CIR equipment management method and device, equipment and storage medium

Abstract

The invention relates to the technical field of equipment management, and discloses a CIR equipment management method and device, equipment and a storage medium, and the method comprises the steps: carrying out the correction and normalization processing of the historical temperature data of CIR equipment, and obtaining a historical panoramic temperature feature; the method comprises the following steps of: training a Siamese-MobileNet network to obtain a trained network model; performing similarity calculation based on the trained network model to obtain a fault diagnosis result; constructing a dual deep Q network based on a fault diagnosis result for training to obtain a dynamic scheduling optimization model; performing parameter online updating on the dynamic scheduling optimization model to obtain an updated optimization model; on the basis of the updated optimization model, multi-target optimization calculation is carried out in combination with a probability penalty moment boundary intersection algorithm, a scheduling strategy of the CIR equipment is obtained, historical data and real-time information can be comprehensively utilized, early diagnosis and prediction of faults can be achieved, and multi-target optimization scheduling is carried out on the basis.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment management, and in particular to a CIR equipment management method, device, equipment and storage medium. Background Art

[0002] As the complexity of CIR equipment continues to increase, its management and maintenance are facing unprecedented challenges. Traditional equipment management methods often rely on manual experience and fixed maintenance plans, which are difficult to adapt to the dynamic changes and potential failures of CIR equipment. Especially in terms of early signs of failure such as abnormal temperature, the existing management methods lack the ability to identify them in a timely and accurate manner, resulting in a high equipment failure rate, affecting production efficiency and product quality.

[0003] In addition, the operation of CIR equipment involves the balance of multiple objectives, such as production efficiency, energy consumption, and equipment life, which are often contradictory and difficult to optimize simultaneously. Traditional single-objective optimization methods are difficult to meet actual production needs, while multi-objective optimization methods often face problems such as high computational complexity and slow convergence. Especially in a dynamic production environment, how to achieve real-time optimization and scheduling of CIR equipment to adapt to the ever-changing production needs and equipment status has become a difficult problem that needs to be solved urgently. Summary of the invention

[0004] The present invention provides a CIR equipment management method, device, equipment and storage medium for comprehensively utilizing historical data and real-time information to achieve early diagnosis and prediction of faults, and perform multi-objective optimization scheduling on this basis.

[0005] In a first aspect, the present invention provides a CIR device management method, the CIR device management method comprising: Correct and normalize the historical temperature data of the CIR device to obtain the historical panoramic temperature characteristics; Inputting the historical panoramic temperature features into a preset Siamese-MobileNet network for training to obtain a trained network model; Based on the trained network model, similarity calculation is performed on the real-time panoramic temperature features of the CIR device to obtain a fault diagnosis result; Building a dual deep Q network based on the fault diagnosis result, and training the dual deep Q network to obtain a dynamic scheduling optimization model; Based on the recursive least squares method, the parameters of the dynamic scheduling optimization model are updated online to obtain an updated optimization model; Based on the updated optimization model, a multi-objective optimization calculation is performed in combination with a probability penalty moment boundary intersection algorithm to obtain a scheduling strategy for the CIR device.

[0006] In a second aspect, the present invention provides a CIR device management device, the CIR device management device comprising: The correction module is used to correct and normalize the historical temperature data of the CIR device to obtain the historical panoramic temperature characteristics; A training module, used for inputting the historical panoramic temperature features into a preset Siamese-MobileNet network for training to obtain a trained network model; A calculation module, used for performing similarity calculation on the real-time panoramic temperature features of the CIR device based on the trained network model to obtain a fault diagnosis result; A construction module, used for constructing a dual deep Q network based on the fault diagnosis result, training the dual deep Q network, and obtaining a dynamic scheduling optimization model; An updating module, used for performing online parameter updating of the dynamic scheduling optimization model based on a recursive least squares method to obtain an updated optimization model; The optimization module is used to perform multi-objective optimization calculation based on the updated optimization model in combination with the probability penalty moment boundary intersection algorithm to obtain the scheduling strategy of the CIR device.

[0007] The third aspect of the present invention provides a CIR device management device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the CIR device management device executes the above-mentioned CIR device management method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned CIR device management method.

[0009] In the technical solution provided by the present invention, by correcting and normalizing the historical temperature data of the CIR device and combining the training of the Siamese-MobileNet network, the temperature characteristics of the device can be captured more accurately, thereby improving the accuracy of fault diagnosis. Based on the trained network model, the similarity calculation of the real-time panoramic temperature characteristics can be performed to quickly identify the potential fault type and realize the early warning of the fault. By constructing a dual deep Q network and combining the recursive least squares method to update the parameters online, the scheduling strategy can be dynamically adjusted according to the real-time status of the equipment and production needs to improve the adaptability of the system. The probability penalty moment boundary intersection algorithm is used for multi-objective optimization calculation, which can simultaneously consider multiple objectives such as production efficiency, energy consumption and equipment life to achieve more comprehensive optimization. By using lightweight network structures such as MobileNetV3 and combining the fast convergence characteristics of the recursive least squares method, this method greatly reduces the computational complexity while ensuring the optimization effect, and is suitable for online real-time applications. Through the application of Siamese network structure and contrastive learning, small sample problems can be better handled and the generalization ability of the model in the face of new faults can be improved. Through timely fault diagnosis and optimized scheduling strategies, the unplanned downtime of CIR equipment can be effectively reduced, and the overall reliability and production efficiency can be improved.

[0010] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0011] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A schematic diagram of an embodiment of a CIR device management method in an embodiment of the present invention; Figure 2 A schematic diagram of an embodiment of a CIR device management device in an embodiment of the present invention; Figure 3 The figure is a schematic diagram of an embodiment of a CIR device management device in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0015] To facilitate understanding of this embodiment, a CIR device management method disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps: 101. Correct and normalize the historical temperature data of the CIR device to obtain historical panoramic temperature characteristics; It is understandable that the execution subject of the present invention may be a CIR device management device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0016] Specifically, temperature sensors are installed at multiple target locations of the CIR device, and historical temperature data of each location is collected through the sensors. The historical temperature data is curve fitted to generate the corresponding original temperature curve. The original temperature curve is compensated for the ambient temperature to eliminate the interference of the temperature data caused by the change of the external environment, and the compensated temperature curve is obtained. The compensated temperature curve is corrected for sensor sensitivity to eliminate the error caused by the inconsistent sensor sensitivity, and the corrected temperature curve is obtained. The corrected temperature curve is normalized to the maximum and minimum values, and each temperature data is mapped to the same range to form a normalized temperature curve. The normalization process can eliminate the deviation caused by the different ranges of data collected by different sensors, so that all data can be compared and analyzed under the same standard. The normalized temperature curve is arranged in chronological order to construct a one-dimensional vector to represent the panoramic temperature feature in a single time series. The panoramic temperature feature of a single time series is processed by sliding window, and the entire time series is divided into multiple subsequences of fixed length, and local dynamic features are extracted in the time dimension. Fourier transform is performed on each subsequence to obtain the corresponding frequency domain panoramic temperature feature. Fourier transform can convert time domain features into frequency domain features, allowing the system to capture the periodicity and frequency information in temperature data. The frequency domain panoramic temperature features are spliced ​​with the panoramic temperature features of a single time series, and the time domain and frequency domain features are combined to construct a more comprehensive historical panoramic temperature feature.

