CPU junction temperature control method and system

Abstract

The invention discloses a CPU junction temperature control method and system, and the method comprises the steps: obtaining first temperature data collected by an internal temperature sensor and n pieces of second temperature data collected by n ambient temperature sensors disposed around, constructing a temperature field grid topological graph, and carrying out the calculation to obtain standardized temperature field data; carrying out first-order and second-order derivative calculation of time dimension, constructing spatio-temporal characteristic sequence data, and calculating to obtain topological characteristic vector data of the temperature field; the attention weight is adopted, and fusion feature vector data are obtained through calculation; calculating to obtain system stability index data by adopting a gating circulation unit network; constructing a workload feature vector, and performing classification processing to obtain load type index data; calculating to obtain multi-level early warning threshold set data; and a prediction model is constructed and optimized to obtain control parameter set data, and a final control instruction sequence is generated. According to the invention, the temperature control precision is improved, and the reliability and the use efficiency of the power utilization information acquisition terminal are improved.

Description

Technical Field

[0001] The invention belongs to the field of measurement and control of CPU junction temperature, in particular to a method and system for controlling CPU junction temperature. Background Art

[0002] As a key device in the perception layer of the smart grid, the power consumption information collection terminal undertakes the important tasks of collecting, processing and transmitting power consumption data. Such terminals usually need to operate stably for a long time in a complex and changeable external environment, and their working environment often faces harsh conditions such as high temperature, humidity, and dust. In this case, the CPU, as the core processing unit of the terminal, has a direct impact on the reliability and service life of the equipment due to the stability of its junction temperature. Especially in southern my country, the ambient temperature often exceeds 40°C in summer, and the heat generated by the equipment itself makes it easy for the CPU junction temperature to exceed the safety threshold. Therefore, studying the control method of the CPU junction temperature of the power consumption information collection terminal is of great practical significance for ensuring the reliability of power grid data collection and improving the stability of terminal equipment.

[0003] At present, the temperature control of the CPU of the power consumption information collection terminal mainly adopts passive heat dissipation and simple threshold control methods. At the hardware level, common heat dissipation solutions include passive heat dissipation measures such as using PCB boards with good thermal conductivity, increasing the heat sink area, and optimizing device layout. At the software level, it is mainly through setting a fixed temperature threshold. When it is detected that the CPU temperature exceeds the preset threshold, simple temperature control is performed by reducing the CPU main frequency and reducing data processing tasks. Some terminals use fans to assist in heat dissipation, but the control strategy is relatively simple, often using a single switch control or a graded control based on a fixed temperature range. These traditional methods can meet basic needs under general working conditions, but there are still many shortcomings in the face of complex and changeable actual working environments.

[0004] Specifically, the existing technology has the following major problems: First, in terms of temperature sampling and preprocessing, there is a lack of a collaborative analysis mechanism for multi-point temperature data, which makes it difficult to accurately identify and filter abnormal temperature data, resulting in insufficient reliability of temperature monitoring. Second, in terms of temperature field reconstruction, the traditional linear interpolation method cannot accurately describe the spatial distribution characteristics of temperature, especially in hot spots. Distortion is prone to occur. Third, the existing feature extraction method mainly focuses on the absolute value of temperature, ignores the spatiotemporal correlation characteristics of temperature changes, and is difficult to predict temperature anomaly trends in a timely manner. Fourth, the fixed temperature threshold setting cannot adapt to the dynamic changes of ambient temperature and workload, which can easily lead to false alarms or missed alarms. Fifth, simple switch control or PID control methods are difficult to balance control accuracy and energy consumption, and temperature fluctuations are prone to occur when the load changes sharply. Sixth, there is a lack of a smooth processing mechanism for control instructions. Frequent control switching not only affects system stability, but also increases energy consumption. The existence of these technical problems seriously restricts the reliability and efficiency of power consumption information collection terminals. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for controlling CPU junction temperature to solve the above problems existing in the prior art.

[0006] Technical solution, a method for controlling CPU junction temperature, comprising the following steps:

[0007] S1, obtaining first temperature data collected by a CPU internal temperature sensor and n second temperature data collected by n ambient temperature sensors arranged around the CPU, and processing the first temperature data and the n second temperature data by sliding median filtering to obtain filtered first temperature sequence data and n second temperature sequence data; constructing a temperature field grid topology map based on the filtered first temperature sequence data and the n second temperature sequence data; and calculating standardized temperature field data based on the temperature data in the temperature field grid topology map by using a Z-score normalization method; wherein n is a natural number greater than 1;

[0008] S2. Calculate the first-order and second-order derivatives of the standardized temperature field data in the time dimension to obtain temperature change rate field data and temperature change acceleration field data; construct spatiotemporal feature sequence data based on the standardized temperature field data, the temperature change rate field data and the temperature change acceleration field data; calculate and obtain the temperature field topological feature vector data based on the temperature distribution relationship in the standardized temperature field data;

[0009] S3. Based on the spatiotemporal feature sequence data and the temperature field topological feature vector data, the attention weight is used to calculate the fused feature vector data; based on the historical sequence of the fused feature vector data, the gated recurrent unit network is used to calculate the predicted sequence data of the next m time points; based on the predicted sequence data, the system stability index data is calculated; where m is a natural number greater than 1;

[0010] S4, obtaining the CPU usage data and memory usage data collected by the CPU status monitoring module; constructing a workload feature vector based on the fused feature vector data, the CPU usage data and the memory usage data; obtaining load type index data based on the workload feature vector after classification processing; calculating the environmental impact coefficient based on the filtered n second temperature series data; calculating and obtaining multi-level warning threshold set data based on the load type index data, the environmental impact coefficient and the system stability index data;

[0011] S5. Based on the fused feature vector data and the multi-level warning threshold set data, a prediction model is constructed and optimized to obtain the control parameter set data; based on the control parameter set data, a piecewise linear control instruction sequence is generated; the power consumption data collected by the power consumption monitoring module is obtained, and based on the power consumption data, the piecewise linear control instruction sequence is optimized to obtain the final control instruction sequence.

[0012] A control system for CPU junction temperature, comprising:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the method for controlling the CPU junction temperature.

[0016] Beneficial effects: The present invention realizes accurate identification and filtering of abnormal temperature data, improves the effectiveness of temperature monitoring, realizes dynamic adjustment of warning thresholds and global optimization of control parameters, improves temperature control accuracy, and also improves the reliability and efficiency of electricity consumption information collection terminals. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the present invention.

[0018] Figure 2 This is a flow chart of step S1 of the present invention.

[0019] Figure 3 This is a flow chart of step S2 of the present invention.

[0020] Figure 4 This is a flow chart of step S3 of the present invention.

[0021] Figure 5 This is a flow chart of step S4 of the present invention.

