtering out redundant feature components.

[0052] Discriminant feature representation The input and output fully connected layers have a weight matrix denoted as: The bias vector is denoted as The linear mapping of the output layer is calculated as follows: This yields a two-dimensional output vector. ,in The original output value corresponding to the source domain, The original output value corresponding to the target domain.

[0053] To convert the output neuron's value into a neighborhood attribution probability, the output vector is normalized using the softmax activation function. The mathematical expression for the softmax function is: ,in This represents the index of the output neuron. After softmax activation, the neighborhood assignment probability vector is obtained. ,in This represents the probability that a feature sample belongs to the source domain. This represents the probability that a feature sample belongs to the target domain, and satisfies... .

[0054] The domain identifier corresponding to the output neuron with the highest domain attribution probability is selected as the predicted domain attribution label. Specifically, if... If the feature sample is from the source domain, the predicted label is 0; if If the predicted feature sample originates from the target domain, the predicted label is 1. The predicted label is denoted as... .

[0055] The domain discriminator network is trained using a binary cross-entropy loss function. For a single feature sample, its true domain identifier is denoted as... The predicted domain affiliation probability is The discriminant loss for this sample is calculated as follows: For the entire mixed feature sample set, the discriminant loss is the average of the discriminant losses of all samples.

[0056] In the adversarial training framework, the domain discriminator network and the feature encoder form a game-like relationship. The domain discriminator's optimization goal is to maximize the discrimination accuracy, i.e., correctly distinguishing source domain features from target domain features; while the feature encoder's optimization goal is to minimize the discrimination accuracy, even if the extracted features cannot distinguish the source domain from the perspective of the domain discriminator. This adversarial mechanism is achieved through a gradient inversion layer. During forward propagation, the gradient inversion layer does not change the feature representation; during backward propagation, the gradient inversion layer multiplies the gradient from the discriminator by a negative number and then passes it to the feature encoder, thus achieving adversarial training. The mathematical expression for gradient inversion is: ,in To counteract the training intensity coefficient, it is usually set to a value between 0.1 and 1.0.

[0057] Through the construction of the aforementioned domain discriminator network and the adversarial training mechanism, the feature encoder can learn domain-invariant temperature pattern feature representations, enabling the model trained on source domain data to be effectively transferred to the target domain cooler temperature curve anomaly detection task, significantly reducing the amount of target domain labeled data and improving the model's generalization ability on different engineering machinery equipment.

[0058] In one optional implementation, the target domain cooler's limited labeled temperature time-series data is feature-encoded using the domain-invariant feature representation to obtain a target domain feature vector. A time-series weight distribution is calculated based on the time-series dependencies in the target domain feature vector, and a temperature pattern prediction result is generated by combining the cooler's current operating state parameters. Input a small amount of labeled temperature time-series data of the target domain cooler into the feature encoder that has completed the migration update to obtain the target domain feature vector sequence; A temporal self-attention computation unit is constructed to calculate the similarity score between each feature vector in the target domain feature vector sequence and all feature vectors in the sequence, thereby obtaining the temporal attention matrix; The temporal attention matrix is ​​normalized row by row to obtain a temporal weight distribution. The target domain feature vector sequence is weighted and aggregated according to the temporal weight distribution. The aggregated feature vector is then concatenated with the cooler operating state parameters at the current moment to obtain a fused feature vector. The fused feature vector is input into the temperature pattern classifier, and mapped to the temperature pattern category probability distribution through a fully connected layer and activation function. The category with the highest probability is selected as the temperature pattern prediction result.

[0059] After completing the domain-adaptive feature encoding, a refined analysis of the temperature time-series data of the target domain cooler is required to achieve accurate temperature pattern recognition. The temperature time-series data collected from the target domain cooler is input into the feature encoder, which has undergone transfer updating. The encoder's weight parameters have been optimized through domain adversarial training, enabling it to extract domain-invariant temperature change pattern features. Assume the dimension of the target domain input data is... ,in This represents the number of time steps in the target domain samples. As the feature dimension, after processing by the feature encoder, the target domain feature vector sequence is obtained. Each feature vector The dimension is ,in Indicates the time step index.