[0017] 102. Input the historical panoramic temperature features into the preset Siamese-MobileNet network for training to obtain a trained network model; Specifically, the historical panoramic temperature features are randomly paired to construct a training sample set of positive sample pairs and negative sample pairs. A positive sample pair represents two samples from the same category, while a negative sample pair represents two samples from different categories. The generation method of sample pairs helps the network learn the similarities and differences between different samples during the training process. A Siamese-MobileNet network structure is constructed, which consists of two MobileNetV3 subnetworks with shared parameters and a distance metric layer. The two MobileNetV3 subnetworks are fully parameter-shared, that is, during the training process, they share the same weight parameters, thereby ensuring that for the same pair of inputs, the two parts of the network can output consistent feature vector representations. When configuring the MobileNetV3 subnetwork, each subnetwork contains an initial convolution layer, five inverted residual blocks, and a global average pooling layer. The initial convolution layer uses a 3x3 convolution kernel, the stride is set to 2, and the number of output channels is 16, which helps to perform a relatively rough feature extraction of the input panoramic temperature features in the initial stage. The configuration of the inverted residual block is more complex. Each block applies 1x1 convolution, depthwise separable convolution, and 1x1 convolution in sequence. After each convolution operation, a batch normalization layer is added to stabilize the training process, and the ReLU6 activation function is used to introduce nonlinear features. The use of depthwise separable convolution reduces the number of network parameters, allowing MobileNetV3 to improve computational efficiency while maintaining high accuracy. The distance metric layer of the Siamese-MobileNet network is set to use cosine similarity as the distance metric. Cosine similarity is a commonly used similarity metric that measures the similarity of two feature vectors by calculating the cosine value of the angle between them. The closer the value is to 1, the more similar the two feature vectors are. Through this metric, the network can determine the similarity between the two input samples. The positive and negative sample pairs in the training sample set are input into the two MobileNetV3 subnetworks respectively to obtain two feature vectors. The two feature vectors are calculated by the cosine similarity of the distance metric layer to obtain the similarity score of the sample pair. Based on the similarity score and the true label of the sample pair (i.e., whether they belong to the same category), the contrast loss function value of the network is calculated. The contrast loss function can measure the performance of the network in distinguishing positive and negative sample pairs. The goal of the network is to minimize this loss value, thereby improving the ability to identify similar samples and distinguish different samples. To optimize the network parameters, the Adam optimizer is used for backpropagation and gradient update. The Adam optimizer is widely used in the training of deep learning models due to its good convergence performance. After each training cycle, the network will be evaluated on the validation set. Common performance indicators include accuracy, precision, recall, etc. When the performance indicators on the validation set no longer improve, the training process stops, and the trained Siamese-MobileNet network model is finally obtained.

[0018] 103. Based on the trained network model, similarity calculation is performed on the real-time panoramic temperature features of the CIR device to obtain the fault diagnosis result; Specifically, real-time temperature data of CIR devices are collected. The real-time temperature data is preprocessed, including data smoothing, noise filtering and other steps, to ensure the accuracy and consistency of the data and obtain high-quality real-time panoramic temperature features. The real-time panoramic temperature features are input into the first MobileNetV3 subnetwork of the trained Siamese-MobileNet network model. The subnetwork processes the input features in a predetermined structural order, and sequentially processes the initial convolution layer, five inverted residual blocks and the global average pooling layer to generate a real-time feature vector representing the real-time temperature data. Multiple fault temperature feature samples of known types are selected from the pre-established fault sample library. Each fault sample is sequentially input into the second MobileNetV3 subnetwork of the Siamese-MobileNet network, and the fault feature vectors corresponding to these fault samples are obtained through the same convolution processing flow. Since the two subnetworks of the Siamese network share parameters, the processing process of the fault sample is symmetrical with that of the real-time data, ensuring the consistency and comparability of the feature vectors. The cosine similarity of the distance metric layer is calculated for the real-time feature vector and each fault feature vector. Using cosine similarity as the distance metric, a set of similarity scores is obtained by calculating the cosine value of the angle between the real-time feature vector and each fault feature vector. Each score reflects the similarity between the real-time data and a known fault type. The higher the score, the more similar the two are. Based on multiple similarity scores, a fault probability distribution matrix is ​​constructed, and each element of the matrix corresponds to the similarity between the real-time feature and a fault feature. In order to convert the similarity score into a more explanatory probability value, the matrix is ​​operated by the softmax function. The softmax function can normalize the similarity score to a probability value, ensuring that the sum of the probabilities of all fault types is 1, and a normalized fault probability distribution is obtained, indicating the probability that the real-time feature corresponds to different fault types. A fault diagnosis threshold is set to filter the normalized fault probability distribution. By comparing with the threshold, the fault types with probability values ​​greater than the preset threshold are filtered out, thereby narrowing the possible fault range. The filtered fault types can be classified by multiple labels, that is, a CIR device may correspond to multiple faults at the same time. These fault types are presented in the form of multiple labels, and finally the fault diagnosis results are generated.

[0019] 104. Construct a dual deep Q network based on the fault diagnosis result, train the dual deep Q network, and obtain a dynamic scheduling optimization model; Specifically, the operating status, fault diagnosis results, and production task information of the CIR equipment are encoded. By encoding this information into a numerical state vector, the actual state of the equipment is converted into input data that can be processed by the machine learning model. The state vector is input into the dual deep Q network, which consists of a main network and a target network. The two networks share a feature extraction layer but have independent value function estimation layers and advantage function estimation layers. When configuring the dual deep Q network, its feature extraction layer is designed. The feature extraction layer contains three fully connected layers, each of which uses the ReLU activation function to introduce nonlinear characteristics. The output dimension of the feature extraction layer is gradually reduced to 256, 128, and 64, respectively. The gradually reduced structure helps to extract more refined feature vectors. In the value function estimation layer and the advantage function estimation layer, each layer is configured with two fully connected layers, and the ReLU activation function is also used. In the last layer of these two estimation layers, the output dimension is equal to the size of the action space, so as to ensure that the network can generate corresponding action decisions according to different states. The encoded state vector is input into the feature extraction layer of the main network to obtain a feature vector. This feature vector is simultaneously input to the value function estimation layer and the advantage function estimation layer, and the Q value of each possible action is obtained through mathematical operations. The Q value is used to measure the expected return of performing an action in a specific state. The network guides the dynamic scheduling of CIR devices by selecting the action with the highest Q value. To select actions, the ε-greedy strategy is adopted, which combines the balance of exploration and exploitation. The network has a certain probability of selecting a random action (exploration), and the rest of the time it selects the action with the largest current Q value (exploitation). The selected action will be applied to the CIR device, and the next state of the device and the corresponding reward value will be observed and recorded. The state transition samples (including the current state, action, reward, and next state) are stored in the experience replay buffer for subsequent training. A batch of state transition samples are randomly sampled from the experience replay buffer, and the network parameters are updated by calculating the target Q value and current Q value of these samples. The target Q value is calculated by the target network, while the current Q value is calculated by the main network. In order to evaluate the network performance, the time difference error is calculated based on the difference between the target Q value and the current Q value, and the mean square error is used as the loss function to measure the gap between the model's prediction and the actual return. The parameters of the main network are then optimized using the gradient descent method, updated every 4 time steps. In order to maintain the coordinated development of the main network and the target network, the parameters of the main network are copied to the target network every 100 time steps to achieve soft updates of the target network. This ensures that the target network does not change too frequently and ensures the stability of training. By continuously updating and optimizing the network, the network will eventually converge to obtain an optimization model that can effectively deal with the dynamic scheduling problem of CIR devices.In actual operation, this model can intelligently select the best scheduling strategy based on the equipment status, fault conditions and task requirements, thereby improving the operating efficiency and safety of the equipment.