[0022] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, the present application proposes a method for controlling the junction temperature of a CPU, comprising the following steps:

[0024] S1, obtaining first temperature data collected by a CPU internal temperature sensor and n second temperature data collected by n ambient temperature sensors arranged around the CPU, and processing the first temperature data and the n second temperature data by sliding median filtering to obtain filtered first temperature sequence data and n second temperature sequence data; constructing a temperature field grid topology map based on the filtered first temperature sequence data and the n second temperature sequence data; and calculating standardized temperature field data based on the temperature data in the temperature field grid topology map by using a Z-score normalization method; wherein n is a natural number greater than 1;

[0025] S2. Calculate the first-order and second-order derivatives of the standardized temperature field data in the time dimension to obtain temperature change rate field data and temperature change acceleration field data; construct spatiotemporal feature sequence data based on the standardized temperature field data, the temperature change rate field data and the temperature change acceleration field data; calculate and obtain the temperature field topological feature vector data based on the temperature distribution relationship in the standardized temperature field data;

[0026] S3. Based on the spatiotemporal feature sequence data and the temperature field topological feature vector data, the attention weight is used to calculate the fused feature vector data; based on the historical sequence of the fused feature vector data, the gated recurrent unit network is used to calculate the predicted sequence data of the next m time points; based on the predicted sequence data, the system stability index data is calculated; where m is a natural number greater than 1;

[0027] S4, obtaining the CPU usage data and memory usage data collected by the CPU status monitoring module; constructing a workload feature vector based on the fused feature vector data, the CPU usage data and the memory usage data; obtaining load type index data based on the workload feature vector after classification processing; calculating the environmental impact coefficient based on the filtered n second temperature series data; calculating and obtaining multi-level warning threshold set data based on the load type index data, the environmental impact coefficient and the system stability index data;

[0028] S5. Based on the fused feature vector data and the multi-level warning threshold set data, a prediction model is constructed and optimized to obtain the control parameter set data; based on the control parameter set data, a piecewise linear control instruction sequence is generated; the power consumption data collected by the power consumption monitoring module is obtained, and based on the power consumption data, the piecewise linear control instruction sequence is optimized to obtain the final control instruction sequence.

[0029] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0030] S11, obtaining first temperature data T1(t) collected by a CPU internal temperature sensor at time t and n second temperature data T2i(t) collected by n ambient temperature sensors arranged around the CPU at time t; wherein the value range of i is 1 to n; performing median filtering on the first temperature data and the n second temperature data using a sliding window of length k to obtain filtered first temperature sequence data T1'(t) and n second temperature sequence data T2i'(t);

[0031] S12, taking the sampling points T1'(t) of the filtered first temperature sequence data and the sampling points of the n second temperature sequence data T2i'(t) as vertices, and using the Delaunay triangulation algorithm to construct a temperature field grid topology graph G(t) having a vertex set V and an edge set E; wherein the vertex set V includes the sampling points of the first temperature sequence data T1'(t) and the n second temperature sequence data T2i'(t), and the edge set E includes edges generated based on the Delaunay triangulation algorithm;

[0032] S13. Based on the pre-stored historical temperature data, the temperature mean μ and standard deviation σ are calculated; based on the temperature mean μ and standard deviation σ, the temperature data of the temperature field grid topology diagram G(t) is standardized according to the formula Z=(x-μ) / σ to obtain the standardized temperature field data G'(t); wherein x is the temperature data to be standardized, μ represents the temperature mean, σ represents the standard deviation, and Z represents the standardized temperature field data.

[0033] In one embodiment of the present application, the CPU internal temperature sensor data T1(t) and the multi-point ambient temperature sensor data {T2i(t)} (i=1, 2, ..., n) on the board are obtained, and the sliding median filter is used to remove outliers, and the cleaned temperature sequence T1'(t) and {T2i'(t)} are output. Based on the cleaned temperature sequence T1'(t) and {T2i'(t)}, the Delaunay triangulation algorithm is used to construct a temperature field grid G(V, E), where V is a set of temperature sampling points and E is an edge set, and the temperature field topology G(t) is output. Based on the temperature values ​​in the temperature field topology G(t), Z-score standardization is used: Z = (x-μ) / σ, where μ is the historical temperature mean and σ is the standard deviation, and the standardized temperature field G'(t) is output.

[0034] This embodiment uses sliding median filtering technology to pre-process the CPU internal temperature sensor data and multiple ambient temperature sensor data, which can effectively remove sudden interference and outliers in the sampling process and ensure the reliability of the data; combined with the Z-score standardization method, the temperature data of different dimensions are unified into the standard normal distribution interval, so that the data is comparable and consistent. In particular, for devices such as power consumption information collection terminals exposed to complex external environments, this embodiment can accurately reflect the actual working status of the terminal through the collaborative processing of multi-point temperature data, avoiding the misjudgment that may be caused by single-point temperature data. By constructing a standardized temperature field containing position and temperature information, a reliable data basis is provided for subsequent feature extraction and status evaluation, so that the system can maintain a stable monitoring effect under different environmental conditions and working conditions. In practical applications, this embodiment can increase the effectiveness of temperature monitoring from 90% to more than 98%, improving the system's anti-interference ability and adaptability.

[0035] According to one aspect of the present application, step S12 is further:

[0036] S121, based on the filtered first temperature series data T1'(t) and the n second temperature series data T2i'(t), calculate the spatial coordinate deviation value di=[(xi-x0) of each sampling point 2 +(yi-y0) 2 ] 1 / 2 , where (xi, yi) is the coordinate of the i-th sampling point, and (x0, y0) is the coordinate of the center of the region; based on the spatial coordinate deviation value di and the preset temperature threshold θ, the abnormal sampling points are deleted to obtain the valid sampling point set V1;

[0037] S122. Based on the valid sampling point set V1, an improved Bowyer-Watson algorithm is used to construct an initial Delaunay triangulation to obtain a triangle set T1; the minimum internal angle θmin and the maximum side length ratio R of each triangle in the triangle set are calculated to obtain a triangle quality evaluation index set Q1;

[0038] S123, based on the triangle set T1 and the triangle quality evaluation index set Q1, filter out triangles whose quality evaluation index is lower than a preset threshold, and insert new Steiner points into them to form an encrypted point set V2; merge the encrypted point set V2 with the valid sampling point set V1 to obtain an optimized sampling point set V3;

[0039] S124, re-execute Delaunay triangulation based on the optimized sampling point set V3 to obtain an optimized triangle set T2; extract all edges of the optimized triangle set T2 to obtain an edge set E1; process the edge set E1 according to a preset boundary constraint condition to obtain a final edge set E2;

[0040] S125. Based on the optimized sampling point set V3 and the final edge set E2, the piecewise cubic Hermite interpolation method is used to calculate the temperature value T(x, y) of any point (x, y) inside the grid. The formula is T(x, y)=Σ[hi(x, y)·Ti], where hi(x, y) is the interpolation basis function and Ti is the temperature value of the sampling point. A temperature field grid topology graph G(t) is constructed, where the vertex set V contains the position and temperature information of the sampling point set V3, and the edge set E contains the connection relationship of the final edge set E2.

[0041] This embodiment solves the accuracy and efficiency problems of temperature field reconstruction of the power consumption information collection terminal by constructing an improved Delaunay triangulation and adaptive grid optimization method. Specifically, an abnormal point detection mechanism based on spatial coordinate deviation and temperature threshold is adopted to effectively remove unreliable sampling points and improve the reliability of temperature field reconstruction; by introducing the Steiner point optimization strategy, the grid density of the area with large temperature gradient is adaptively encrypted, and the fine characterization of the hot spot area is achieved. Especially when the temperature difference between different components of the terminal is large, this embodiment realizes the smooth reconstruction of the temperature field through the segmented cubic Hermite interpolation method, avoiding the distortion problem of the traditional linear interpolation method in the temperature jump area. The grid quality evaluation mechanism based on the minimum internal angle and the maximum side length ratio is adopted to ensure the numerical stability of the temperature field reconstruction. This embodiment improves the accuracy of temperature field reconstruction to 0.1°C, the spatial resolution reaches the millimeter level, and the reconstruction time is shortened to less than 10ms. At the same time, the temperature gradient calculation error in the hot spot area is reduced by 75%, which improves the accuracy and real-time performance of temperature field reconstruction.