[0060] To capture the dependencies between different time points in temperature time-series data, a temporal self-attention computation unit is constructed. This unit quantifies the strength of temporal dependencies by calculating the similarity between feature vectors. For any feature vector in the feature vector sequence... The similarity score between the vector and all feature vectors in the sequence is calculated. Specifically, this is first achieved through three independent linear transformation matrices. , and The feature vectors are mapped to query vectors, key vectors, and value vectors, respectively, with each of these three transformation matrices having a dimension of 1. ,in This represents the dimension of the attention computation space. Query vector. The key vector is used to represent the query request at the current moment. Feature identifiers and value vectors used to represent each time step. Used to represent the actual characteristics at each moment, where Represents the index of any time step in the sequence.

[0061] Calculate query vector With all key vectors The dot product is used to obtain the original similarity score. To avoid the gradient vanishing effect caused by an excessively large dot product, the similarity score is scaled. The scaled score is... The scaled similarity scores for all time steps are organized into a matrix to obtain the temporal attention matrix. Elements in the matrix Indicates time step time step The original intensity of attention.

[0062] The temporal attention matrix is ​​normalized row-wise to convert the original similarity scores into a probability distribution. For the nth row of the matrix... Okay, let's apply the softmax function for normalization. The calculation formula is: ,in For summation index variables, Indicates time step time step The normalized attention weights. The normalized weights satisfy... This ensures the probabilistic nature of the weight distribution. All normalized weights are organized into a time-series weight distribution matrix. Matrix elements .

[0063] The target domain feature vector sequence is weighted and aggregated according to the temporal weight distribution, for each time step. Its weighted aggregated feature vector is calculated as follows: This vector integrates feature information from all time points in the sequence, and the aggregation weights are determined by temporal dependencies. This weighted aggregation mechanism enables the model to automatically focus on historical moments that are strongly correlated with the current temperature pattern, thereby improving the accuracy of temperature pattern recognition.

[0064] To further enhance the model's predictive ability, the weighted aggregated feature vector is fused with the cooler's operating state parameters at the current moment. These parameters include physical quantities such as coolant flow rate, ambient temperature, engine load rate, and fan speed, which directly affect cooler temperature changes. Assume the current operating state parameter vector is... ,in This represents the dimension of the state parameters. To ensure consistency in feature dimensions, a linear transformation layer is first used to map the state parameter vector to the same dimensional space as the weighted aggregated feature vector. The mapped state features are... ,in The transformation matrix is... This is the bias vector. The mapped state features are concatenated with the weighted aggregated feature vector to obtain the fused feature vector. The semicolon indicates a vector concatenation operation.

[0065] The fused feature vector is input into the temperature pattern classifier for final pattern recognition. The classifier employs a multi-layer fully connected neural network structure. The first fully connected layer maps the fused feature vector to the intermediate representation space, and the calculation formula is as follows: ,in This is the first layer weight matrix. For bias vectors, This represents the hidden layer dimension, and ReLU is the modified linear unit activation function. To prevent overfitting, dropout regularization is applied after the first layer output, with probability... The activation values ​​of some neurons are randomly discarded. The second fully connected layer further extracts abstract features, calculated using the following formula: ,in This is the weight matrix for the second layer. This is the bias vector.

[0066] The final fully connected layer maps features to the temperature pattern category space, assuming there are a total of temperature patterns. The output layer is calculated using the following formula: (Categories: Normal operation mode, Mild anomaly mode, Moderate anomaly mode, and Severe anomaly mode). ,in This is the output layer weight matrix. For bias vectors, This is the original output vector. Applying the softmax activation function to the original output vector transforms it into a temperature pattern category probability distribution, calculated using the following formula: ,in Indicates category index, For summation index variables, Represents the output vector The One element, This indicates that, given a fused feature vector Under the condition, time step Belongs to the The probability of each temperature pattern category.