[0020] 105. Based on the recursive least squares method, the parameters of the dynamic scheduling optimization model are updated online to obtain an updated optimization model; Specifically, the parameters of the dynamic scheduling optimization model are initialized. During initialization, the forgetting factor λ and the initial covariance matrix P are set. The value range of the forgetting factor λ is usually between (0, 1]. Its function is to gradually reduce the influence of old data over time, so that the model can adapt to new changes more quickly. The initial covariance matrix P is set to a large positive number multiplied by the unit matrix to ensure that the adjustment of model parameters in the initial stage has greater flexibility. Based on the real-time status of the CIR device and the currently executed action, the model calculates the Q value at the current moment. The Q value represents the expected return of performing an action in the current state. The calculation of the actual Q value depends on the current device status, action and related feedback data, reflecting the effectiveness of the current strategy. At the same time, based on the dynamic scheduling optimization model, the system predicts the next The maximum Q value that may appear at the moment. By combining the current reward value r, the target Q value is calculated, where the target Q value is equal to the reward r plus the discount factor γ multiplied by the maximum Q value predicted at the next moment. The discount factor γ is usually used to measure the importance of future returns and takes a value between 0 and 1. The actual Q value is compared with the calculated target Q value, and the difference between the two is the time difference error δ. δ is equal to the target Q value minus the actual Q value. The size of δ reflects the deviation between the current model's estimate of future returns and the actual situation. This error value is the key basis for subsequent updating of model parameters. Based on the current state and action of the device, the current feature vector x is extracted from the model. Transpose the eigenvector x and multiply it with the covariance matrix P, then perform the inner product operation with the current eigenvector x to obtain an intermediate variable z. The role of the intermediate variable z is to measure the correlation between the current eigenvector and the covariance matrix. Using the forgetting factor λ and the intermediate variable z, calculate the gain vector k. The gain vector k is the core part of the recursive least squares method, which is equal to the product of the covariance matrix P and the eigenvector x, divided by (λ plus z). By introducing the gain vector, the model can flexibly adjust its parameters according to the current error. The size of the gain vector directly affects the amplitude of the parameter update, ensuring that the model can effectively learn new information in each iteration. Based on the gain The update amount of θ is equal to the gain vector k multiplied by the time difference error δ. This updating process makes the parameters of the model gradually approach the optimal solution to better adapt to the operating status and scheduling requirements of the CIR equipment. At the same time, the covariance matrix P is updated to ensure that the model can continue to effectively adjust the parameters in the next iteration. By recursively updating the covariance matrix and parameters θ, the system can optimize its own performance under continuous real-time data input. This process will continue until the change of parameter θ converges, or the preset number of iterations is reached, and finally the updated dynamic scheduling optimization model is obtained.

[0021] 106. Based on the updated optimization model, the multi-objective optimization calculation is performed in combination with the probability penalty moment boundary intersection algorithm to obtain the scheduling strategy of the CIR equipment.

[0022] Specifically, the key performance indicators of CIR equipment are modeled. Production efficiency, energy consumption and equipment life are the three main objectives that affect the operation of CIR equipment. Therefore, by establishing mathematical models for these three, a set of objective functions for multi-objective optimization problems is constructed. The objective function reflects the trade-off relationship between key indicators such as equipment operation efficiency, energy consumption and service life under different scheduling strategies. Based on the updated optimization model, N initial solutions are generated, each of which represents a candidate scheduling strategy for CIR equipment. These initial solutions are diverse and can cover different possible states of equipment operation. After generating these solutions, they are non-dominated sorted. Through non-dominated sorting, solutions that perform balanced performance on various objectives are found to form a Pareto front solution set. The solutions in the Pareto front solution set cannot be optimized by improving one objective without compromising another objective, so they are candidate optimal solutions in multi-objective optimization. Two solutions are randomly selected from the Pareto front solution set, denoted as a1 and a2. The simulated binary crossover operator is used to perform a crossover operation on these two solutions, and two new offspring solutions y1 and y2 are generated from their gene combinations. The purpose of the crossover operation is to generate offspring solutions with better performance by combining the characteristics of the parent solution, thereby improving the diversity of solutions and global search capabilities. When generating offspring solutions y1 and y2, polynomial mutation operations are performed on the two offspring solutions. Through mutation, some genes in the offspring solutions are locally randomly adjusted to increase the diversity of solutions and avoid the trap of local optimality. After the mutation operation, mutated solutions z1 and z2 are generated, which may perform better than the parent on some objectives. The mutated solutions z1 and z2 are merged with the original N initial solutions to form a new solution set. For each solution in the new solution set, its probability penalty boundary intersection value is calculated. The probability penalty boundary intersection algorithm is an effective multi-objective optimization algorithm that measures the quality of each solution by considering the performance of the solution on different objectives and combining it with a penalty function. Based on the probability penalty boundary intersection value of each solution, all solutions in the solution set are sorted, and the N optimal solutions are retained as the next generation population. This process will be iterated continuously, and each generation will generate a new solution set through crossover, mutation and selection, and select the best solution through the probability penalty boundary intersection algorithm. This iterative process continues until the preset number of iterations is reached or the convergence condition is met. The system will output an optimized CIR equipment scheduling strategy that can find the best balance between production efficiency, energy consumption and equipment life to achieve efficient operation of the equipment.

[0023] In the embodiment of the present invention, by correcting and normalizing the historical temperature data of the CIR device and combining the training of the Siamese-MobileNet network, the temperature characteristics of the device can be captured more accurately, thereby improving the accuracy of fault diagnosis. Based on the trained network model, the similarity calculation of the real-time panoramic temperature characteristics can be performed to quickly identify the potential fault type and realize the early warning of the fault. By constructing a dual deep Q network and combining the recursive least squares method to update the parameters online, the scheduling strategy can be dynamically adjusted according to the real-time status of the equipment and production demand, thereby improving the adaptability of the system. The probability penalty moment boundary intersection algorithm is used for multi-objective optimization calculation, which can simultaneously consider multiple objectives such as production efficiency, energy consumption and equipment life, and achieve more comprehensive optimization. By using lightweight network structures such as MobileNetV3 and combining the fast convergence characteristics of the recursive least squares method, this method greatly reduces the computational complexity while ensuring the optimization effect, and is suitable for online real-time applications. Through the application of Siamese network structure and contrastive learning, small sample problems can be better handled, and the generalization ability of the model in the face of new faults can be improved. Through timely fault diagnosis and optimized scheduling strategies, the unplanned downtime of CIR equipment can be effectively reduced, and the overall reliability and production efficiency can be improved.

[0024] In a specific embodiment, the process of executing step 101 may specifically include the following steps: Temperature sensors are installed at multiple target locations of the CIR device to collect historical temperature data, and curve fitting is performed on the historical temperature data to obtain the original temperature curve; Performing ambient temperature compensation on the original temperature curve to obtain a compensated temperature curve, and performing sensor sensitivity correction on the compensated temperature curve to obtain a corrected temperature curve; The corrected temperature curve is normalized to its maximum and minimum values ​​to obtain a normalized temperature curve, and the normalized temperature curve is arranged in chronological order to form a one-dimensional vector to obtain the panoramic temperature characteristics of a single time series; Perform sliding window processing on the panoramic temperature features of a single time series to obtain multiple subsequences of fixed length, and perform Fourier transform on the multiple subsequences of fixed length to obtain frequency domain panoramic temperature features; The frequency domain panoramic temperature features are concatenated with the panoramic temperature features of a single time series to obtain the historical panoramic temperature features.