[0042] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0043] S21. Use the central difference method to calculate the first-order derivative of the standardized temperature field data G'(t) in the time dimension to obtain the temperature change rate field data Gv(t); the calculation formula of the first-order derivative is v(t)=[G'(t+Δt)-G'(t-Δt)] / (2Δt); use the central difference method to calculate the second-order derivative of the standardized temperature field data G'(t) in the time dimension to obtain the temperature change acceleration field data Ga(t), and the calculation formula of the second-order derivative is a(t)=[G'(t+Δt)-2G'(t)+G'(t-Δt)] / (Δt 2 ), where Δt is the sampling time interval, and G'(t) represents the standardized temperature field data at time t;

[0044] S22, calculating the spatial gradient of the standardized temperature field data G'(t) to obtain the temperature gradient field data ΔT(t); combining the standardized temperature field data G'(t), the temperature change rate field data Gv(t), the temperature change acceleration field data Ga(t) and the temperature gradient field data ΔT(t) into a four-dimensional vector at each sampling point to obtain the spatiotemporal feature sequence data F(t);

[0045] S23. Based on the temperature value of each sampling point in the standardized temperature field data G'(t), calculate the degree distribution characteristic value ki(t), clustering coefficient ci(t) and centrality index bi(t) of each sampling point; combine the degree distribution characteristic value ki(t), clustering coefficient ci(t) and centrality index bi(t) to form the temperature field topological characteristic vector data P(t). Wherein ki(t) represents the number of sampling points directly connected to the i-th sampling point; ci(t) is equal to the actual number of connections between the sampling points connected to the i-th sampling point divided by the maximum number of possible connections; bi(t) represents the ratio of the number of shortest paths passing through the i-th sampling point to the total number of shortest paths.

[0046] In one embodiment of the present application, based on the normalized temperature field G'(t), the first-order derivative v(t) = dT / dt and the second-order derivative a(t) = d 2 T / dt 2, where T represents the temperature scalar value, and outputs the temperature change rate field Gv(t) and the acceleration field Ga(t). Based on the standardized temperature field G'(t), the temperature change rate field Gv(t) and the acceleration field Ga(t), construct the feature vector F(t) =[T(t), v(t), a(t), ΔT(t)], where ΔT(t) is the temperature gradient field, and T(t) represents the function of temperature change over time, and output the spatiotemporal feature sequence F(t). Based on the temperature field topology graph G'(t), calculate the graph features (degree distribution, clustering coefficient, centrality), and output the topological feature vector P(t).

[0047] This embodiment achieves a comprehensive capture of the dynamic change law of CPU junction temperature by introducing a dual feature extraction mechanism of spatiotemporal feature sequence and topological feature vector. Specifically, by calculating the temperature change rate field and the acceleration field, it can not only reflect the instantaneous change trend of temperature, but also predict the change trend of temperature, providing a time margin for system early warning; by constructing the temperature gradient field, the unevenness of temperature in spatial distribution is accurately characterized, and the hot spot area and heat conduction path are effectively identified. This embodiment is particularly suitable for power consumption information collection terminals, which need to work stably for a long time in complex environments. In the case of rapid temperature changes, the temperature abnormality trend can be predicted 200ms in advance, so that the system has enough time for preventive control; in terms of spatial distribution, the spatial resolution of the temperature field can be increased to the millimeter level, and the heat source position can be accurately located, laying the foundation for subsequent precise control. This embodiment increases the prediction accuracy of temperature anomalies from the original 85% to more than 95%, improving the early warning capability of the system.

[0048] According to one aspect of the present application, step S22 is further:

[0049] S221, extract the temperature sequence {Tp(t-w+1), ..., Tp(t)} of each sampling point p(x, y) in the preset time window in the standardized temperature field data G'(t); calculate the statistical characteristics of the temperature sequence, including the mean μp(t), standard deviation σp(t), skewness sp(t) and kurtosis kp(t); based on the statistical characteristics, construct a time feature vector FT(t)=[μp(t), σp(t), sp(t), kp(t)];

[0050] S222, based on the standardized temperature field data G'(t), calculate the first-order spatial derivatives ΨT / Ψx and ΨT / Ψy of the temperature field in the x-direction and the y-direction, and the second-order spatial derivative Ψ 2 T / Ψx 2 , 2 T / Ψy 2 and 2T / ΨxΨy; Based on the first-order and second-order spatial derivatives, construct the spatial eigenvector FS(t); where Ψ represents the partial derivative;

[0051] S223, based on the temperature change rate field data Gv(t) and the temperature change acceleration field data Ga(t), calculate the change rate field spatial covariance matrix Cv(t) and the acceleration field spatial covariance matrix Ca(t); perform eigenvalue decomposition on the change rate field spatial covariance matrix and the acceleration field spatial covariance matrix, and take the largest k eigenvalues ​​to construct a dynamic eigenvector FD(t); wherein k is a natural number greater than 1;

[0052] S224, calculate the Laplace operator ΔT=Ψ of the standardized temperature field data G'(t) 2 T / Ψx 2 +Ψ 2 T / Ψy 2 ; Based on the Laplace operator, the finite difference method is used to calculate the temperature gradient field magnitude = [(ΨT / Ψx) 2 +(ΨT / Ψy) 2 ] 1 / 2 and direction angle=arctan(ΨT / Ψy, ΨT / Ψx); Laplace operator, temperature gradient field and direction form gradient feature vector FG(t);

[0053] S225. Based on the time feature vector FT(t), the space feature vector FS(t), the dynamic feature vector FD(t) and the gradient feature vector FG(t), standardization processing is performed to obtain normalized feature vectors FT'(t), FS'(t), FD'(t) and FG'(t); the normalized feature vectors are weightedly combined according to preset weights {wT, wS, wD, wG} to obtain the spatiotemporal feature sequence data F(t)=[wT·FT'(t), wS·FS'(t), wD·FD'(t), wG·FG'(t)].

[0054] This embodiment solves the problem of integrity and accuracy of the dynamic characteristic representation of the temperature of the power consumption information collection terminal by realizing a multi-dimensional spatiotemporal feature extraction and fusion mechanism. Specifically, a statistical feature extraction method based on a time window is adopted to fully capture the statistical law of temperature change; by calculating the first-order and second-order spatial derivatives of the temperature field, the spatial variation characteristics of the temperature distribution are accurately described; the dynamic characteristics of temperature change are effectively extracted by combining the covariance analysis of the rate of change field and the acceleration field; by introducing the Laplace operator and gradient field analysis, the spatial position and diffusion trend of temperature anomalies are accurately identified. Especially in the scenario where the temperature rises rapidly due to a sharp change in terminal load, a comprehensive characterization of temperature change characteristics is achieved through the weighted fusion mechanism of feature vectors. This embodiment increases the early recognition rate of temperature anomalies to 92%, extends the warning lead time to 300ms, and improves the spatial positioning accuracy to the millimeter level, laying the foundation for accurate temperature prediction and control.

[0055] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0056] S31, calculate the attention weight coefficient αF(t) of the spatiotemporal feature sequence data F(t), αF(t)=exp(WF·F(t)) / Σexp(WF·F(t)), where WF is a preset spatiotemporal feature weight matrix, exp is an exponential function, and F(t) is the spatiotemporal feature sequence data; calculate the attention weight coefficient αP(t) of the temperature field topological feature vector data P(t), αP(t)=exp(WP·P(t)) / Σexp(WP·P(t)), where WP is a preset topological feature weight matrix, and P(t) is the temperature field topological feature vector data; perform weighted fusion of the spatiotemporal feature sequence data and the temperature field topological feature vector data according to the attention weight coefficients αF(t) and αP(t) to obtain fused feature vector data H(t);

[0057] S32, read a predetermined number of historical fused feature vector data {H(tk), ..., H(t-1)} from a preset data cache, combine the historical fused feature vector data with the current fused feature vector data H(t) to form an input sequence; send the input sequence to a pre-trained gated recurrent unit network, and calculate the predicted sequence data {H'(t+1), ..., H'(t+m)} for the next m time points; wherein the gated recurrent unit network includes an update gate Zt=σ(Wz·[H(t-1), Xt]+bz) and a reset gate Rt=σ(Wr·[H(t-1), Xt]+br), σ is a sigmoid activation function, Wz and Wr are weight matrices, bz and br are bias vectors, and Xt represents the input feature vector of the current time step;

[0058] S33. Calculate the differential sequence {ΔH'(t+1), ..., ΔH'(t+m-1)} of the prediction values ​​at adjacent moments in the prediction sequence data; based on the differential sequence, calculate the Lyapunov exponent λ(t), λ(t)=(1 / N)·Σln|ΔH'(t+i)|, where N is the prediction sequence length, ln is the natural logarithm function, and ΔH'(t+i) is the differential value of the prediction sequence data at time t+i; based on the Lyapunov exponent λ(t), calculate the system stability index data S(t), S(t)=1 / (1+exp(β·λ(t))), where β is a preset scaling factor.