[0067] The category with the highest probability is selected as the temperature model prediction result, i.e. ,in Indicates time step The model predicts temperature pattern categories. In this way, it can comprehensively utilize temporal dependencies and operational status information to accurately identify the temperature pattern of the target domain cooler. In practical applications, corresponding early warning mechanisms can be triggered based on the prediction results. For example, when the prediction result indicates a severely abnormal pattern, an alarm message can be immediately sent to the operator, suggesting measures such as reducing the load or shutting down for inspection, thereby effectively preventing equipment damage caused by cooling system failures.

[0068] In one optional implementation, a temporal self-attention computation unit is constructed to calculate the similarity score between each feature vector in the target domain feature vector sequence and all feature vectors in the sequence, resulting in a temporal attention matrix including: The target domain feature vector sequence is linearly transformed by the query mapping matrix, key mapping matrix and value mapping matrix respectively to obtain the query vector sequence, key vector sequence and value vector sequence; For the first query vector in the query vector sequence, calculate the inner product of the first query vector and all key vectors in the key vector sequence to obtain the similarity score vector corresponding to the first query vector; The similarity score vector is scaled by dividing each score value by the square root of the key vector dimension to avoid the inner product value being too large and causing the gradient to vanish. The scaled similarity score vector is subjected to exponential normalization to obtain the attention weight vector corresponding to the first query vector. A temporal position encoding bias term is introduced, and a position encoding vector is generated based on the temporal position index of the feature vector in the sequence. The position encoding vector is then added to the calculation process of the similarity score vector, so that feature vectors that are temporally adjacent obtain higher similarity scores. Repeat the above calculation process to traverse all query vectors in the query vector sequence, and stack the attention weight vectors corresponding to all query vectors row by row to obtain the temporal attention matrix.

[0069] After obtaining the feature vector sequence of the target domain, a temporal self-attention computation unit needs to be constructed to capture the temporal dependencies within the sequence. The core idea of ​​this computation unit is to enable each feature vector in the sequence to pay attention to all other feature vectors in the sequence, thereby establishing a global temporal correlation pattern.

[0070] First, the target domain feature vector sequence is input into three independent linear transformation layers. It is assumed that the target domain feature vector sequence contains... There are n feature vectors, each feature vector having a dimension of n. Define the query mapping matrix. Key mapping matrix and value mapping matrix ,in , and These represent the dimensions of the query vector, key vector, and value vector, respectively. In practical applications, they are typically set to... To simplify calculations. For the first... eigenvectors The corresponding query vector is obtained through linear transformation. Key vector and value vector After traversing the entire sequence, the query vector sequence is obtained. Key vector sequence and value vector sequence .

[0071] Next, we will analyze the first query vector in the query vector sequence. Perform similarity calculation. The similarity score vector is obtained by multiplying the product of the key vectors with the dot product of all key vectors in the key vector sequence. Specifically, for the th key vector in the key vector sequence... Key vectors Calculate the inner product After iterating through all key vectors, the similarity score vector corresponding to the first query vector is obtained. Each element of this vector represents the similarity between the first query vector and the corresponding key vector; a larger value indicates a higher similarity.

[0072] Because inner product operations produce large values, especially when the key vector dimension is large. When the value is large, it can lead to the vanishing gradient problem in subsequent softmax operations. To avoid this, the similarity score vector is scaled. Each element in the score vector is divided by... The scaled similarity score is obtained. This scaling operation keeps the inner product result within a reasonable range, ensuring that the gradient remains stable during backpropagation.

[0073] The scaled similarity score vector is then subjected to exponential normalization, i.e., a softmax operation, to obtain the attention weight vector corresponding to the first query vector. The specific calculation formula is as follows: ,in The summation index variable. After softmax transformation, the attention weight vector... All elements in the sequence are positive and sum to 1, representing the distribution of attention given to each position in the sequence by the first query vector.