[0025] Specifically, temperature sensors are installed at key locations of the CIR device. These sensors can be distributed in different areas of the device to monitor temperature changes in various locations. Each sensor continuously collects temperature data during the operation of the device to form a set of historical temperature data. Curve fitting technology is applied to the historical temperature data to smooth the data and remove noise. Curve fitting fits the original temperature data to a smooth curve by constructing a mathematical model, which more clearly reflects the actual temperature change trend. Common curve fitting methods include polynomial fitting or least squares fitting. For example, assuming that the relationship between temperature data and time can be expressed by a polynomial, the fitted temperature curve can be expressed as: ; in, Indicates time The temperature of the moment, is the fitting coefficient of the polynomial. Through fitting, the original temperature curve obtained can better describe the temperature change of the device during operation. The original temperature curve is compensated for the ambient temperature. The compensation method can use linear regression or other methods to eliminate the influence of the ambient temperature based on the data of the ambient temperature sensor around the device. Assuming that there is a certain linear relationship between the ambient temperature and the surface temperature of the device, the compensated temperature can be expressed as: ; in, is the temperature after compensation, is the original temperature, is the ambient temperature, is the compensation coefficient, which reflects the influence ratio of ambient temperature on device temperature. The compensated temperature curve may still be limited by the characteristics of the sensor itself, such as the sensitivity error of the sensor. Perform sensor sensitivity correction on the compensated temperature curve to reduce the error caused by the difference in sensitivity of different sensors. Assume that the sensor sensitivity is , the corrected temperature can be expressed as: ; in, is the corrected temperature, is the sensitivity factor of the sensor. After sensitivity correction, the data collected by different sensors are ensured to be consistent. The corrected temperature curve is normalized to the maximum and minimum values, and the data is mapped to a fixed interval, usually [0, 1], to eliminate the influence caused by the different data ranges of different sensors. The normalized temperature curve is arranged in chronological order to form a one-dimensional vector, and the panoramic temperature feature of a single time series is obtained. The time series contains the temperature changes of the equipment at different time points and can reflect the operating status of the equipment. In order to extract the time series features in the data, the panoramic temperature features of a single time series are processed by sliding window. The sliding window processing divides the time series into multiple subsequences by setting a fixed-length window. These subsequences can capture the local features in the time series and help to better analyze the short-term temperature changes during the operation of the equipment. Fourier transform is performed on multiple fixed-length subsequences to convert the signal in the time domain into the signal in the frequency domain, revealing the periodic features and frequency components in the data, and obtaining the frequency domain panoramic temperature features, which reflect the frequency distribution of the equipment temperature data. By analyzing these frequency domain features, it is possible to identify whether the equipment has periodic faults or fluctuations during operation. The frequency domain panoramic temperature features obtained by Fourier transform are spliced ​​with the panoramic temperature features in the time series to obtain the final historical panoramic temperature features. By combining the time domain and frequency domain features, the system can analyze the operating status of the equipment more comprehensively. For example, the time series features can reflect the trend of temperature changes, while the frequency domain features can reveal the periodicity and suddenness of temperature changes.

[0026] In a specific embodiment, the process of executing step 102 may specifically include the following steps: Randomly pair the historical panoramic temperature features to obtain a training sample set containing positive sample pairs and negative sample pairs; Construct the Siamese-MobileNet network structure. The Siamese-MobileNet network includes two parameter-sharing MobileNetV3 sub-networks and a distance measurement layer. Configure the MobileNetV3 subnetworks. Each MobileNetV3 subnetwork contains an initial convolution layer, five inverted residual blocks, and a global average pooling layer. The initial convolution layer uses a 3x3 convolution kernel with a stride of 2 and 16 output channels. The inverted residual blocks use 1x1 convolution, depthwise separable convolution, and 1x1 convolution in sequence, and add a batch normalization layer and ReLU6 activation function after each convolution operation. Set the distance metric layer, use cosine similarity as the distance metric, and calculate the similarity of the output feature vectors of the two MobileNetV3 subnetworks; Input the sample pairs in the training sample set into two MobileNetV3 sub-networks respectively to obtain two training feature vectors; Calculate the cosine similarity of the two training feature vectors to obtain the similarity score of the sample pair, and calculate the contrast loss function value based on the similarity score and the true label of the sample pair; The Adam optimizer is used to optimize the parameters of the Siamese-MobileNet network, and the network performance indicators are calculated on the validation set. When the performance indicators on the validation set no longer improve, the training process is stopped to obtain the trained network model.

[0027] Specifically, historical panoramic temperature features are randomly paired to obtain a training sample set containing positive sample pairs and negative sample pairs. Positive sample pairs are composed of samples of the same category, that is, the device states represented by the two panoramic temperature features are similar or the same; negative sample pairs are samples of different categories, that is, panoramic temperature features representing different device states or fault types. By randomly pairing temperature features, a training sample set is constructed. The goal of positive sample pairs is to allow the model to recognize similar temperature features, while negative sample pairs help the model distinguish different device states. Construct a Siamese-MobileNet network structure. The core of the network is two MobileNetV3 subnetworks with shared parameters and a distance measurement layer. The Siamese network calculates the similarity between the feature vectors of two input samples by comparing them. The two MobileNetV3 subnetworks share the same weight parameters, which means that they are updated simultaneously during the learning process to ensure that similar feature representations are generated for the same input. The distance measurement layer of the network is used to calculate the similarity of two feature vectors, which is usually measured by cosine similarity. In the specific configuration of the MobileNetV3 subnetwork, each subnetwork contains an initial convolution layer, five inverted residual blocks, and a global average pooling layer. The initial convolution layer uses a 3x3 convolution kernel, a stride of 2, and 16 output channels, which enables the convolution layer to extract preliminary information about the input features. The design of the five inverted residual blocks is more complex, using 1x1 convolution, depthwise separable convolution, and 1x1 convolution in sequence. The advantage of depthwise separable convolution is that it can reduce computational complexity while maintaining high accuracy. After each convolution operation, a batch normalization layer is added to stabilize the training process, and the ReLU6 activation function is used to introduce nonlinear characteristics to enhance the network's expressiveness. The final global average pooling layer maps the two-dimensional features to a lower-dimensional vector space by globally averaging the features. This step helps reduce the complexity of the features and reduces the possibility of overfitting. While building the network, the distance metric layer is set. Cosine similarity, as a distance metric, can calculate the similarity between two feature vectors. The formula for calculating cosine similarity is as follows: ; in, and represent two eigenvectors, is their dot product, and is the modulus of the vector (i.e., the Euclidean length). The value of cosine similarity is between [-1, 1], where 1 means that the two vectors are exactly the same, 0 means that they are orthogonal (completely dissimilar), and -1 means that they are opposite. During the training process, each pair of samples in the training sample set is input into two MobileNetV3 subnetworks of the Siamese-MobileNet network. The two subnetworks output a feature vector respectively, and these two feature vectors represent the high-dimensional feature expression of their respective input samples. Then these two feature vectors are input into the distance metric layer, and the similarity score of the sample pair is calculated by cosine similarity. The higher the similarity score, the more similar the two samples are; the lower the similarity score, the greater the difference between the two samples. Based on the similarity score and the true label of the sample pair, the contrast loss function is calculated. The contrast loss function guides the network to learn the distinguishing ability of features by minimizing the distance of similar sample pairs and maximizing the distance of dissimilar sample pairs. The formula of the contrast loss function can be expressed as: ; in, is the loss value, is the label of the sample pair, is the Euclidean distance between two eigenvectors, is the preset distance threshold. For the positive sample pair (i.e. ), the network hopes to minimize their distance so that As close to 0 as possible. For negative sample pairs (i.e. ), the network hopes that their distance is greater than the threshold , thereby maximizing their discrimination. During the training process, the Adam optimizer is used to optimize the parameters of the Siamese-MobileNet network. The Adam optimizer is a commonly used optimization algorithm that accelerates convergence and improves the stability of training by adaptively adjusting the learning rate of each parameter. The network performs backpropagation according to the feedback of the contrast loss function at each iteration to update the weights of the model. In order to evaluate the performance of the network, performance indicators such as accuracy, precision, and recall are calculated on the validation set. When the performance indicators on the validation set no longer improve, it means that the model has reached the optimal state and the training process can be stopped. The trained Siamese-MobileNet network model is obtained. The trained model is used for real-time fault diagnosis. By inputting the real-time temperature characteristics of the device and comparing the similarity with the historical data, the operating status or fault type of the device can be accurately identified.