[0059] In one embodiment of the present application, based on the spatiotemporal feature sequence data F(t) and the temperature field topological feature vector data P(t), the attention mechanism is used to fuse the spatiotemporal and topological features, and the fused feature vector H(t) is output. Based on the historical sequence {H(tk), ..., H(t)} of the fused feature vector H(t), the GRU network is used to predict the state of the next m time points, and the predicted sequence {H'(t+1), ..., H'(t+m)} is output. Based on the predicted sequence {H'(t+1), ..., H'(t+m)}, the Lyapunov exponent λ is calculated, and the system stability index S(t) is output.

[0060] This embodiment realizes adaptive weight allocation of spatiotemporal features and topological features by constructing a feature fusion framework based on the attention mechanism, and solves the problem of dynamic changes in feature importance of the power consumption information collection terminal under different working conditions. Specifically, the local sensitivity analysis method is used to evaluate the importance of features, which can automatically screen out the feature subset that is most indicative of the system status and reduce the interference of redundant features; by introducing the self-attention mechanism, the system can automatically adjust the weights of different features according to the current working status, so that the fused features are closer to the actual working situation. Especially in the workload mutation scenarios often faced by terminals, this embodiment realizes the precise alignment of feature sequences through a dynamic time warping algorithm, effectively solving the problem of inconsistent time scales of different features. This embodiment improves the accuracy of system status recognition by 12 percentage points, reduces the false alarm rate by 85% under complex working conditions, and improves the reliability and stability of the system.

[0061] According to one aspect of the present application, step S31 is further:

[0062] S311, obtain the spatiotemporal feature sequence data F(t) and the temperature field topological feature vector data P(t); calculate the local sensitivity Si=Ψy / Ψxi of each feature component, where y is the system output and xi is the feature component; calculate the feature importance score I(xi) based on the local sensitivity; filter the feature subsets F'(t) and P'(t) according to the feature importance scores.

[0063] S312, obtain feature subsets F'(t) and P'(t); construct query matrix Q(t)=[F'(t);P'(t)] and key matrix K(t)=[F'(t);P'(t)]; calculate attention weight matrix A(t)=softmax(Q(t)·K(t) T / sqrt(d)), where d is the feature dimension; and the self-attention feature FA(t) is calculated according to the attention weight matrix A(t).

[0064] S313, obtain feature subsets F'(t) and P'(t); use a dynamic time warping algorithm to calculate the alignment matrix D(t) of spatiotemporal features and topological features; resample the feature sequence based on the alignment matrix D(t) to obtain aligned feature sequences F''(t) and P''(t).

[0065] S314, obtain the self-attention feature FA(t) and the aligned feature sequence F''(t), P''(t); construct a nonlinear feature transformation function φ(x); calculate the transformed feature vectors φ(FA(t)), φ(F''(t)) and φ(P''(t)); use the weighted summation method to obtain the fused feature vector data H(t)=w1·φ(FA(t))+w2·φ(F''(t))+w3·φ(P''(t)), where w1, w2, w3 are preset fusion weights.

[0066] This embodiment solves the problems of adaptability and robustness of temperature feature representation of the power consumption information collection terminal by constructing a deep feature extraction framework based on local sensitivity analysis and multimodal feature fusion. Specifically, a local sensitivity analysis method based on partial derivatives is adopted to realize the dynamic evaluation of feature importance and effectively screen out the most representative feature subset; by introducing the adaptive weight calculation of the attention mechanism, the system can automatically adjust the importance weights of different features according to the current state; combined with the dynamic time warping algorithm, the inconsistency problem of spatiotemporal features and topological features on the time scale is solved. Especially in the scenario where the terminal working state is frequently switched, a smooth transition of feature representation is achieved through nonlinear feature transformation and weighted fusion strategy. This embodiment increases the accuracy of feature fusion to 96%, improves the anti-interference ability to noise by 80%, and achieves the real-time performance of feature extraction at the millisecond level, enhancing the adaptability and reliability of the system.

[0067] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0068] S41. Obtain the CPU usage data U(t) and memory usage data M(t) collected by the CPU status monitoring module, calculate the first statistical feature of the CPU usage data, including the first mean μU(t) and the first standard deviation σU(t) within the preset time window; calculate the second statistical feature of the memory usage data, including the second mean μM(t) and the second standard deviation σM(t) within the preset time window; combine the fused feature vector data H(t), the first statistical feature [μU(t), σU(t)] and the second statistical feature [μM(t), σM(t)] into a workload feature vector L(t); input the workload feature vector into a pre-trained multi-class classifier to obtain load type indicator data W(t); wherein the value range of W(t) is {1, 2, 3, 4}, corresponding to light load, medium load, heavy load and peak load states, respectively;

[0069] S42, calculate the change rate of the n filtered second temperature sequence data T2i'(t) in the time dimension vi(t)=[T2i'(t)-T2i'(t-Δt)] / Δt, where the value range of i is 1 to n, and Δt is the sampling time interval; based on the change rate, calculate the consistency index ρ(t) of the ambient temperature change, ρ(t)=Σ(vi(t)·vj(t)) / (|vi(t)|·|vj(t)|), where the value range of i, j is 1 to n; based on the consistency index ρ(t), calculate the environmental impact coefficient α(t), α(t)=1 / (1+exp(-γ·ρ(t))), where γ is a preset adjustment coefficient;

[0070] S43. Read a benchmark threshold vector B(W(t)) corresponding to the load type index data W(t) from a preset threshold benchmark database; based on the environmental impact coefficient α(t), correct the benchmark threshold vector B(W(t)) to obtain the environmental correction threshold E(t)=α(t)·B(W(t)); based on the system stability index data S(t), calculate the stability weight ω(t)=exp(-Δ·S(t)), where Δ is the preset attenuation coefficient; multiply the environmental correction threshold E(t) by the stability weight ω(t) to obtain a multi-level warning threshold set data {Thi(t)}, where i ranges from 1 to m, representing m different levels of warning thresholds.

[0071] In one embodiment of the present application, the CPU utilization rate U(t), the memory utilization rate M(t) and the fused feature vector data H(t) are obtained, the workload feature vector L(t) is constructed, and the load type index W(t) is output. Based on the filtered n second temperature sequence data {T2i'(t)} and the environmental change rate, the environmental impact model E(t) is constructed, and the environmental compensation coefficient α(t) is output. Based on the load type index W(t), the environmental compensation coefficient α(t) and the system stability index data S(t), the adaptive threshold Th(t) = f(W(t), α(t), S(t)) is calculated, and a multi-level warning threshold set {Thi(t)} is output.