[0074] To enhance the model's ability to perceive temporal location information, a temporal location encoding bias term is introduced. This term is applied to positions in the sequence with index 0. From the feature vector, generate the corresponding position encoding vector. The positional encoding vector can be generated using sine and cosine functions. For the positional encoding vector... Each dimension, when Calculate when it is even ,when Calculate when it is an odd number Adding the position-encoded vector to the key vector results in an enhanced key vector. When calculating the similarity score, the enhanced key vector is used instead of the original key vector; that is, the calculation... In this way, temporally adjacent feature vectors will receive higher similarity scores due to the similarity of their positional encodings, thus occupying a larger proportion in the attention weight distribution. The introduction of positional encoding enables the model to distinguish feature vectors at different positions in the sequence, even if their content features are similar, and can be differentiated based on their temporal positional relationship.

[0075] Repeat the above calculation process to process all query vectors in the query vector sequence, for the first... query vectors The similarity score vector is obtained by calculating its inner product with all enhancement key vectors. After scaling and softmax transformation, the corresponding attention weight vector is obtained. After traversing all After obtaining a query vector, the result will be The attention weight vectors are stacked row-wise to form a temporal attention matrix. The first of the matrix Line 1 Column elements Indicates the first The query vector pairs the first The attention weights of each key vector reflect their position in the sequence. Position The level of attention given to it.

[0076] The temporal attention matrix has a clear physical meaning. Each row of the matrix represents the attention distribution of a time step across the entire sequence, and the sum of the elements within a row is 1. Diagonal elements typically have larger values, indicating higher attention from each time step. Due to the introduction of positional encoding bias, elements near the diagonal also exhibit larger values, reflecting the influence of temporal proximity. By analyzing the numerical distribution pattern of the attention matrix, key time periods and anomalous moments in the temperature curve can be identified. For example, if the attention weights in a certain row are concentrated on a few time steps, it indicates that the temperature characteristics at that moment mainly depend on information from these specific time steps, suggesting abrupt changes or anomalous patterns.

[0077] In practical implementations, to improve computational efficiency, matrix operations are typically used to process all query vectors in batches. The query vector sequence, key vector sequence, and value vector sequence are each organized into matrix form. , and Through matrix multiplication Calculate the similarity scores between all query vectors and all key vectors in one step to obtain the similarity score matrix. The temporal attention matrix is ​​directly obtained after scaling and applying a softmax operation to the score matrix. This batch computation method can fully utilize the parallel computing power of the GPU, significantly improving the computation speed.

[0078] After the temporal attention matrix is ​​calculated, it is multiplied by the value vector sequence to obtain the weighted feature representation. Specifically, for the Each query vector corresponds to an output feature vector. The output feature vector incorporates information from all positions in the sequence, with the fusion weights determined by the attention weights. Combining all output feature vectors yields a feature sequence enhanced by temporal self-attention. This sequence better captures the global temporal dependencies of the temperature curve, providing a more discriminative feature representation for subsequent anomaly detection.

[0079] A second aspect of the present invention provides a system for detecting abnormal temperature curves in engineering machinery coolers, comprising: The data encoding unit is used to acquire historical temperature time-series data of the source domain cooler and the corresponding operating condition label, as well as a small amount of labeled temperature time-series data of the target domain cooler, and to perform multi-layer time-series feature encoding on the historical temperature time-series data of the source domain cooler to obtain the source domain temperature pattern feature representation. The migration update unit is used to migrate and update the feature encoding weights by minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, so as to obtain a domain-invariant feature representation. The prediction generation unit is used to encode a small amount of labeled temperature time series data of the target domain cooler using the domain-invariant feature representation to obtain a target domain feature vector, calculate the time series weight distribution based on the time series dependency relationship in the target domain feature vector, and generate a temperature pattern prediction result by combining the cooler operating status parameters at the current moment. The reverse optimization unit is used to perform joint reverse optimization of the feature encoding weights and temporal weight distributions based on the error between the temperature pattern prediction results and the actual temperature pattern labels.