[0028] In a specific embodiment, the process of executing step 103 may specifically include the following steps: Collect the real-time temperature data of the CIR device and pre-process the real-time temperature data to obtain the real-time panoramic temperature characteristics; The real-time panoramic temperature features are input into the first MobileNetV3 sub-network of the trained network model, and processed sequentially through the initial convolution layer, five inverted residual blocks, and the global average pooling layer to obtain the real-time feature vector; Select multiple fault temperature feature samples of known types from the pre-established fault sample library, and input each fault sample into the second MobileNetV3 sub-network of the trained network model in turn to obtain multiple fault feature vectors; Perform cosine similarity calculation on the distance metric layer for the real-time feature vector and each fault feature vector to obtain multiple similarity scores; Based on multiple similarity scores, a fault probability distribution matrix is ​​constructed, where each element in the matrix represents the similarity between the real-time feature and the preset fault type; Perform softmax function operation on the fault probability distribution matrix, convert the similarity score into a probability value, and obtain the normalized fault probability distribution; A fault diagnosis threshold is set, and the normalized fault probability distribution is threshold screened to screen out the fault types whose probability values ​​are greater than the fault diagnosis threshold. Multi-label classification is performed on the screened fault types to obtain the fault diagnosis results.

[0029] Specifically, temperature sensors are installed at multiple target parts of the equipment to continuously collect temperature data of different parts during the operation of the equipment. The collected real-time temperature data is preprocessed. The preprocessing process may include data denoising, interpolation, normalization and other operations to ensure that the obtained temperature data can accurately reflect the actual operating status of the equipment and obtain processed real-time panoramic temperature features. The real-time panoramic temperature features are input into a trained network model. In this implementation, the core of the network model is the MobileNetV3 subnetwork. MobileNetV3 has a lightweight structure and is suitable for running on embedded or resource-limited devices. The input of the network passes through an initial convolution layer, which uses a 3x3 convolution kernel, a step size of 2, and an output channel number of 16. The convolution operation can effectively extract the spatial information of the temperature feature. The data passes through five inverted residual blocks, each of which contains a 1x1 convolution, a depth-separable convolution, and a 1x1 convolution. The inverted residual block extracts the deep information of temperature features layer by layer through efficient convolution operations, and normalizes and activates the output of each layer through batch normalization layer and ReLU6 activation function to ensure the stability and nonlinear expression ability of the network. After processing by these layers, the feature map is input into the global average pooling layer to map the multi-dimensional features into a lower-dimensional vector space to obtain the real-time feature vector, which represents the temperature features under the current device state. Multiple known types of fault temperature feature samples are selected from the pre-established fault sample library. Each sample in the fault sample library corresponds to a specific fault type and has been marked and stored by previous data. For each fault sample, it is sequentially input into the second sub-network of the trained MobileNetV3 network, and the structure of this sub-network is exactly the same as the sub-network that processes the real-time features. After sequential processing by the initial convolution layer, the inverted residual block and the global average pooling layer, each fault sample is also converted into a feature vector, namely the fault feature vector. The similarity between the real-time feature vector and each fault feature vector is calculated. The cosine similarity is used to measure the similarity between two vectors by comparing the angle between them. The formula for cosine similarity is: ; in, and Represent the real-time feature vector and the fault feature vector respectively, represents the dot product of two vectors, and Respectively represent the modulus (i.e., length) of the two vectors. The value range of cosine similarity is [-1, 1]. When the value is close to 1, it means that the two vectors are very similar, when the value is close to 0, it means that they are unrelated, and when the value is -1, it means that they are completely opposite. In this way, the similarity between each fault feature vector and the real-time feature vector is calculated to obtain multiple similarity scores. Based on the similarity scores, a fault probability distribution matrix is ​​constructed. Each element of the matrix represents the degree of similarity between the real-time feature and different fault types, that is, the higher the similarity score, the closer the real-time temperature feature is to the fault feature. By performing a softmax operation on the matrix, the similarity score is converted into a probability value. The softmax function can normalize multiple similarity values ​​into a probability distribution. The formula is as follows: ; in, Indicates the fault type The probability of Indicates the fault type The similarity score of is the total number of fault types. Through the softmax function, the sum of the probabilities of all fault types is 1, indicating the relative possibility of each fault type. Set a fault diagnosis threshold. This threshold is used to filter out fault types with a higher probability of occurrence. If the probability of a certain fault type is greater than the set threshold, it is considered that the fault type may occur. In this way, all possible fault types are filtered out and multi-label classification is performed. The result of multi-label classification may contain multiple fault types, especially when the device may have multiple faults at the same time, and finally a fault diagnosis result is generated.

[0030] In a specific embodiment, the process of executing step 104 may specifically include the following steps: The operating status, fault diagnosis results and production task information of the CIR equipment are encoded to obtain a state vector, which is used as the input of the dual deep Q network. The dual deep Q network includes a main network and a target network. Each network consists of a shared feature extraction layer and an independent value function estimation layer and advantage function estimation layer. The feature extraction layer is configured to include three fully connected layers, each using the ReLU activation function, with output dimensions of 256, 128, and 64 respectively; the value function estimation layer and the advantage function estimation layer are configured to include two fully connected layers each, using the ReLU activation function, and the output dimension of the last layer is equal to the size of the action space; The state vector is input into the feature extraction layer of the main network to obtain the feature vector, and the feature vector is input into the value function estimation layer and the advantage function estimation layer respectively, and the Q value is obtained through mathematical operations; Select actions based on the ε-greedy strategy, apply the selected actions to the CIR device, observe the next state and reward value, and store the state transition samples in the experience replay buffer; Randomly sample a batch of state transition samples from the experience replay buffer and calculate the target Q value and current Q value, where the target Q value is calculated by the target network; The temporal difference error is calculated based on the target Q value and the current Q value, and the mean square error is used as the loss function to perform gradient descent optimization on the main network parameters, with an update cycle of every 4 time steps; The parameters of the main network are copied to the target network every 100 time steps to achieve soft update of the target network until convergence, and a dynamic scheduling optimization model is obtained.

[0031] Specifically, the operating status, fault diagnosis results and production task information of the CIR equipment are encoded. The operating status of the CIR equipment includes sensor data such as the temperature and pressure of the current equipment, while the fault diagnosis result is the current fault type of the equipment or its health status. The production task information includes the type of task that the equipment is currently to complete, the task priority and resource requirements. By converting multi-dimensional information into a numerical state vector, it provides input for the subsequent reinforcement learning algorithm. The state vector is dynamic and is continuously updated as the operating status, fault conditions and task requirements of the equipment change. A dual deep Q network is used to process the state vector. The dual deep Q network consists of a main network (main Q network) and a target network (target Q network). The two networks have similar structures but different weights. The main advantage of the dual deep Q network is that it can reduce the deviation in Q value estimation and improve the stability of learning. Each network contains a shared feature extraction layer and an independent value function estimation layer and advantage function estimation layer. The task of the feature extraction layer is to extract useful features from the input state vector, while the value function estimation layer and the advantage function estimation layer are used to estimate the value and advantage of different actions respectively, so as to generate the final Q value through combination. The configuration of the feature extraction layer includes three fully connected layers, each of which uses the ReLU activation function. The output dimension of the first fully connected layer is 256, the output dimension of the second fully connected layer is 128, and the output dimension of the third fully connected layer is 64. The ReLU activation function can effectively solve the gradient vanishing problem and enhance the nonlinear expression ability of the network. Through these layers, the input state vector is gradually converted into feature vectors for subsequent value function and advantage function estimation layers to process. The value function estimation layer and the advantage function estimation layer are configured. Each estimation layer contains two fully connected layers, which also use the ReLU activation function. The task of the value function estimation layer is to estimate the overall value of the current state, while the task of the advantage function estimation layer is to estimate the advantage relative to the average value under different actions. The dimension of the final output layer is equal to the size of the action space, which means that the network will output the Q value of each possible action. When the state vector is input to the feature extraction layer of the main Q network, a feature vector is generated. This feature vector is then input to the value function estimation layer and the advantage function estimation layer respectively, and the final Q value is calculated by the following formula: ; in, Yes Status and actions The Q value, Indicates status The value function of Indicates action Relative to the state The advantage function of is the action space, which represents the set of all possible actions. Through this formula, the network can calculate the value of different actions and guide the scheduling decision of CIR devices. In order to select an action, use -30 strategy. This strategy combines the balance of exploration and utilization, allowing the network to randomly select actions (exploration) with a certain probability to prevent falling into the local optimal solution; the rest of the time, the action with the largest current Q value is selected (utilization), so as to make decisions according to the current optimal strategy. The ε value is usually larger in the early stage of training, allowing more exploration, and gradually decreases as training progresses. The selected action will be applied to the CIR device, and the device will perform the corresponding scheduling operation. Observe the next state and reward value after executing the action, and store the state transition samples (including the current state, action, reward and next state) in the experience replay buffer. In order to improve training efficiency and stability, a batch of state transition samples are randomly sampled from the experience replay buffer. For each sample, the main network calculates the current Q value, and the target network calculates the target Q value. The calculation formula for the target Q value is: ; in, is the reward value, is the discount factor (usually between 0 and 1, indicating the weight of future rewards), For the next state, is the best action in the next state. The time difference error (TD error) is calculated by the difference between the target Q value and the current Q value, that is: ; The temporal difference error reflects the deviation between the model's predicted value and the actual observed reward, and the goal of the network is to minimize this error. In order to optimize the network parameters, the mean square error is used as the loss function and optimized by the gradient descent method. The formula for the mean square error is: ; in, is the loss value, is the batch size, The time difference error for each sample. The main network is optimized by gradient descent in each training cycle, and the update cycle is every 4 time steps. In order to maintain the synchronization between the main network and the target network, the parameters of the main network are copied to the target network every 100 time steps to achieve soft update of the target network, thereby avoiding unstable training caused by too frequent updates of the target network, and ensuring that the target network can reflect the latest learning results of the main network. As the training progresses, the network parameters gradually converge, and finally an optimization model that can effectively cope with the dynamic scheduling of CIR devices is obtained.