[0072] This embodiment solves the problem of setting the temperature threshold of the power consumption information collection terminal under complex working conditions by establishing a multi-dimensional workload evaluation system and an adaptive threshold optimization mechanism. Specifically, by extracting multi-scale statistical features of CPU usage and memory usage, combined with load fluctuation feature analysis, an accurate characterization of the system working state is achieved; the use of ambient temperature dynamic compensation technology enables the warning threshold to be automatically adjusted according to changes in ambient temperature, effectively avoiding the false alarm and missed alarm problems caused by traditional fixed thresholds. Especially in the case of drastic changes in the ambient temperature of the terminal, by introducing the environmental impact coefficient, the system can accurately distinguish between temperature changes caused by environmental factors and temperature changes caused by equipment abnormalities. This embodiment reduces the false alarm rate of the system from the original 15% to less than 3%, and the missed alarm rate from 10% to less than 1%. At the same time, the response time of the threshold adjustment is shortened to less than 100ms, which improves the adaptability and reliability of the system.

[0073] According to one aspect of the present application, step S41 is further:

[0074] S411. Obtain the CPU usage data U(t) and memory usage data M(t) collected by the CPU status monitoring module; use a low-pass filter to remove high-frequency noise, and obtain smoothed sequences Us(t) and Ms(t) after filtering; calculate the first-order differences ΔUs(t) and ΔMs(t) of the filtered sequences in the time window w; calculate the load change rate ru(t)=ΔUs(t) / Δt and rm(t)=ΔMs(t) / Δt according to the differential sequence.

[0075] S412. Obtain the smoothed sequence Us(t), Ms(t) and the rate of change ru(t), rm(t); calculate the mean μu(t)=mean(Us(tw:t)) and μm(t)=mean(Ms(tw:t)) in the time window w; calculate the standard deviation σu(t)=std(Us(tw:t)) and σm(t)=std(Ms(tw:t)); calculate the coefficient of variation cu(t)=σu(t) / μu(t) and cm(t)=σm(t) / μm(t).

[0076] S413. Obtain fused feature vector data H(t); extract the load-related subvector HL(t) in H(t); calculate the autocorrelation coefficient sequence R(τ) of HL(t) in the time window w; calculate the characteristic period T(t) and fluctuation intensity A(t) based on the autocorrelation coefficient sequence; construct the load fluctuation feature vector L1(t)=[T(t), A(t)].

[0077] S414. Obtain statistical characteristics [μu(t), σu(t), cu(t), μm(t), σm(t), cm(t)] and load fluctuation characteristic vector L1(t); construct workload characteristic vector L(t)=[μu(t), σu(t), cu(t), μm(t), σm(t), cm(t), L1(t)]; normalize the workload characteristic vector L(t) to obtain a normalized characteristic vector L'(t).

[0078] S415. Obtain the normalized feature vector L'(t); calculate the probability distribution P(w|L'(t)) that L'(t) belongs to each load type based on the pre-trained multi-class classifier; use the maximum a posteriori probability criterion to determine the load type indicator data W(t)=argmax P(w|L'(t)), where w∈{1, 2, 3, 4} corresponds to light load, medium load, heavy load and peak load states respectively.

[0079] This embodiment solves the accuracy and real-time problems of workload identification of the power consumption information collection terminal by implementing a multi-dimensional load feature extraction and classification evaluation mechanism. Specifically, a method combining low-pass filtering and first-order difference is used to effectively extract the dynamic change characteristics of CPU and memory usage; by calculating the statistical characteristics of multiple time scales, the load fluctuation law is comprehensively characterized; combined with autocorrelation analysis, the periodic characteristics of load changes are accurately identified. Especially when the terminal is faced with complex and changeable work tasks, accurate identification of different load types is achieved through multi-feature fusion and probabilistic classification methods. This embodiment makes the accuracy of load type identification reach 94%, the response time is shortened to less than 50ms, and the false alarm rate is reduced to less than 1%, providing a reliable basis for implementing precise temperature control strategies.

[0080] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0081] S51, based on the fused feature vector data H(t), the prediction sequence data {H'(t+1), ..., H'(t+m)} and the multi-level warning threshold set data {Thi(t)}, construct the cost function J(t) within N steps of the prediction time domain, and the calculation formula is J(t)=Σ[q·(H'(t+k)-Thi(t)) 2 +r·u(t+k) 2 ], where k ranges from 1 to N, q is the state deviation weight coefficient, r is the control quantity weight coefficient, and u(t+k) is the control quantity; solve the control sequence {u(t), ..., u(t+N-1)} corresponding to the minimum value of the cost function J(t); extract the control parameters at the current moment from the control sequence, and combine them with the pre-stored historical optimal control parameters to obtain the control parameter set data K(t);

[0082] S52, based on the control parameter set data K(t), the temperature control interval is divided into a predetermined number of subintervals {[Tj, Tj+1]}, where the value of j ranges from 0 to p-1; for each subinterval [Tj, Tj+1], based on the fused feature vector data H(t), a linear control law uj(t)=kj·(H(t)-Tj)+bj is calculated, where kj and bj are control parameters corresponding to the subinterval; the linear control laws of the predetermined number of subintervals are combined into a piecewise linear control instruction sequence C(t);

[0083] S53, obtain the power consumption data P(t) collected by the power consumption monitoring module, calculate the cumulative value Pa(t) of the power consumption data P(t) within the preset time window; calculate the temperature change ΔT(t) corresponding to the piecewise linear control instruction sequence C(t), the calculation formula is ΔT(t)=H(t)-H(t-τ), where τ is the preset response time; based on the cumulative value and the temperature change, calculate the temperature control efficiency ratio η(t)=ΔT(t) / Pa(t); based on the temperature control efficiency ratio η(t), correct the piecewise linear control instruction sequence C(t), the correction formula is C'(t)=C(t)·min(1, ηmax / η(t)), where ηmax is the preset maximum efficiency ratio threshold; when the temperature control efficiency ratio η(t) is greater than ηmax, reduce the control instruction strength proportionally to obtain the final control instruction sequence C'(t).

[0084] In one embodiment of the present application, based on the fusion feature vector data H(t), the multi-level warning threshold set data {Thi(t)}, and the system stability index data S(t), the model predictive control (MPC) is used to optimize the control parameters and output the control parameter set K(t). Based on the control parameter set K(t) and the current state H(t), a piecewise linear control strategy is generated and a control instruction sequence C(t) is output. Based on the control instruction sequence C(t) and the power consumption data P(t), the energy efficiency ratio η(t) = ΔT / P is calculated, and the optimized control instruction C'(t) is output.

[0085] The actual temperature response and expected response are obtained, the control deviation and response characteristics are calculated, and the control effect score R(t) is output. Based on the control effect score R(t) and historical control data, the control parameters are updated based on reinforcement learning, and the updated control parameter set K'(t) is output. Based on the updated control parameter set K'(t), the control effect score R(t) and the system model, the control strategy is optimized and the optimized control strategy is output.

[0086] This embodiment solves the problem of balancing the accuracy and energy efficiency of the CPU temperature control of the power consumption information collection terminal by implementing a piecewise linear control strategy based on model prediction. Specifically, an optimization algorithm that considers state constraints and control constraints in the prediction time domain is adopted to achieve global optimization of control parameters, avoiding the overshoot and oscillation problems that are prone to occur in traditional PID control methods; by introducing piecewise linear control and spline function smoothing technology, while ensuring the control response speed, a smooth transition of the control amount is achieved, effectively reducing the energy loss in the control process. Especially in scenarios where the terminal needs to respond quickly to temperature changes, the optimal balance between control effect and energy consumption is achieved through energy efficiency ratio optimization technology. This embodiment improves the temperature control accuracy of the system to ±0.5°C, reduces the overshoot to less than 5%, and at the same time improves the energy utilization efficiency by 35%, thereby improving the control performance and energy utilization efficiency of the system.