[0080] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0081] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0082] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting abnormal temperature curves in engineering machinery coolers, characterized in that, include: The historical temperature time series data and corresponding operating condition labels of the source domain cooler are obtained, as well as a small amount of labeled temperature time series data of the target domain cooler. Multi-layer time series feature encoding is performed on the historical temperature time series data of the source domain cooler to obtain the source domain temperature pattern feature representation. By minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, the feature encoding weights are migrated and updated to obtain a domain-invariant feature representation. The domain-invariant feature representation is used to encode a small amount of labeled temperature time series data of the target domain cooler to obtain a target domain feature vector. The time series weight distribution is calculated based on the time series dependency relationship in the target domain feature vector, and the temperature pattern prediction result is generated by combining the cooler operating status parameters at the current moment. Based on the error between the temperature pattern prediction result and the actual temperature pattern label, the feature encoding weight and temporal weight distribution are jointly optimized in reverse.

2. The method according to claim 1, characterized in that, Multi-layer time-series feature encoding is performed on the historical temperature time-series data of the source domain cooler to obtain the source domain temperature pattern feature representation, including: The historical temperature time series data of the source domain cooler is divided into multiple time series segments according to the time window. The first time series convolution operation is performed on each time series segment. The local change pattern between temperature values ​​at adjacent time moments is captured by the sliding convolution kernel to obtain the first layer feature mapping. Based on the first layer feature mapping, a multi-scale temporal receptive field is constructed. By parallel convolution operations with different convolution strides, short-term temperature fluctuation features and long-term temperature trend features are captured respectively, resulting in a multi-scale feature representation. The multi-scale feature representation is recalibrated at the channel level, and the channel attention coefficient is calculated based on the contribution of each scale feature to temperature pattern recognition. The channel features are then adaptively weighted. The weighted multi-scale features are spliced ​​and fused, and the fused features are projected onto a feature space of a unified dimension through a fully connected mapping to obtain the source domain temperature pattern feature representation.

3. The method according to claim 1, characterized in that, By minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, the feature encoding weights are updated through migration, resulting in a domain-invariant feature representation including: A small amount of labeled temperature time-series data of the target domain cooler is encoded using the same multi-layer time-series feature encoding process as the source domain to obtain the temperature pattern feature representation of the target domain. A domain discriminator network is constructed, and the source domain temperature pattern feature representation and the target domain temperature pattern feature representation are respectively input into the domain discriminator network. The domain belonging label of the predicted feature sample is output through binary classification. The domain classification loss of the domain discriminator network is calculated. The gradient of the domain classification loss is inverted by a gradient inversion layer, and the inverted gradient is backpropagated to the weight parameters of the feature encoder, so that the optimization direction of the feature encoder is against the optimization direction of the domain discriminator. During the adversarial optimization process, the source domain temperature pattern classification loss is minimized and the domain classification confusion is maximized simultaneously. When the domain discriminator network cannot effectively distinguish between the features of the source domain and the target domain, the iterative optimization is terminated. Extract the feature encoder weight parameters after the adversarial optimization convergence, and use the feature encoder weight parameters to perform feature encoding on the temperature time series data of the source domain and the target domain to obtain the neighborhood-invariant feature representation.

4. The method according to claim 3, characterized in that, A domain discriminator network is constructed, and the source domain temperature pattern feature representation and the target domain temperature pattern feature representation are respectively input into the domain discriminator network. The predicted domain attribution labels of the feature samples are output through binary classification, including: A domain discriminator network containing multiple fully connected layers is constructed. The source domain temperature pattern feature representation and the target domain temperature pattern feature representation are merged to form a hybrid feature sample set. Each feature sample is labeled with a domain source identifier. The source domain feature sample is labeled as the first domain identifier, and the target domain feature sample is labeled as the second domain identifier. The feature samples in the hybrid feature sample set are sequentially input into the first fully connected layer of the domain discriminator network, and the feature vectors are linearly mapped and nonlinearly transformed by nonlinear activation units to obtain the first layer of hidden feature representation. The first hidden feature representation is input into the second fully connected layer for further feature space mapping. By compressing the feature dimension layer by layer, the discriminative feature representation is obtained. The discriminative feature is represented as the input-output fully connected layer. The value of the output neuron is converted into the domain belonging probability through the activation function. The domain identifier corresponding to the output neuron with the highest domain belonging probability is selected as the predicted domain belonging label.