[0032] In a specific embodiment, the process of executing step 105 may specifically include the following steps: Initialize the parameters of the dynamic scheduling optimization model, set the forgetting factor λ and the initial covariance matrix P, where the value range of λ is (0, 1], and P is a large positive number multiplied by the unit matrix; Based on the real-time status and actions of the CIR device, the current Q value is calculated to obtain the actual Q value; Based on the dynamic scheduling optimization model, the maximum Q value at the next moment is predicted, and the target Q value is calculated by combining the reward value r, where the target Q value is equal to r plus the discount factor γ multiplied by the predicted maximum Q value; Compare the actual Q value with the target Q value and calculate the time difference error δ, which is equal to the target Q value minus the actual Q value; Based on the current state and action, extract the current feature vector x, transpose the current feature vector x, multiply it with the covariance matrix P, and then do the inner product with the current feature vector x to obtain the intermediate variable z; Using the forgetting factor λ and the intermediate variable z, calculate the gain vector k, which is equal to the product of the covariance matrix P and the eigenvector x divided by (λ plus z); Based on the gain vector k and the time difference error δ, update the parameter θ of the dynamic scheduling optimization model, and the update amount of θ is equal to k multiplied by δ; The covariance matrix P is updated until the parameters converge or the preset number of iterations is reached to obtain an updated optimization model.

[0033] Specifically, the parameters of the dynamic scheduling optimization model are initialized, and the forgetting factor λ and the initial covariance matrix P are set. The forgetting factor λ ranges from (0, 1], which is used to control the degree of "forgetting" of historical data by the model. A smaller λ value means that the model "forgets" old data faster and focuses more on new data; a larger λ value allows the model to retain more historical information. The covariance matrix P is used to measure the uncertainty of the model parameters, and is initially set to a large positive number multiplied by the unit matrix. This setting ensures that the model has greater flexibility in updating parameters in the initial stage, thereby accelerating the early convergence process. In the actual operation of the CIR device, based on the current state of the device and the actions taken, the model calculates the Q value at the current moment. The Q value represents the expected return of taking an action in a specific state. By combining the real-time sensor data, fault information and production task conditions of the device, the current state vector is formed and input into the dynamic scheduling optimization model to calculate the actual Q value. This step reflects the model's instant evaluation of the current scheduling strategy. The model determines the future scheduling strategy by predicting the maximum Q value at the next moment. Based on the current state of the device, the model can predict the maximum Q value of the device in a possible future state, combined with the device's instant reward value. , calculate the target Q value. The calculation formula of the target Q value is: ; in, is the current reward value, is the discount factor (usually between 0 and 1, representing the discount rate for future returns), For the state of the next moment, is the optimal action in the next state. The closer the discount factor is to 1, the more the model values ​​future rewards. The target Q value can help the model update its estimate of future states and gradually correct the current scheduling strategy. In order to continuously optimize the model, the actual Q value is compared with the predicted target Q value. The difference between the two is the time difference error The calculation formula is: ; Time difference error It measures the deviation between the model's estimate of the current state and action and the future expectation. The goal of the model is to minimize this error by adjusting the parameters. In order to update the parameters of the model, the feature vector is extracted from the current state and action. The feature vector can capture the key features of the current state and action and serve as an important basis for updating the model. After transposition and covariance matrix Multiply by the eigenvector Perform inner product operation to obtain intermediate variables The formula is: ; in, is the feature vector The transpose of is the covariance matrix. Intermediate variables Reflects the correlation between the current eigenvector and the covariance matrix. Gain vector It is a key element in recursive least squares and is used to adjust the parameters of the model. and intermediate variables , the calculation formula of the gain vector is: ; Gain Vector It reflects the importance of the current feature vector in the parameter update process, and its size directly affects the update amplitude of the model parameters. and time difference error , the parameters of the model The update amount of the parameter is equal to the gain vector Multiply by the time difference error , the formula is: ; Through this formula, the parameters of the model are In each time step, corrections are made based on the deviation between the actual and predicted values, so that the model can gradually learn a better scheduling strategy. Update, covariance matrix It also needs to be adjusted in real time. The update formula of the covariance matrix is: ; This update formula ensures that the covariance matrix can gradually converge over time, and the uncertainty of the parameters is continuously reduced during the model learning process. Through the iterative update of the recursive least squares method, the parameters of the model gradually stabilize until convergence or the preset number of iterations is reached. As a result of convergence, the model can make more accurate scheduling decisions based on the real-time status of the CIR device and environmental changes, thereby improving the overall performance of the device.

[0034] In a specific embodiment, the process of executing step 106 may specifically include the following steps: Model the production efficiency, energy consumption and equipment life of CIR equipment and construct a set of objective functions for multi-objective optimization problems; Based on the updated optimization model, N initial solutions are generated, each of which corresponds to a set of candidate scheduling strategies for CIR devices, and the N generated initial solutions are sorted non-dominated to obtain a Pareto frontier solution set; Randomly select two solutions a1 and a2 from the Pareto front solution set, use the simulated binary crossover operator to perform a crossover operation on a1 and a2 to generate two offspring solutions y1 and y2; Perform polynomial mutation operator processing on the offspring solutions y1 and y2 to generate mutated solutions z1 and z2, and merge the mutated solutions z1 and z2 with the N initial solutions to form a new solution set; The probability penalty boundary intersection value is calculated for each solution in the solution set, and the solution set is sorted and selected based on the probability penalty boundary intersection value, and the N optimal solutions are retained as the next generation population until the preset number of iterations or convergence conditions are reached, and the scheduling strategy of the CIR device is obtained.