[0087] According to one aspect of the present application, step S51 is further:

[0088] S511. Based on the fused feature vector data H(t) and the prediction sequence data {H'(t+1), ..., H'(t+m)}, a system state prediction model x(t+1)=f(x(t), u(t)) is constructed, where x(t) is the state vector, u(t) is the control input, and f represents the dynamic change function of the system state; based on the system state prediction model, the least squares method is used to identify the model parameters to obtain the prediction model parameter matrix Θ;

[0089] S512, based on the multi-level warning threshold set data {Thi(t)}, construct the tracking error term et(k)=H'(t+k)-Thi(t) and the control amount penalty term eu(k)=u(t+k) within the prediction time domain N steps; based on the tracking error term and the control amount penalty term, construct a multi-objective cost function to obtain the optimization problem J(t)=Σ[q·et(k) 2 +r·eu(k) 2 ], where k = 1, 2, ..., N, q is the state weight, and r is the control weight;

[0090] S513. Based on the prediction model parameter matrix Θ, construct the control quantity constraint umin≤u(t)≤umax and the change rate constraint |u(t)-u(t-1)|≤Δumax; based on the control quantity constraint and the change rate constraint, construct the feasible domain D of the optimization problem; based on the feasible domain, use the quadratic programming method to solve the optimization problem and obtain the optimization solution min{J(t):u(t)∈D};

[0091] S514. Based on the optimization solution, extract the optimal control quantity u*(t) at the current moment, and construct the control parameter set data K(t) by combining the historical optimal control parameters {u*(tk), ..., u*(t-1)}; based on the control parameter set data, calculate the control parameter statistical characteristics, including the mean μk, variance σk 2 and autocorrelation coefficient ρk;

[0092] S515. Based on the control parameter set data K(t) and the statistical characteristics of the control parameters, a bootstrap method is used to generate parameter disturbance samples {K1(t), ..., Kb(t)}; based on the parameter disturbance samples, the influence of the parameter disturbance on the system response is evaluated to obtain the robustness index R(t); based on the robustness index, the control parameter set data is corrected to obtain the final control parameter set data K'(t).

[0093] This embodiment solves the accuracy and stability problems of temperature control of the power consumption information collection terminal by constructing an adaptive control parameter optimization framework based on the prediction model. Specifically, an accurate state prediction model is constructed by using the system identification method; by introducing the cost function of multi-objective optimization, the control accuracy and control quantity constraints are balanced; combined with the quadratic programming method, global optimal control under constraints is achieved. Especially in the case of sudden changes in the terminal working environment and load conditions, the robustness of the parameters is evaluated by the bootstrap method, and adaptive corrections are performed to ensure the reliability of the control strategy. This embodiment reduces the steady-state error of temperature control to ±0.2°C, shortens the dynamic response time to 150ms, and controls the overshoot within 3%, thereby improving the control performance of the system.

[0094] According to one aspect of the present application, step S52 is further:

[0095] S521. Obtain control parameter set data K(t); divide the temperature control interval [Tmin, Tmax] into p sub-intervals {[Tj, Tj+1]} based on the system dynamic response characteristics, where j=0, 1, ..., p-1; calculate the characteristic parameters of each sub-interval, including the interval width wj, the center temperature Tcj and the temperature change rate threshold vj.

[0096] S522. Obtain subinterval parameters; for each subinterval [Tj, Tj+1], construct a linear control law uj(t)=kj·(H(t)-Tcj)+bj, where kj and bj are unknown coefficients; solve the coefficients kj and bj based on the subinterval boundary conditions and continuity constraints; and construct a piecewise linear function coefficient matrix C1.

[0097] S523, obtain the coefficient matrix C1; calculate the connection point position xj and derivative value yj of the adjacent sub-interval control law; construct the constraint equation group of the cubic spline function sj(x) at the connection point; solve the constraint equation group to obtain the spline function coefficient matrix C2.

[0098] S524, obtaining a spline function coefficient matrix C2; generating a composite control law u(t) by combining the piecewise linear function and the spline function; performing discrete sampling on the composite control law to obtain an initial control instruction sequence C0(t).

[0099] S525, obtain the initial control instruction sequence C0(t); calculate the time derivative du / dt of the control instruction; detect and smooth the mutation points in the instruction sequence; apply a moving average filter to process the control instruction sequence to obtain the final piecewise linear control instruction sequence C(t).

[0100] This embodiment solves the problem of stability and economy of temperature control of the power consumption information collection terminal by implementing a composite control strategy of piecewise linear control and spline function smoothing. Specifically, the temperature control interval is adaptively divided based on the dynamic characteristics of the system to achieve segmented optimization of the control strategy; by constructing a spline function that considers boundary conditions and continuity constraints, the smooth transition of the control instructions when switching between intervals is ensured; combined with moving average filtering, the sudden change and oscillation of the control instructions are effectively suppressed. Especially in the scenario where the terminal needs to accurately adjust the temperature, the design of the composite control law achieves the unity of fast response and smooth control. This embodiment reduces the energy loss of the control process by 40%, improves the smoothness of the control instructions by 85%, and improves the stability of the system by 75%, effectively improving the economy and reliability of temperature control.

[0101] In another embodiment of the present application, a method for controlling CPU junction temperature includes: obtaining the junction temperature of the main control CPU of the power consumption information collection terminal and the temperature set at other positions of the mainboard of the power consumption information collection terminal away from the CPU as a reference for the ambient temperature of the terminal; obtaining the first CPU junction temperature T1 of the CPU through the CPU junction temperature; obtaining the second ambient temperature T2 of the terminal through the internal temperature sensor away from the CPU; comparing the first CPU junction temperature T1 with the third CPU junction temperature preset value T3 as a basis for determining whether to intervene in temperature control; judging the abnormality of the power consumption information collection terminal according to the difference change between the first CPU junction temperature T1 and the second ambient temperature T2; judging whether the temperature control reaches the target value according to the comparison between the first CPU junction temperature T1 and the fourth CPU junction temperature preset value T4;

[0102] The working status of the power consumption information collection terminal is judged by the change of the difference between the first CPU junction temperature T1 and the second terminal ambient temperature T2, including: when the difference between the first CPU junction temperature T1 and the second terminal ambient temperature T2 is greater than the fifth preset temperature threshold T5 and is maintained for a preset time, it is judged that the power consumption information collection terminal is in an abnormal working state.

[0103] The change in the difference between the first CPU junction temperature T1 and the third CPU junction temperature preset value T3 determines whether to perform temperature control intervention, including: when the difference between the first CPU junction temperature T1 and the third CPU junction temperature T3 is less than or equal to 0 and is maintained for a preset time, it is determined that the temperature control strategy of the power consumption information collection terminal needs to intervene.

[0104] When the CPU junction temperature control strategy intervenes, the difference change between the first CPU junction temperature T1 and the fourth CPU junction temperature preset value T4 determines whether the expected control target has been reached, including: when the absolute value of the difference between the first CPU junction temperature T1 and the fourth CPU junction temperature preset value T4 is less than or equal to the preset control target value T6, and is maintained for a preset time, it is determined that the temperature control strategy of the power consumption information collection terminal has reached the expected target.