5. The method according to claim 1, characterized in that, The target domain feature vector is obtained by using the domain-invariant feature representation to encode a small amount of labeled temperature time-series data of the cooler in the target domain. The temporal weight distribution is calculated based on the temporal dependencies in the target domain feature vector, and a temperature pattern prediction result is generated by combining the cooler's operating state parameters at the current moment. Input a small amount of labeled temperature time-series data of the target domain cooler into the feature encoder that has completed the migration update to obtain the target domain feature vector sequence; A temporal self-attention computation unit is constructed to calculate the similarity score between each feature vector in the target domain feature vector sequence and all feature vectors in the sequence, thereby obtaining the temporal attention matrix; The temporal attention matrix is ​​normalized row by row to obtain a temporal weight distribution. The target domain feature vector sequence is weighted and aggregated according to the temporal weight distribution. The aggregated feature vector is then concatenated with the cooler operating state parameters at the current moment to obtain a fused feature vector. The fused feature vector is input into the temperature pattern classifier, and mapped to the temperature pattern category probability distribution through a fully connected layer and activation function. The category with the highest probability is selected as the temperature pattern prediction result.

6. The method according to claim 5, characterized in that, A temporal self-attention computation unit is constructed to calculate the similarity score between each feature vector in the target domain feature vector sequence and all feature vectors in the sequence, resulting in a temporal attention matrix including: The target domain feature vector sequence is linearly transformed by the query mapping matrix, key mapping matrix and value mapping matrix respectively to obtain the query vector sequence, key vector sequence and value vector sequence; For the first query vector in the query vector sequence, calculate the inner product of the first query vector and all key vectors in the key vector sequence to obtain the similarity score vector corresponding to the first query vector; The similarity score vector is scaled by dividing each score value by the square root of the key vector dimension to avoid the inner product value being too large and causing the gradient to vanish. The scaled similarity score vector is subjected to exponential normalization to obtain the attention weight vector corresponding to the first query vector. A temporal position encoding bias term is introduced, and a position encoding vector is generated based on the temporal position index of the feature vector in the sequence. The position encoding vector is then added to the calculation process of the similarity score vector, so that feature vectors that are temporally adjacent obtain higher similarity scores. Repeat the above calculation process to traverse all query vectors in the query vector sequence, and stack the attention weight vectors corresponding to all query vectors row by row to obtain the temporal attention matrix.

7. An abnormal temperature curve detection system for engineering machinery coolers, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data encoding unit is used to acquire historical temperature time-series data of the source domain cooler and the corresponding operating condition label, as well as a small amount of labeled temperature time-series data of the target domain cooler, and to perform multi-layer time-series feature encoding on the historical temperature time-series data of the source domain cooler to obtain the source domain temperature pattern feature representation. The migration update unit is used to migrate and update the feature encoding weights by minimizing the distribution difference metric between the source domain temperature pattern feature representation and the target domain temperature pattern feature representation in the common feature space, so as to obtain a domain-invariant feature representation. The prediction generation unit is used to encode a small amount of labeled temperature time series data of the target domain cooler using the domain-invariant feature representation to obtain a target domain feature vector, calculate the time series weight distribution based on the time series dependency relationship in the target domain feature vector, and generate a temperature pattern prediction result by combining the cooler operating status parameters at the current moment. The reverse optimization unit is used to perform joint reverse optimization of the feature encoding weights and temporal weight distributions based on the error between the temperature pattern prediction results and the actual temperature pattern labels.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images