[0035] Specifically, the production efficiency, energy consumption and equipment life of CIR equipment are modeled. For production efficiency, the objective function can be defined as the output of equipment per unit time, denoted as ,in represents the current scheduling strategy. The higher the production efficiency, the greater the value of the objective function. For energy consumption, the objective function can be defined as the amount of energy used by the equipment per unit time, recorded as , this function should be minimized, that is, the less energy consumption, the better. The equipment life is closely related to the maintenance and operating conditions of the equipment. The objective function It can be defined as the consumption of equipment life based on the current operating state, with the goal of minimizing the consumption of equipment life. By modeling the three objectives of production efficiency, energy consumption, and equipment life, a multi-objective optimization problem is constructed, in which each objective function may reach a balance under mutual constraints. In the multi-objective optimization framework, the updated optimization model is used to generate the initial solution. Hypothesis generation Initial solutions, each of which represents a different CIR equipment scheduling strategy. These scheduling strategies include different production task priorities, energy consumption allocations, and maintenance cycle arrangements. Each solution can be represented as a multidimensional vector, where each dimension corresponds to a specific scheduling parameter. The initial solutions can be classified by non-dominated sorting. Non-dominated sorting evaluates the quality of solutions based on the Pareto dominance relationship. Dominate another solution The condition is No worse than , and outperforms at least one objective . By performing non-dominated sorting on all solutions, we determine which solutions are on the Pareto front, that is, these solutions have achieved a balance on multiple objectives and cannot be further optimized by improving one objective without compromising other objectives. Two solutions are randomly selected from the Pareto front solution set, denoted as a1 and a2. These two solutions represent different scheduling strategies. In order to explore new solution spaces, a simulated binary crossover operator (SBX) is used to perform crossover operations on a1 and a2 to generate two new offspring solutions, denoted as y1 and y2. SBX is a commonly used genetic algorithm crossover operator that generates new solutions by simulating the gene crossover process in binary genetic algorithms. The SBX operator generates offspring solutions y1 and y2 based on the characteristics of the parent solutions a1 and a2 according to a certain probability, thereby achieving effective exploration in the solution space. After the offspring solutions are generated, y1 and y2 are processed by the polynomial mutation operator. The role of the mutation operator is to further increase the diversity of solutions by randomly adjusting the parameters of certain solutions. The polynomial mutation operator generates new solutions by applying random perturbations to each dimension of the solution. The mutated solutions are recorded as z1 and z2. This mutation operation can help the algorithm escape from the local optimal solution and increase the globality of the search. The initial solutions are merged to form a new solution set. In the new solution set, the probability penalty boundary intersection value (PB1) is calculated for each solution. PBI is an indicator to measure the distance between the solution and the ideal solution in a multi-objective optimization problem, which can balance the diversity and convergence of the solution. The smaller the PBI value, the closer the solution is to the ideal non-dominated frontier in the target space. The calculation formula of PBI can be expressed as: ; in, represents the distance from the solution to the reference direction, represents the vertical distance of the solution from this direction, It is a tuning parameter used to balance the diversity and convergence of solutions. After calculating the PBI value of each solution, the solution set is sorted based on these values. After sorting, the solution with the smallest PB1 value is selected. The algorithm selects the solutions that perform best in the current iteration and have high optimization potential as the next generation population. The algorithm repeats the above process. Through continuous selection, crossover, mutation, and screening, the quality of the solution is improved in each generation. The newly generated population will continue to evolve until the preset number of iterations is reached or the convergence condition is met. The convergence condition usually means that the Pareto front solution set no longer changes significantly in several iterations, or the objective function value of the algorithm reaches a stable state. At this point, the algorithm believes that it has found a sufficiently good solution and finally obtains the scheduling strategy for the CIR device.

[0036] The CIR device management method in the embodiment of the present invention is described above. The CIR device management apparatus in the embodiment of the present invention is described below. Figure 2 In one embodiment of the present invention, a CIR device management device includes: The correction module 201 is used to correct and normalize the historical temperature data of the CIR device to obtain the historical panoramic temperature characteristics; A training module 202 is used to input the historical panoramic temperature features into a preset Siamese-MobileNet network for training to obtain a trained network model; A calculation module 203 is used to perform similarity calculation on the real-time panoramic temperature features of the CIR device based on the trained network model to obtain a fault diagnosis result; A construction module 204 is used to construct a dual deep Q network based on the fault diagnosis result, train the dual deep Q network, and obtain a dynamic scheduling optimization model; An updating module 205 is used to perform online parameter update of the dynamic scheduling optimization model based on a recursive least squares method to obtain an updated optimization model; The optimization module 206 is used to perform multi-objective optimization calculation based on the updated optimization model in combination with the probability penalty moment boundary intersection algorithm to obtain the scheduling strategy of the CIR device.

[0037] Through the collaboration of the above components, by correcting and normalizing the historical temperature data of the CIR equipment and combining the training of the Siamese-MobileNet network, the temperature characteristics of the equipment can be captured more accurately, thereby improving the accuracy of fault diagnosis. Based on the trained network model, the similarity calculation of the real-time panoramic temperature characteristics can quickly identify the potential fault type and realize the early warning of the fault. By constructing a dual deep Q network and combining the recursive least squares method for online parameter update, the scheduling strategy can be dynamically adjusted according to the real-time status of the equipment and production demand to improve the adaptability of the system. The probability penalty moment boundary intersection algorithm is used for multi-objective optimization calculation, which can simultaneously consider multiple objectives such as production efficiency, energy consumption and equipment life to achieve more comprehensive optimization. By using lightweight network structures such as MobileNetV3 and combining the fast convergence characteristics of the recursive least squares method, this method greatly reduces the computational complexity while ensuring the optimization effect, and is suitable for online real-time applications. Through the application of Siamese network structure and contrastive learning, it can better handle small sample problems and improve the generalization ability of the model when facing new faults. Through timely fault diagnosis and optimized scheduling strategies, the unplanned downtime of CIR equipment can be effectively reduced, and the overall reliability and production efficiency can be improved.

[0038] above Figure 2 The CIR device management apparatus in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The CIR device management device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0039] Figure 3 3 is a schematic diagram of the structure of a CIR device management device provided by an embodiment of the present invention. The CIR device management device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 may be temporary storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the CIR device management device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the CIR device management device 300 to implement the steps of the above-mentioned CIR device management method.

[0040] The CIR device management device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The structure of the CIR device management device shown does not constitute a limitation on the CIR device management device provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0041] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the CIR device management method.

[0042] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0043] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A CIR device management method, characterized in that: The method comprises: Correct and normalize the historical temperature data of the CIR device to obtain the historical panoramic temperature characteristics; Inputting the historical panoramic temperature features into a preset Siamese-MobileNet network for training to obtain a trained network model; Based on the trained network model, similarity calculation is performed on the real-time panoramic temperature features of the CIR device to obtain a fault diagnosis result; Building a dual deep Q network based on the fault diagnosis result, and training the dual deep Q network to obtain a dynamic scheduling optimization model; Based on the recursive least squares method, the parameters of the dynamic scheduling optimization model are updated online to obtain an updated optimization model; Based on the updated optimization model, a multi-objective optimization calculation is performed in combination with a probability penalty moment boundary intersection algorithm to obtain a scheduling strategy for the CIR device.

2. The CIR device management method according to claim 1, characterized in that: The historical temperature data of the CIR device is corrected and normalized to obtain the historical panoramic temperature features, including: Installing temperature sensors on multiple target locations of the CIR device, collecting historical temperature data, and performing curve fitting on the historical temperature data to obtain an original temperature curve; Performing ambient temperature compensation on the original temperature curve to obtain a compensated temperature curve, and performing sensor sensitivity correction on the compensated temperature curve to obtain a corrected temperature curve; Performing maximum and minimum value normalization processing on the corrected temperature curve to obtain a normalized temperature curve, and arranging the normalized temperature curve in chronological order to form a one-dimensional vector to obtain a panoramic temperature feature of a single time series; Performing sliding window processing on the panoramic temperature feature of the single time series to obtain a plurality of subsequences of fixed lengths, and performing Fourier transform on the plurality of subsequences of fixed lengths to obtain frequency domain panoramic temperature features; The frequency domain panoramic temperature feature is concatenated with the panoramic temperature feature of the single time series to obtain a historical panoramic temperature feature.