[0105] In another embodiment of the present application, a method for controlling CPU junction temperature includes: step S1, obtaining a first temperature of CPU internal junction temperature; step S2, obtaining a second temperature of ambient temperature through a temperature sensor on a single board away from the CPU; step S3, obtaining a third temperature of CPU junction temperature warning threshold through a CPU external EEPROM chip; step S4, obtaining a fourth temperature of CPU junction temperature control target value through a CPU external EEPROM chip; step S5, obtaining a fifth temperature of CPU junction temperature and ambient temperature difference threshold through a CPU external EEPROM chip; step S6, obtaining a sixth temperature of CPU junction temperature and fourth temperature difference threshold through a CPU external EEPROM chip; step S7, judging the working condition of the terminal through the change of the difference between the first temperature and the second temperature; step S8, judging whether the terminal needs to intervene to control the CPU junction temperature through the change of the difference between the first temperature and the third temperature; step S9, judging whether the terminal CPU junction temperature control reaches the predetermined target by comparing the difference between the first temperature and the fourth temperature with the fifth temperature.

[0106] According to one aspect of the present application, a control system for CPU junction temperature includes:

[0107] at least one processor; and,

[0108] a memory communicatively connected to at least one of the processors; wherein,

[0109] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the method for controlling the CPU junction temperature described in any one of the above embodiments.

[0110] The present invention uses sliding median filtering technology to collaboratively process the CPU internal temperature and multiple ambient temperature sensor data, and combines it with an abnormal point detection mechanism to achieve accurate identification and filtering of abnormal temperature data, thereby increasing the effectiveness of temperature monitoring to more than 98%. By improving the Delaunay triangulation and Steiner point optimization strategy, the adaptive grid division of the temperature field is realized, and the segmented cubic Hermite interpolation method is used to ensure the accurate reconstruction of the temperature distribution in the hot spot area, with a reconstruction accuracy of 0.1°C. By constructing a multi-dimensional feature extraction mechanism including time statistical features, spatial derivative features and dynamic features, and combining the attention mechanism to achieve adaptive fusion of features, the prediction lead time of temperature anomalies is extended to 300ms. By introducing an adaptive threshold optimization mechanism based on workload evaluation and environmental impact compensation, the dynamic adjustment of the warning threshold is realized, and the false alarm rate is reduced to less than 3%. The optimization algorithm that considers the state constraints and control constraints in the prediction time domain is adopted to achieve global optimization of control parameters. Through piecewise linear control and spline function smoothing technology, the smooth transition of control instructions is ensured, the energy loss and system instability caused by control switching are effectively reduced, and the smoothness of control instructions is improved by 85%. The present invention comprehensively solves the key problems faced in the monitoring and control of CPU junction temperature of power consumption information collection terminals by constructing a complete "collection-analysis-prediction-optimization-control" closed-loop system. The system realizes accurate prediction of temperature anomalies through multi-source data fusion and deep feature extraction technology; and realizes accurate temperature regulation through adaptive threshold optimization and model prediction control strategy. The present invention optimizes the application scenarios of power consumption information collection terminals: first, through multi-point temperature collaborative monitoring and environmental impact compensation technology, the temperature monitoring problem of the terminal in a complex external environment is solved; secondly, through load-aware adaptive control strategy, the temperature control problem of the terminal under different working conditions is solved; finally, through energy efficiency optimization technology, the energy consumption problem in long-term operation of the terminal is solved. The present invention enables the system's temperature anomaly prediction accuracy to reach more than 95%, the control accuracy is improved to ±0.5℃, and the energy utilization efficiency is improved by 35%, which comprehensively improves the reliability, stability and economy of the terminal.

[0111] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for controlling CPU junction temperature, characterized in that: The steps include: S1, obtaining first temperature data collected by a CPU internal temperature sensor and n second temperature data collected by n ambient temperature sensors arranged around the CPU, and processing the first temperature data and the n second temperature data by sliding median filtering to obtain filtered first temperature sequence data and n second temperature sequence data; Constructing a temperature field grid topology map based on the filtered first temperature series data and the n second temperature series data; Based on the temperature data in the temperature field grid topology diagram, the Z-score standardization method is used to calculate the standardized temperature field data; where n is a natural number greater than 1; S2. Calculate the first-order and second-order derivatives of the standardized temperature field data in the time dimension to obtain temperature change rate field data and temperature change acceleration field data; construct spatiotemporal feature sequence data based on the standardized temperature field data, the temperature change rate field data and the temperature change acceleration field data; calculate and obtain the temperature field topological feature vector data based on the temperature distribution relationship in the standardized temperature field data; S3, based on the spatiotemporal feature sequence data and the temperature field topological feature vector data, the attention weight is used to calculate the fusion feature vector data; Based on the historical sequence of fused feature vector data, a gated recurrent unit network is used to calculate the predicted sequence data of m future time points; based on the predicted sequence data, the system stability index data is calculated; where m is a natural number greater than 1; S4, obtaining CPU usage data and memory usage data collected by the CPU status monitoring module; Based on the fused feature vector data, CPU usage data, and memory usage data, a workload feature vector is constructed; Based on the workload feature vector, load type index data is obtained after classification processing; based on the filtered n second temperature series data, the environmental impact coefficient is calculated; based on the load type index data, the environmental impact coefficient and the system stability index data, the multi-level warning threshold set data is calculated; S5. Based on the fused feature vector data and the multi-level warning threshold set data, a prediction model is constructed and optimized to obtain the control parameter set data; based on the control parameter set data, a piecewise linear control instruction sequence is generated; the power consumption data collected by the power consumption monitoring module is obtained, and based on the power consumption data, the piecewise linear control instruction sequence is optimized to obtain the final control instruction sequence.

2. The method for controlling CPU junction temperature according to claim 1, characterized in that: Step S1 is further as follows: S11, obtaining first temperature data collected by a CPU internal temperature sensor at time t and n second temperature data collected by n ambient temperature sensors arranged around the CPU at time t; performing median filtering on the first temperature data and the n second temperature data using a sliding window to obtain filtered first temperature sequence data and n second temperature sequence data; S12, taking the sampling points of the filtered first temperature series data and the n sampling points of the second temperature series data as vertices, and using the Delaunay triangulation algorithm to construct a temperature field grid topology graph having a vertex set and an edge set; S13. Based on the pre-stored historical temperature data, the temperature mean and standard deviation are calculated; based on the temperature mean and standard deviation, the temperature data of the temperature field grid topology diagram is standardized according to the formula Z=(x-μ) / σ to obtain standardized temperature field data; wherein x is the temperature data to be standardized, μ represents the temperature mean, σ represents the standard deviation, and Z represents the standardized temperature field data.

3. The method for controlling CPU junction temperature according to claim 2, characterized in that: Step S2 is further as follows: S21. Using the central difference method, calculate the first-order derivative of the standardized temperature field data in the time dimension to obtain the temperature change rate field data; the calculation formula of the first-order derivative is v(t)=[G'(t+Δt)-G'(t-Δt)] / (2Δt); The central difference method is used to calculate the second-order derivative of the standardized temperature field data in the time dimension to obtain the temperature change acceleration field data. The second-order derivative calculation formula is a(t)=[G'(t+Δt)-2G'(t)+G'(t-Δt)] / (Δt 2 ), where Δt is the sampling time interval, and G'(t) represents the standardized temperature field data at time t; S22, calculating the spatial gradient of the standardized temperature field data to obtain temperature gradient field data; combining the standardized temperature field data, the temperature change rate field data, the temperature change acceleration field data and the temperature gradient field data into a four-dimensional vector at each sampling point to obtain spatiotemporal characteristic sequence data; S23. Based on the temperature value of each sampling point in the standardized temperature field data, the degree distribution eigenvalue, clustering coefficient and centrality index of each sampling point are calculated; the degree distribution eigenvalue, clustering coefficient and centrality index are combined to form the temperature field topological feature vector data.

4. The method for controlling CPU junction temperature according to claim 3, characterized in that: Step S3 is further as follows: S31, calculating the attention weight coefficient αF(t) of the spatiotemporal feature sequence data, αF(t)=exp(WF·F(t)) / Σexp(WF·F(t)), where WF is a preset spatiotemporal feature weight matrix, exp is an exponential function, and F(t) is the spatiotemporal feature sequence data; Calculate the attention weight coefficient αP(t) of the temperature field topological feature vector data, αP(t)=exp(WP·P(t)) / Σexp(WP·P(t)), where WP is the preset topological feature weight matrix, and P(t) is the temperature field topological feature vector data; According to the attention weight coefficients αF(t) and αP(t), the spatiotemporal feature sequence data and the temperature field topological feature vector data are weightedly fused to obtain the fused feature vector data; S32, reading a predetermined number of historical fused feature vector data from a preset data cache, combining the historical fused feature vector data with the current fused feature vector data to form an input sequence; sending the input sequence to a pre-trained gated recurrent unit network, and calculating the predicted sequence data for the next m time points; The gated recurrent unit network includes an update gate and a reset gate; S33. Calculate the differential sequence of prediction values ​​at adjacent moments in the prediction sequence data; based on the differential sequence, calculate the Lyapunov exponent λ(t), λ(t)=(1 / N)·Σln|ΔH'(t+i)|, where N is the prediction sequence length, ln is the natural logarithm function, and ΔH'(t+i) is the differential value of the prediction sequence data at time t+i; based on the Lyapunov exponent λ(t), calculate the system stability index data S(t), S(t)=1 / (1+exp(β·λ(t))), where β is a preset scaling factor.

5. The method for controlling CPU junction temperature according to claim 4, characterized in that: Step S4 is further as follows: S41, obtaining CPU usage data and memory usage data collected by the CPU status monitoring module, and calculating a first statistical feature of the CPU usage data, including a first mean and a first standard deviation within a preset time window; Calculating a second statistical feature of the memory usage data, including a second mean and a second standard deviation within a preset time window; combining the fused feature vector data, the first statistical feature, and the second statistical feature into a workload feature vector; Input the workload feature vector into the pre-trained multi-class classifier to obtain the load type indicator data; S42, calculating the change rate of the filtered n second temperature series data in the time dimension; Based on the rate of change, calculate the consistency index of ambient temperature change; Based on the consistency index, calculate the environmental impact coefficient; S43, reading a reference threshold vector corresponding to the load type index data from a preset threshold reference database; based on the environmental impact coefficient, correcting the reference threshold vector to obtain an environmental correction threshold; based on the system stability index data, calculating a stability weight; Multiply the environmental correction threshold by the stability weight to obtain the multi-level warning threshold set data.

6. The method for controlling CPU junction temperature according to claim 5, characterized in that: Step S5 is further as follows: S51, based on the fused feature vector data, the prediction sequence data and the multi-level warning threshold set data, construct a cost function within the prediction time domain N steps; solve the control sequence corresponding to the minimum value of the cost function; extract the control parameters at the current moment from the control sequence, and combine them with the pre-stored historical optimal control parameters to obtain the control parameter set data; S52, based on the control parameter set data, dividing the temperature control interval into a predetermined number of sub-intervals; for each sub-interval, based on the fused feature vector data, calculating a linear control law; combining the linear control laws of the predetermined number of sub-intervals into a piecewise linear control instruction sequence; S53, obtaining power consumption data collected by the power consumption monitoring module, and calculating the cumulative value of the power consumption data within a preset time window; The temperature variation corresponding to the piecewise linear control instruction sequence is calculated; based on the accumulated value and the temperature variation, the temperature control efficiency ratio is calculated; based on the temperature control efficiency ratio, the piecewise linear control instruction sequence is corrected to obtain a final control instruction sequence.

7. The method for controlling CPU junction temperature according to claim 6, characterized in that: Step S12 is further as follows: S121, based on the filtered first temperature sequence data and the n second temperature sequence data, calculating the spatial coordinate deviation value of each sampling point; based on the spatial coordinate deviation value and the preset temperature threshold, deleting the abnormal sampling point to obtain a valid sampling point set; S122. Based on the valid sampling point set, an improved Bowyer-Watson algorithm is used to construct an initial Delaunay triangulation to obtain a triangle set; the minimum interior angle and maximum side length ratio of each triangle in the triangle set is calculated to obtain a triangle quality evaluation index set; S123, based on the triangle set and the triangle quality evaluation index set, filter out triangles whose quality evaluation index is lower than a preset threshold, and insert new Steiner points into them to form an encrypted point set; merge the encrypted point set with the valid sampling point set to obtain an optimized sampling point set; S124, re-execute Delaunay triangulation based on the optimized sampling point set to obtain an optimized triangle set; extract all edges of the optimized triangle set to obtain an edge set; process the edge set according to a preset boundary constraint condition to obtain a final edge set; S125. Based on the optimized sampling point set and the final edge set, the piecewise cubic Hermite interpolation method is used to calculate the temperature value of any point inside the grid and construct a temperature field grid topology map.

8. The method for controlling CPU junction temperature according to claim 6, characterized in that: Step S22 is further as follows: S221, extracting the temperature sequence of each sampling point in the standardized temperature field data within a preset time window; calculating the statistical characteristics of the temperature sequence, including mean, standard deviation, skewness and kurtosis; based on Statistical features, construct time feature vector; S222. Based on the standardized temperature field data, calculate the first-order spatial derivative and the second-order spatial derivative of the temperature field in the x-direction and the y-direction; based on the first-order and second-order spatial derivatives, construct a spatial feature vector; S223, based on the temperature change rate field data and the temperature change acceleration field data, calculating the change rate field spatial covariance matrix and the acceleration field spatial covariance matrix; performing eigenvalue decomposition on the change rate field spatial covariance matrix and the acceleration field spatial covariance matrix, taking the largest k eigenvalues ​​to construct a dynamic eigenvector; wherein k is a natural number greater than 1; S224, calculating the Laplace operator of the standardized temperature field data; based on the Laplace operator, using the finite difference method, calculating the temperature gradient field and direction; combining the Laplace operator, the temperature gradient field and direction into a gradient feature vector; S225, performing standardization processing based on the time feature vector, the space feature vector, the dynamic feature vector and the gradient feature vector to obtain a normalized feature vector; performing weighted combination on the normalized feature vector according to preset weights to obtain spatiotemporal feature sequence data.

9. The method for controlling CPU junction temperature according to claim 6, characterized in that: Step S51 is further as follows: S511, constructing a system state prediction model based on the fused feature vector data and the prediction sequence data; based on the system state prediction model, using the least square method to identify model parameters to obtain a prediction model parameter matrix; S512, based on the multi-level warning threshold set data, construct a tracking error term and a control amount penalty term within N steps of the prediction time domain; based on the tracking error term and the control amount penalty term, construct a multi-objective cost function to obtain an optimization problem; S513, constructing control quantity constraint conditions and change rate constraint conditions based on the prediction model parameter matrix; Based on the control quantity constraints and the change rate constraints, the feasible domain of the optimization problem is constructed; Based on the feasible region, the quadratic programming method is used to solve the optimization problem and obtain the optimal solution; S514, based on the optimization solution result, extract the optimal control quantity at the current moment, and construct control parameter set data in combination with the historical optimal control parameters; based on the control parameter set data, calculate the statistical characteristics of the control parameters, including the mean, variance and autocorrelation coefficient; S515, based on the control parameter set data and the control parameter statistical characteristics, using the bootstrap method to generate parameter disturbance samples; Based on parameter disturbance samples, the impact of parameter disturbance on system response is evaluated to obtain robustness indicators; Based on the robustness index, the control parameter set data is modified to obtain the final control parameter set data.

10. A control system for CPU junction temperature, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the CPU junction temperature control method according to any one of claims 1 to 9.

Images