3. The CIR device management method according to claim 1, characterized in that: The inputting the historical panoramic temperature features into a preset Siamese-MobileNet network for training to obtain a trained network model includes: Randomly pairing the historical panoramic temperature features to obtain a training sample set including positive sample pairs and negative sample pairs; Construct a Siamese-MobileNet network structure, wherein the Siamese-MobileNet network includes two parameter-sharing MobileNetV3 subnetworks and a distance measurement layer; The MobileNetV3 subnetwork is configured, each MobileNetV3 subnetwork includes an initial convolution layer, five inverted residual blocks and a global average pooling layer; the initial convolution layer uses a 3x3 convolution kernel, a step size of 2, and 16 output channels; the inverted residual block uses 1x1 convolution, depthwise separable convolution, and 1x1 convolution in sequence, and a batch normalization layer and a ReLU6 activation function are added after each convolution operation; The distance measurement layer is set, and cosine similarity is used as a distance measurement indicator to calculate the similarity of the output feature vectors of the two MobileNetV3 sub-networks; Input the sample pairs in the training sample set into two MobileNetV3 sub-networks respectively to obtain two training feature vectors; Performing cosine similarity calculation on the two training feature vectors to obtain a similarity score of the sample pair, and calculating a contrast loss function value based on the similarity score and the true label of the sample pair; The Adam optimizer is used to optimize the parameters of the Siamese-MobileNet network, and the network performance index is calculated on the validation set. When the performance index on the validation set no longer improves, the training process is stopped to obtain the trained network model.

4. The CIR device management method according to claim 3, characterized in that: The similarity calculation is performed on the real-time panoramic temperature features of the CIR device based on the trained network model to obtain a fault diagnosis result, including: Collecting real-time temperature data of the CIR device, and preprocessing the real-time temperature data to obtain real-time panoramic temperature features; Inputting the real-time panoramic temperature feature into the first MobileNetV3 subnetwork of the trained network model, and sequentially processing through an initial convolutional layer, five inverted residual blocks, and a global average pooling layer to obtain a real-time feature vector; Selecting multiple fault temperature feature samples of known types from a pre-established fault sample library, and inputting each fault sample into the second MobileNetV3 subnetwork of the trained network model in turn to obtain multiple fault feature vectors; Performing cosine similarity calculation on the real-time feature vector and each fault feature vector at a distance measurement layer to obtain multiple similarity scores; Based on the multiple similarity scores, construct a fault probability distribution matrix, each element in the matrix represents the similarity between the real-time feature and the preset fault type; Performing a softmax function operation on the fault probability distribution matrix, converting the similarity scores into probability values, and obtaining a normalized fault probability distribution; A fault diagnosis threshold is set, the normalized fault probability distribution is subjected to threshold screening, fault types with probability values ​​greater than the fault diagnosis threshold are screened out, and multi-label classification is performed on the screened fault types to obtain a fault diagnosis result.

5. The CIR device management method according to claim 1, characterized in that: The step of constructing a dual deep Q network based on the fault diagnosis result and training the dual deep Q network to obtain a dynamic scheduling optimization model includes: Encoding the operating status, fault diagnosis results and production task information of the CIR equipment to obtain a state vector, which is used as the input of a dual deep Q network, wherein the dual deep Q network includes a main network and a target network, each network consisting of a shared feature extraction layer and independent value function estimation layers and advantage function estimation layers; The feature extraction layer is configured to include three fully connected layers, each layer uses a ReLU activation function, and the output dimensions are 256, 128, and 64 respectively; the value function estimation layer and the advantage function estimation layer are configured to include two fully connected layers each, use a ReLU activation function, and the output dimension of the last layer is equal to the size of the action space; The state vector is input into the feature extraction layer of the main network to obtain a feature vector, and the feature vector is respectively input into the value function estimation layer and the advantage function estimation layer to obtain the Q value through mathematical operation; Select actions based on the ε-greedy strategy, apply the selected actions to the CIR device, observe the next state and reward value, and store the state transition samples in the experience replay buffer; Randomly sampling a batch of state transition samples from the experience replay buffer, and calculating a target Q value and a current Q value, wherein the target Q value is calculated by the target network; Calculate the temporal difference error based on the target Q value and the current Q value, use the mean square error as the loss function, perform gradient descent optimization on the main network parameters, and update the period every 4 time steps; The parameters of the main network are copied to the target network every 100 time steps to achieve soft update of the target network until convergence, and a dynamic scheduling optimization model is obtained.

6. The CIR device management method according to claim 1, characterized in that: The online updating of parameters of the dynamic scheduling optimization model based on the recursive least squares method to obtain an updated optimization model includes: Initialize the parameters of the dynamic scheduling optimization model, set the forgetting factor λ and the initial covariance matrix P, where the value range of λ is (0, 1], and P is a large positive number multiplied by the unit matrix; Based on the real-time status and actions of the CIR device, the current Q value is calculated to obtain the actual Q value; Based on the dynamic scheduling optimization model, the maximum Q value at the next moment is predicted, and the target Q value is calculated by combining the reward value r, where the target Q value is equal to r plus the discount factor γ multiplied by the predicted maximum Q value; Compare the actual Q value with the target Q value and calculate the time difference error δ, which is equal to the target Q value minus the actual Q value; Based on the current state and action, extract the current feature vector x, transpose the current feature vector x, multiply it with the covariance matrix P, and then do the inner product with the current feature vector x to obtain the intermediate variable z; Using the forgetting factor λ and the intermediate variable z, calculate the gain vector k, which is equal to the product of the covariance matrix P and the eigenvector x divided by (λ plus z); Based on the gain vector k and the time difference error δ, update the parameter θ of the dynamic scheduling optimization model, and the update amount of θ is equal to k multiplied by δ; The covariance matrix P is updated until the parameters converge or the preset number of iterations is reached to obtain an updated optimization model.

7. The CIR device management method according to claim 1, characterized in that: The multi-objective optimization calculation is performed based on the updated optimization model in combination with the probability penalty moment boundary intersection algorithm to obtain the scheduling strategy of the CIR device, including: Model the production efficiency, energy consumption and equipment life of CIR equipment and construct a set of objective functions for multi-objective optimization problems; Based on the updated optimization model, N initial solutions are generated, each solution corresponds to a candidate scheduling strategy for a group of CIR devices, and the generated N initial solutions are non-dominated sorted to obtain a Pareto front solution set; Randomly select two solutions a1 and a2 from the Pareto front solution set, perform a crossover operation on a1 and a2 using a simulated binary crossover operator, and generate two offspring solutions y1 and y2; Perform polynomial mutation operator processing on the offspring solutions y1 and y2 to generate mutated solutions z1 and z2, and merge the mutated solutions z1 and z2 with the N initial solutions to form a new solution set; The probability penalty boundary intersection value is calculated for each solution in the solution set, and the solution set is sorted and selected based on the probability penalty boundary intersection value, and N optimal solutions are retained as the next generation population until a preset number of iterations or convergence conditions are reached to obtain the scheduling strategy of the CIR device.

8. A CIR equipment management device, characterized in that: The device is used to execute the CIR device management method according to any one of claims 1 to 7, comprising: The correction module is used to correct and normalize the historical temperature data of the CIR device to obtain the historical panoramic temperature characteristics; A training module, used for inputting the historical panoramic temperature features into a preset Siamese-MobileNet network for training to obtain a trained network model; A calculation module, used for performing similarity calculation on the real-time panoramic temperature features of the CIR device based on the trained network model to obtain a fault diagnosis result; A construction module, used for constructing a dual deep Q network based on the fault diagnosis result, training the dual deep Q network, and obtaining a dynamic scheduling optimization model; An updating module, used for performing online parameter updating of the dynamic scheduling optimization model based on a recursive least squares method to obtain an updated optimization model; The optimization module is used to perform multi-objective optimization calculation based on the updated optimization model in combination with the probability penalty moment boundary intersection algorithm to obtain the scheduling strategy of the CIR device.

9. A CIR device management device, characterized in that: The CIR device management device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory to enable the CIR device management device to execute the CIR device management method according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the CIR device management method according to any one of claims 1 to 7 is implemented.

Images